AI Aulis

AI NEWS IN ENGLISH · 15 November 2026

The most important AI news every morning. AI writes and checks the stories against the original sources – check important facts in the sources.

26 stories · next issue tomorrow morning · How the paper is made →

Issue · 15 November 2026

Today's essentials

Lead story

Cursor: leaked Fireworks key led to malicious Claude Code extension

AI-generated contentWritten by AI

bleepingcomputer.com, cybersecuritynews.com, ai.info, paktechstan.com, cursor.com · 5 sources · Sources published: 29 October–14 November 2026

Cursor said an attacker used a key issued to Fireworks AI to obtain coding agent traces. Credentials found in them gave the attacker access to Cursor’s internal services and eventually to the extension marketplace cache. On 29 October the marketplace served a backdoored Claude Code extension for about 11 hours.

Cursor, the company behind the Cursor code editor, explained how the October breach unfolded in an update published on 13 November. BleepingComputer, Cyber Security News and ai.info, which reported on the breach, rely on Cursor’s own account.

From a key to credentials and the extension marketplace

Cursor said the attacker obtained a key that the company had issued to Fireworks AI, which provides training and inference for AI models. The key gave read access to storage that Cursor used to share data with Fireworks.

The storage held coding agent traces, mostly from 2025. They came from users who did not have Privacy Mode enabled, whose data Cursor was allowed to use for model training. Some traces still contained credentials.

With these, the attacker reached online services used by Cursor and obtained, among other things, customer support messages, credentials that customers had sent to support, and archived source code repositories that are not used in production.

Fewer than 0.05 per cent of users had traces exposed and fewer than 0.01 per cent of customers had API credentials exposed, Cursor said. The figures do not show how many people received the malware.

Backdoor in the Claude Code extension

On 29 October the attacker used a stolen credential to place modified copies of Anthropic’s Claude Code extension in the cache from which Cursor’s VS Code extension marketplace serves extensions. The affected builds were versions 2.1.312 and 2.1.313 for Windows and macOS. Anthropic’s own extension and Open VSX, the marketplace from which Cursor fetches extensions, were not compromised, Cursor said.

In a security notice published on 29 October, Cursor said the extension sent the attacker information about the machine, including the names of environment variables but not their values. It also added the attacker’s SSH key to the machine and tried to open a reverse SSH tunnel. In such a tunnel the infected machine itself connects to the attacker’s server, which lets the attacker log in to the machine through it. The malicious code tried to persist on the machine using a LaunchAgent on macOS and a scheduled task on Windows.

Up to 5,000 users who requested these versions during a window of about 11 hours may have received a malicious version, Cursor said. The figure is the potential distribution, not the number of backdoors actually installed. Coding errors and missing administrator rights prevented remote login on most machines, the company said. Microsoft’s threat intelligence unit said the code may still have written the key and the startup files, ai.info reported. The attacker’s server has been taken offline.

Cursor said the breach did not reach production systems, official updates, customer repositories or the code of Privacy Mode users. The company moved inference away from Fireworks, rotated credentials and rebuilt the extension marketplace. The attacker’s last observed activity, on 1 November, was blocked, and Cursor believes the breach is contained.

How to check your machine

The guidance applies to users who installed or updated either version from Cursor’s extension marketplace on macOS or Windows on 29 October between 02.27 and 13.35 Pacific Time. An infection shows up as a key named svc-telemetry in the SSH authorized_keys file. Uninstalling the extension may not be enough, Cursor said, because scripts left by the malicious code keep running.

Cursor’s notice includes scripts for checking and cleaning the machine. The company advises using a separate, trusted machine to rotate any credentials that were accessible from the affected machine, because the attacker may have reached them later through the backdoor. It also advises considering reinstalling the machine.

What this is based on

  1. Cursor confirms security breach led to malicious Claude Code extension

    Publisher: bleepingcomputer.com · published 14 November 2026 at 19:27 (UTC+02:00) · retrieved 15 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

  2. Cursor Reveals How a Leaked Fireworks AI Key Led to a Backdoored Claude Code Extension

    Publisher: cybersecuritynews.com · published 14 November 2026 at 11:09 (UTC+02:00) · retrieved 15 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

  3. Cursor traces security incident to exposed Fireworks key

    Publisher: ai.info · published 14 November 2026 at 04:07 (UTC+02:00) · retrieved 15 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

  4. Cursor Traces Its October Breach to a Leaked Fireworks Key, Closing a Two-Week Investigation — Paktechstan

    Publisher: paktechstan.com · published 14 November 2026 at 07:59 (UTC+02:00) · retrieved 15 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

  5. Security notice — malicious version of Anthropic.claude-code (2.1.312 and 2.1.313) identified in the Cursor VS Code Extension Marketplace · Cursor

    Publisher: cursor.com · published 29 October 2026 at 14:00 (UTC+02:00) · retrieved 15 November 2026 at 05:08 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Top stories

Anthropic advises leaving ‘think hard’ out of prompts for Opus 5.5

Written by AI

claude.com + 1 source · 14 November 2026

AI-generated contentWritten by AI

claude.com, the-decoder.com · 2 sources · Sources published: 12–14 November 2026

Opus 5.5 always reasons on its own, Anthropic says. In the company’s tests, replies started sooner without lines such as ‘think hard’, with no clear drop in quality. The company recommends starting at the default reasoning effort, Medium.

On 12 November Anthropic published a guide to using Claude Opus 5.5 for knowledge work such as research, analysis and writing reports. Opus 5.5 is now the default model in Claude on Pro, Max, Ultra and Team plans, and the company recommends it for everyday work.

The technology site The Decoder summarised the guide on 14 November. Its article is based on Anthropic’s own material, so it does not independently confirm the company’s claims.

No need to ask for reasoning

The guide says Opus 5.5 always reasons before answering. The model decides how much to reason based on the request and the selected reasoning effort. Reasoning effort is a setting that lets users control how long the model may think before it replies.

Anthropic advises removing lines such as ‘think hard’ and ‘think carefully’ from prompts. In the company’s tests, replies started sooner without them, with no clear drop in quality. Anthropic does not give figures from the test. The Decoder wrote that quality did not drop measurably, which is stronger than the guide’s own wording.

The default reasoning effort is Medium. In Anthropic’s coding and knowledge-work tests, Opus 5.5 at Medium matched or beat Opus 5 at High. Higher levels reason for longer and cost more, but they do not always produce better work. The company therefore advises raising the level to High only when Medium is not enough. Users can change the level by clicking the model name next to the send button and choosing Effort.

Zimu Li of Factory, an early tester quoted in the guide, said that in Factory’s tests Opus 5.5 at Medium performed as well as Opus 5 at High while producing 20–25 per cent fewer tokens.

A cheaper model can cost almost as much

Anthropic urges teams that have used Sonnet for everyday work and saved Opus for heavy tasks to reconsider. A more capable model can finish a task with less work, which can offset much of its higher price per token, the company said.

Under token-based billing, Sonnet 5.5 costs $2 per 1 million input tokens and $10 per 1 million output tokens; Opus 5.5 costs $4 and $20. When Sonnet’s reasoning effort is raised to Extra high or Max, its cost per task approaches that of Opus 5.5, the guide says. On subscription plans, more expensive models use up the usage limit faster.

If results are weak, Anthropic advises trying a higher reasoning effort first. Moving to a larger model is worthwhile if the model clearly tried and still failed despite having the information it needed, the company said.

The guide also advises describing the goal and the finished output in the first message, attaching a chart as a file instead of typing out the figures, and naming precisely the styles you do not want. It is worth connecting Claude to your own files and systems with connectors. On Team and Enterprise plans, an owner must first enable a connector.

What this is based on

  1. Which Claude model is best to use for knowledge work

    Publisher: claude.com · published 12 November 2026 · retrieved 15 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

  2. Anthropic says outdated prompting habits are slowing down Claude Opus 5.5

    Publisher: the-decoder.com · published 14 November 2026 at 18:42 (UTC+02:00) · retrieved 15 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Top stories

Gemini 4 Carbon leads MathArena, but a run costs about 15 times as much as GPT-6.1 Sol’s

Written by AI

the-decoder.com + 3 sources · 14 November 2026

AI-generated contentWritten by AI

the-decoder.com, tbreak.com, dailyaibrief.com, oossa.com · 4 sources · Sources published: 14 November 2026

Google’s Gemini 4 Carbon achieved the best score on the MathArena maths benchmark, but its lead over Claude Fable 5.2 and GPT-6.1 Sol falls within the error margins. Carbon’s test run cost about 15 times as much as GPT-6.1 Sol’s.

On 14 November, The Decoder reported on MathArena’s new results, citing an X update by Jasper Dekoninck. The AI Brief website cites the same MathArena leaderboard figures. The site describes MathArena as an independent benchmark in which AI models solve new problems from maths competitions.

Tbreak and AI Brief reported that MathArena added Anthropic’s Claude Fable 5.2 and Google’s Gemini 4 Carbon to the benchmark on 13 November. The Decoder said Carbon is the first non-OpenAI model in a long time to reach the top. It overtook OpenAI’s GPT-6.1 Sol, and Fable 5.2 came second.

The gaps between the top three fall within the error margins

The table shows MathArena’s expected score and expected cost at the reasoning effort used for each model.

Model (reasoning effort) Company Expected score Expected cost
Gemini 4 Carbon (high) Google 85.4% (±3.0) $15.70 (±$1.33)
Claude Fable 5.2 (max) Anthropic 82.8% (±5.3) $40.20 (±$2.59)
GPT-6.1 Sol (max) OpenAI 80.6% (±2.7) $1.07 (±$0.018)
Kimi K3.1 Pro (max) Moonshot AI 54.3% (±5.8) $3.70 (±$0.27)

Source: MathArena, via The Decoder and Tbreak. Score ranges are in percentage points.

The ranges of the top three frontier models overlap. For that reason, Tbreak and AI Brief said, the contest is closer than the order in the table suggests. Fable 5.2 has the widest range of the three. Kimi K3.1 Pro, the best open-weight model, finished about 31 percentage points behind Carbon.

The Decoder points out that the OpenAI models claimed to have recently solved demanding equations such as the Navier–Stokes equations are not yet publicly available.

The cost covers the whole run, not the price per token

Tbreak said the 15-fold difference refers to the cost of the whole benchmark run at the reasoning effort used in testing, not to the price per token. The Oossa website presented the same figures as prices per thousand tokens, which contradicts Tbreak’s explanation. The Decoder put the cost of a Fable 5.2 run at about 40 times that of a GPT-6.1 Sol run, and Tbreak at about 38 times (the figures in the table give about 38).

Tbreak cites The Decoder’s observations on how much work Carbon does to explain the cost gap. Carbon’s most common failure was exceeding the time limit, and the model almost never gave up. In one case it launched a script and left it running for more than 11 hours without checking its progress. The Decoder said this persistence can be seen as both a strength and a weakness.

Google made Carbon available through the Gemini API on 10 November. Tbreak said the company is aiming the model mainly at cybersecurity and long-running tasks. Yahoo, as cited by Tbreak, reported a launch price of $2 per million input tokens and $10 per million output tokens. After the introductory period, the prices rise to $4 and $20.

What this is based on

  1. Gemini 4 Carbon takes the MathArena crown but costs 15 times more than GPT-6.1 Sol

    Publisher: the-decoder.com · published 14 November 2026 at 18:48 (UTC+02:00) · retrieved 15 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

  2. Gemini 4 Carbon tops MathArena maths benchmark

    Publisher: tbreak.com · published 14 November 2026 at 21:57 (UTC+02:00) · retrieved 15 November 2026 at 05:09 (UTC+02:00) · discovery channel: web

  3. Gemini 4 Carbon tops MathArena's maths ranking, ahead of Claude Fable 5.2 and GPT-6.1 Sol - AI Brief

    Publisher: dailyaibrief.com · published 14 November 2026 at 22:36 (UTC+02:00) · retrieved 15 November 2026 at 05:10 (UTC+02:00) · discovery channel: web

  4. Google’s Gemini 4 Carbon tops MathArena benchmark

    Publisher: oossa.com · published 14 November 2026 at 18:52 (UTC+02:00) · retrieved 15 November 2026 at 05:10 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Top stories

Dynatrace MCP server’s read tools let callers inject their own DQL query stages

Written by AI

github.com + 1 source · 13 November 2026

AI-generated contentWritten by AI

github.com, i10x.ai · 2 sources · Sources published: 13 November 2026 · Some source publication dates are unknown

The read tools in Dynatrace MCP Server inserted caller-supplied values into queries without sufficient checks, so tools whose calls some MCP clients approve without asking the user could run more than they were meant to. According to i10x, version 2.1.1 contains the fix.

On 13 November the AI news site i10x.ai reported a now-fixed injection vulnerability in Dynatrace’s MCP server, tracked as CVE-2026-123810. Dynatrace had already published security advisory GHSA-pqh8-p93p-2rx7 about it on 9 June, the site said. It put the severity at medium (CVSS 4.3).

Dynatrace MCP Server is an MCP server, a program that an AI agent uses to retrieve data from Dynatrace’s monitoring service.

How the injection worked

Five of the server’s read tools built queries in DQL, Dynatrace’s own query language, and inserted caller-supplied values into them without sufficient checks. These values included the names of monitored entities, the timeframe and cluster identifiers. According to i10x, a suitably crafted value could close a string, add new stages to the query or skip the rest of the query with a // comment. In the site’s example, the timeframe value ‘30d, to: now() | fieldsAdd internal_secret //’ adds a stage of its own before the tool’s filter.

This let a caller bypass the tools’ limits on which fields are returned, which time period is covered and how many results are shown. The tools carried the readOnlyHint annotation, which indicates that a tool changes nothing. Some MCP clients approve calls to such tools without asking the user. i10x assessed that an attacker would typically exploit the flaw through prompt injection, hiding instructions in content that the language model reads.

However, the flaw gave no access to data that the user’s token could not already reach, i10x said. The site, summarising the advisory, said the same token also gives access to the execute_dql tool, which can run any DQL query. The harm, then, is that neither the limits the read tools promised nor automatic approval could be relied on. Access rights did not expand.

What the fix prevents

The fix is in GitHub commit 15d3546, made on 14 July according to i10x. The commit adds three checks. The timeframe must be a positive integer and a unit, such as 24h or 7d. The additional filter must not contain a pipe character or a line break. Backslashes and quotation marks in values embedded in strings are escaped with a backslash so that a value cannot close the string. Each of the five files received the checks relevant to it.

A code comment says the additional filter check does not block subqueries written in square brackets, such as x in [fetch ...]. Each tool’s OAuth permissions limit these, and the developers consider this acceptable and intentional.

i10x said the flaw affects versions up to and including 1.8.5, and the site advises upgrading to version 2.1.1 or newer. It does not say whether the versions in between are affected.

What this is based on

  1. github.com

    Publisher: github.com · source publication time unknown · retrieved 15 November 2026 at 05:10 (UTC+02:00) · discovery channel: GitHub

  2. CVE-2026-123810: DQL Injection in Dynatrace MCP Server Bypasses Read-Only Safeguards

    Publisher: i10x.ai · published 13 November 2026 at 21:28 (UTC+02:00) · retrieved 15 November 2026 at 05:10 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Models

Artificial Analysis: Perplexity’s Decider v1.3 scores 70 on decision index, Flash scores 64

AI-generated contentWritten by AI

artificialanalysis.ai, x.com, casrai.org · 5 sources · Sources published: 14–15 November 2026 · Some source publication dates are unknown

Perplexity’s Decider v1.3 scored 70 on the Artificial Analysis Decision Index, and the smaller Decider Flash v1.3 scored 64 at half the cost. According to the chart in the post, the three frontier models at the top scored 73 but cost more than Decider v1.3.

Independent tester Artificial Analysis posted results on X on 14 November for Perplexity’s new decision models, Decider v1.3 and Decider Flash v1.3. Perplexity released the models the previous day, the CASRAI website reported.

A decision model chooses from fixed options

A decision model does not write a free-form answer the way a large language model does. Artificial Analysis describes it as choosing from fixed options and returning a probability for each option in a single call. In an example on the company’s page, the model decides which team a customer message should go to and whether the matter is urgent. The company said this makes the model cheap and fast, and makes its calibration measurable, meaning whether its confidence matches its accuracy.

Decision Index v2.1 combines six benchmarks. AA-Paperclip carries the most weight (30 per cent): in it, the model makes decisions and extracts information from CVs, receipts, invoices and research papers. The other benchmarks measure judgements about workplace documents, tool calls, knowledge and hallucinations, reasoning, and whether a decision stays the same when the question is rephrased.

Index scores are not hit rates. Five of the benchmarks use the Decision Score as their metric: a correct answer earns 50–100 points and a wrong one 0–50. The more confident the model, the further the score moves from 50 in either direction.

Results and costs

Decider v1.3 scored 70 on the index and Flash v1.3 scored 64. Decider v1.3 held sixth place, as its predecessor v1.2 did, Artificial Analysis said. It scored 69 on AA-Paperclip (+2) and 89 on the EvalSafe-O*NET-AA benchmark (+1).

Flash v1.3 scored four points more than v1.2. Artificial Analysis attributed the gain to a 12-point improvement on both AA-Symmetry, which measures whether a decision stays the same, and the When2Call tool-call benchmark.

At current prices, 1 million index tasks cost $68 with Decider v1.3 and $34 with Flash v1.3. Decision models are billed for input tokens only. Artificial Analysis said both models sit on the Pareto frontier: no cheaper model scored higher on the index. The chart in the post shows three models at the top with 73 points, all of them more expensive than Decider v1.3.

The three-point gap to the top is a difference in a weighted average, not a gap of three percentage points in accuracy.

Availability

Artificial Analysis lists the models at 27 billion and 4 billion parameters, with text and image input and an Apache 2.0 licence. They are available through Perplexity’s own API at $0.02 and $0.01 per 1 million input tokens. Perplexity also released the model weights on Hugging Face, CASRAI reported. The sources disagree on the size of the context window: Artificial Analysis gives 256,000 tokens, CASRAI 30,720 tokens.

What this is based on

  1. Artificial Analysis Decision Index v2.1 | Artificial Analysis

    Publisher: artificialanalysis.ai · source publication time unknown · retrieved 14 November 2026 at 05:15 (UTC+02:00) · discovery channel: web

  2. Decider v1.3 and Decider Flash v1.3 from Perplexity score 70 and 64, respectively, on the Artificial Analysis Decision Index v2.1 @perplexity_ai has released Decider v1.3 and Decider Flash v1.3…

    ArtificialAnlys · Publisher: x.com · published 14 November 2026 at 15:27 (UTC+02:00) · retrieved 15 November 2026 at 04:13 (UTC+02:00) · discovery channel: X

  3. Decider v1.3 and Decider Flash v1.3 cost $68 and $34, respectively, per one million Artificial Analysis Decision Index tasks. https://t.co/ZAkCL7jw6y

    ArtificialAnlys · Publisher: x.com · published 14 November 2026 at 15:27 (UTC+02:00) · retrieved 15 November 2026 at 04:13 (UTC+02:00) · discovery channel: X

  4. Explore the full results on Artificial Analysis: https://t.co/egA1Vz0le2

    ArtificialAnlys · Publisher: x.com · published 14 November 2026 at 15:27 (UTC+02:00) · retrieved 15 November 2026 at 04:13 (UTC+02:00) · discovery channel: X

  5. Perplexity Decider v1.3 & Flash Open Weights — CASRAI

    Publisher: casrai.org · published 15 November 2026 at 01:00 (UTC+02:00) · retrieved 15 November 2026 at 05:27 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Models

Mollick’s SatBench: 63 per cent of benchmarks favoured by AI labs are solved

Written by AI

@emollick · 15 November 2026

AI-generated contentWritten by AI

x.com · Source published: 15 November 2026 at 01:46 (UTC+02:00)

According to SatBench, compiled by Ethan Mollick, 63 per cent of the 173 benchmarks cited by AI labs since 2023 are saturated, and the pace is accelerating. Counting only independent measurements, the share is 24 per cent.

Ethan Mollick published SatBench on X on 14 November, describing it as a quick compilation. It measures how soon major AI benchmarks become saturated, or effectively beaten. ‘AI is getting better faster than our measurements are’, Mollick wrote.

What 63 per cent means

According to the charts in the post, SatBench includes benchmarks cited in at least ten model announcements from AI labs. A benchmark counts as solved once the best AI score reaches 85 per cent. By November, 109 of the 173 benchmarks had passed that threshold. The data come from Benchmark Almanac, which includes Epoch AI’s Benchmarking Hub.

The best score can come from any source, including a lab’s own announcement. A footnote to the chart puts the solved share at 24 per cent when only independent runs are counted. The 39-percentage-point gap comes from benchmarks that have crossed the threshold only in non-independent measurements, such as labs’ own announcements.

The chart shows the solved share holding at about 30 per cent for two years, because new benchmarks arrived as quickly as old ones fell. The 2023 version of SatBench had 44 benchmarks. In September 2026, the chart shows, that version passed the saturation threshold at which 85 per cent of benchmarks are solved. Now 86 per cent of them are solved.

New benchmarks fall faster than before

A second chart groups benchmarks by release year instead. Mollick said saturation is speeding up year on year. Of benchmarks released in 2023, 17 per cent were solved within a year, compared with 30 per cent of 2024 benchmarks and 43 per cent of 2025 benchmarks, the chart shows.

Within six months, 32 per cent of 2026 benchmarks were solved, compared with 0, 18 and 21 per cent in earlier years. Benchmarks already past the threshold at release count as solved immediately.

The 2026 benchmarks have not yet been tracked for a full year. They also enter the count only after their tenth announcement citation, so the chart says the calculation includes too few of this year’s benchmarks. Benchmarks from 2025 took a median of 14 months to be solved, and those from 2023 and 2024 took 20 months.

An OpenAI model has been the first to cross the 85 per cent threshold on 20 benchmarks, more often than any other lab’s model. In 2026 Anthropic leads with 11. This comparison covers benchmarks for which several labs report results. Unsolved benchmarks include Humanity’s Last Exam, where the best score is 74 per cent, as well as MMMU-Pro, SciCode and IFBench.

The calculation is not a forecast

The chart shows 64 benchmarks still unsolved. In the modelling, each has a 5 per cent chance of falling in any given month, matching the pace of the past 12 months. If no new benchmarks arrived, 85 per cent of the current ones would be solved around 2028. The last would fall around 2034, with an 80 per cent range of 2031–2037.

In a second calculation, an average of 4.7 new benchmarks arrive each month and the solving rate speeds up by one percentage point a year. In that case, 85 per cent of SatBench benchmarks would be solved around 2030. Mollick notes in the chart that the calculation is not a forecast.

What this is based on

  1. AI is getting better faster than our measurements are I threw together SatBench, which measures how quickly major AI benchmarks become saturated (effectively beaten). Of 173 benchmarks cited by the…

    emollick · Publisher: x.com · published 15 November 2026 at 01:46 (UTC+02:00) · retrieved 15 November 2026 at 04:16 (UTC+02:00) · discovery channel: X

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Models

The Information: SpaceX laid off about 60 Grok Imagine staff

AI-generated contentWritten by AI

baijiahao.baidu.com, aitimes.com, casrai.org, aitoolranked.com, genativity.com, newsquawk.com, finance.sina.com.cn · 7 sources · Sources published: 13–15 November 2026 · Some source publication dates are unknown

SpaceX’s AI unit has laid off about 60 people working on the image and video generator Grok Imagine and is now focusing on coding tools, The Information reported. The service is still running, and SpaceX has not confirmed the figure.

The technology publication The Information reported the layoffs on 13 November, citing anonymous sources. It said the staff were laid off over the previous week, leaving only a handful of people to maintain the product. Startup Fortune, citing The Information, wrote that more than half the team had been laid off.

The Information’s article is behind a paywall, so its details are known from summaries by other outlets, such as Investing.com and South Korea’s AI Times. The summaries carry no comment from SpaceX, and Genativity reported that the company did not respond to a request for comment.

Grok Imagine creates images and videos within the Grok chatbot. It is developed by SpaceXAI, formerly xAI, which SpaceX acquired in February 2026 in a share deal worth $250 billion.

Focus shifts to coding and Grok Bot

The company’s leadership wants to focus on coding tools and Grok Bot, an agent product built on the Grok chatbot, The Information reported. In August, SpaceX bought Cursor, which develops coding tools, at a valuation of $60 billion.

As recently as March, after OpenAI shut down its Sora video service, Elon Musk wrote on X that the next Grok Imagine update would be great and that the company would invest more in the field, AI Times reported.

The service is widely used. In a filing with the US Securities and Exchange Commission (SEC), SpaceX says Grok Imagine generated an average of more than 2 billion videos and about 10 billion images a month in the first quarter of this year, AI Times reported. AI Times notes that the layoffs cannot be attributed to profitability with certainty. SpaceX has not given its reasons, and the service’s revenue and costs have not been published.

Chinese companies lead

Chinese companies such as ByteDance have pulled ahead in video generation, The Information reported. Grok Imagine Video 1.5, released in June, ranked 17th among image-to-video models and 16th among text-to-video models in the Artificial Analysis rankings, AI Times wrote, citing The Information.

Genativity’s own benchmark rates models using the same prompts on a scale of 0 to 5. Grok Video scored 4.72 and ByteDance’s Seedance 2.0 scored 4.84. The 95 per cent confidence intervals of the results partly overlap.

Google and Meta are now the main US companies developing their own video models, AI Times and Investing.com reported.

Grok Imagine is still available, and xAI’s status page showed it as operational, CASRAI reported. SpaceX has not announced that it will shut down the service or its API. Nor has the company said how many developers remain or whether its development plans for the service will change.

What this is based on

  1. 消息称 SpaceXAI 裁减约 60 名 Grok Imagine 员工

    Publisher: baijiahao.baidu.com · source publication time unknown · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

  2. 스페이스X, 영상 AI '그록 이매진' 개발팀 대부분 해고...AI 코딩에 집중 - AI타임스

    Publisher: www.aitimes.com · published 14 November 2026 at 03:15 (UTC+02:00) · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

  3. SpaceXAI Reportedly Cuts 60 Grok Imagine Staff — CASRAI

    Publisher: casrai.org · published 15 November 2026 at 00:31 (UTC+02:00) · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

  4. Report: SpaceX cut roughly 60 Grok Imagine jobs

    Publisher: aitoolranked.com · published 14 November 2026 at 06:00 (UTC+02:00) · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

  5. Report: ~60 cut from SpaceX's Grok Imagine team | Genativity

    Publisher: genativity.com · published 14 November 2026 at 09:01 (UTC+02:00) · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

  6. SpaceX's (SPCX) AI group reportedly cuts around 60 Grok video generation staff, The Information reports citing sources; the co. has prioritised AI coding tools and Grok bot | Newsquawk

    Publisher: www.newsquawk.com · published 13 November 2026 at 16:22 (UTC+02:00) · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

  7. Grok Imagine遭遇团队大裁员,SpaceXAI调转赛道主攻 AI 编程工具

    Publisher: finance.sina.com.cn · published 14 November 2026 at 05:34 (UTC+02:00) · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Models

Canonical’s Inference Manager installs AI models on Ubuntu behind a single interface

Written by AI

github.com + 1 source · 14 November 2026

AI-generated contentWritten by AI

github.com, heise.de · 2 sources · Sources published: 14 November 2026 · Some source publication dates are unknown

Canonical’s new command-line tool lets users install AI models on Ubuntu and run them on their own machines. Applications reach every model through one local address.

Canonical presented Inference Manager at its own Ubuntu Summit event, the German site heise online reported on 14 November.

Inference Manager is a command-line tool for finding, installing and removing inference snaps. Snap is Ubuntu’s own package format, and an inference snap contains an AI model. Heise reported that the Snap Store offers several such models that run entirely on the user’s device without an internet connection. Canonical’s GitHub documentation says the tool also lists available models and shows the status of services.

Applications reach the installed models through a single local address. An application sends a request to port 8400 on the machine, and Inference Manager passes it to the inference snap running the model. The request names the model in the form provider/model, for example gemma4/gemma4-e4b. Desktop chat assistants, coding agents in editors and speech-to-text programs, for example, can use the same address, heise wrote.

Canonical says the interface is compatible with OpenAI’s API, so existing client programs can connect to it directly. Heise presents the compatibility as Canonical’s claim.

The tool also shows which hardware and software can run each size of a model. Options include the CPU or a GPU from AMD, Intel or Nvidia. In Canonical’s example, all four sizes of Gemma 4 run on the CPU and on AMD and Nvidia GPUs, but only one of them runs on an Intel GPU.

Myna, a dictation program released with the current Ubuntu version 26.10, is built on inference snaps, heise reported. Myna converts speech to text and writes it into any text field.

Canonical plans to build Ubuntu’s AI features on open-weight models that run locally, according to heise. The plans include context-aware help with troubleshooting, speech recognition to improve accessibility and automations based on AI agent workflows. Canonical said the features are meant to arrive as opt-in: users would enable them themselves and could remove them at any time.

Inference Manager is available from the Snap Store and can be installed from the command line or through Ubuntu’s App Center. Installation and usage instructions and the tool’s Go source code are in Canonical’s GitHub repository canonical/inference.

What this is based on

  1. github.com

    Publisher: github.com · source publication time unknown · retrieved 15 November 2026 at 05:36 (UTC+02:00) · discovery channel: GitHub

  2. Inference Manager: neues Ubuntu-Tool verwaltet lokale KI-Modelle

    Publisher: www.heise.de · published 14 November 2026 at 15:46 (UTC+02:00) · retrieved 15 November 2026 at 05:36 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Models

Frontier model improves five decision models’ accuracy by editing their weights in Cmpnd’s tests

Written by AI

github.com + 4 sources · 14 November 2026

AI-generated contentWritten by AI

github.com, marktechpost.com, cmpnd.ai, aikraft.ru, tldr.tech · 5 sources · Sources published: 13–14 November 2026 · Some source publication dates are unknown

With Cmpnd’s dspy-wsp optimiser, a frontier model reads and edits a small model’s weights without training. In the company’s own tests, the accuracy of five decision models rose by 1.9–16.1 percentage points.

Drew Breunig, CEO of Cmpnd, introduced dspy-wsp on the company’s blog on 13 November. dspy-wsp is an optimiser for the DSPy library, and the company itself describes it as highly experimental. DSPy is an open-source library for building programs that use language models and for optimising them against a metric. The new WeightSpace optimiser does not edit prompts. It changes a model’s weights without gradients or training.

The method comes from Sparks of Weight-Space Programming, a blog post by MIT CSAIL researchers Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab. In it, GPT-6-Astra edited 0.016 per cent of Qwen3.5-0.8B’s weights, and the model’s MATH-500 score rose from 7.9 per cent to 48.1 per cent. In the researchers’ experiments, editing removed obstacles but did not add new capabilities: Gemma 3 270M’s score barely changed. Cmpnd says it has not yet reproduced the study’s full result.

The frontier model tests and edits

The frontier model, which Cmpnd calls the writer, reads the small model’s weights with Python code in a sandbox. It runs the user’s program on training examples a limited number of times, works out what is going wrong in the neural network and saves a targeted change. The result is a small patch file containing only the changed values and checksums. It can be used to modify a public base model.

Cmpnd says the method works best with decision models. A decision model reads a situation description and typed questions, then assigns a probability to each answer option. It does not generate text.

Results in Cmpnd’s own tests

The table shows how much the accuracy of five decision models rose on 400 cases from the Typed Decisions benchmark. These cases were not used in optimisation. ReAnchor is an experimental DSPy optimiser that adjusts answer thresholds without touching the weights.

Model Baseline ReAnchor WeightSpace Both
d1-omni-600M 50.9% +9.3 +16.1 +18.3
NanoJev 39.2% +11.6 +13.3 +17.3
d1-2B 60.6% +9.5 +13.4 +17.3
Strands Decider 2B – +13.7 +10.4 +20.0
Clef-Flash (9B) 68.8% +1.5 +1.9 +6.5

Change in percentage points compared with the unedited model. The metric is the accuracy of the answers the program returns. The writer was GPT-6-Astra, which could run each model 25 times during its optimisation. Standard error is 0.7–1.4 points. Source: Cmpnd; NanoJev’s baseline and its ReAnchor and combined results from MarkTechPost, which cites Cmpnd’s results report.

The results do not support the conclusion that editing weights is always the best approach. On Strands Decider 2B, ReAnchor alone improved the result more. MarkTechPost’s claim that WeightSpace beat ReAnchor on every model applies only to the four models in its table.

The choice of writer model also affected the result. Edits by GPT-6.1-Sol changed accuracy by +5.8 percentage points on d1-omni-600M, +12.1 on NanoJev and −0.6 on d1-2B.

By default, the writer may compute the decision layer’s outputs itself in the sandbox, which the original study did not allow. Under the study’s rules, d1-omni-600M’s accuracy rose by only 2.8 ± 0.7 percentage points, and the edit did not improve on ReAnchor’s result.

Costs and privacy

According to the repository’s documentation, optimising a decision model with 25 runs took 68–167 minutes, and the writer model’s calls cost $5.42–10.78. MarkTechPost’s figures (78–167 minutes, $5.42–7.63) are based on three models.

The sandbox stops the writer’s code from using the network or reaching the home directory. The writer model’s provider still sees the model weights and training examples as the writer reads them.

dspy-wsp is on GitHub under the MIT licence and requires Python 3.11 or later, DSPy 3.5 and an API key for the writer model. The d1-omni-600M model weights can be optimised on a Mac or on a GPU with at least 4 GB of memory; installation instructions are in the repository.

What this is based on

  1. github.com

    Publisher: github.com · source publication time unknown · retrieved 15 November 2026 at 05:38 (UTC+02:00) · discovery channel: GitHub

  2. Cmpnd Releases dspy-wsp: Weight-Space Programming Comes to DSPy

    Publisher: www.marktechpost.com · published 14 November 2026 at 18:45 (UTC+02:00) · retrieved 15 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

  3. Weight-Space Programming with DSPy — Cmpnd

    Publisher: www.cmpnd.ai · published 13 November 2026 · retrieved 15 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

  4. dspy-wsp: редактирование весов LLM без градиентов

    Publisher: aikraft.ru · published 14 November 2026 at 21:18 (UTC+02:00) · retrieved 15 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

  5. Weight-Space Programming with DSPy — Cmpnd | TLDR

    Publisher: tldr.tech · published 13 November 2026 · retrieved 15 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Models

World Labs introduces Atlas, a world model that turns images into 3D worlds

AI-generated contentWritten by AI

worldlabs.ai · Source publication date unknown

World Labs says its Atlas model turns photographs or AI-generated images into 3D worlds that users can move a camera through. The model comes with editing tools, conversion into meshes used in games, and an API that connects Atlas to tools such as Blender.

World Labs presents its Atlas model on its website, and the TLDR AI newsletter highlighted the page on 13 November. The company calls Atlas a next-generation world model. A world model creates three-dimensional environments or reconstructs them from images. The description is the company’s own, and the product page offers no test results or comparisons with other models to support it.

World Labs is also releasing tools alongside Atlas. Viewfinder builds a scene from images, and users can view it from new angles, plan camera moves and render them as videos.

In Chisel, users first sketch the structure of a scene with simple shapes, a technique known as grayboxing. Every camera view updates when users move objects, the company says.

A splat representation builds a 3D scene from a large number of small, partly translucent points. Games and 3D programs usually use meshes made of polygons. Splats to Mesh converts splats into meshes with real geometry. World Labs says the converted worlds can be used in games, simulations and the 3D tools users already work with. The Relight tool is for lighting and editing splats.

Through an API and integrations, users can connect Atlas capabilities to their own applications and existing tools such as the 3D program Blender, the company says.

World Labs is marketing Atlas for film and visual effects production, architecture and design, and robotics. In robotics, the company says, the model can build research and simulation environments from real or AI-generated scenes.

Users can try Viewfinder on World Labs’ website by uploading their own photographs or generating images. The page gives no price but links to the API documentation and further details on how the model works.

What this is based on

  1. Atlas

    Publisher: worldlabs.ai · source publication time unknown · retrieved 15 November 2026 at 05:40 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Agents

Grok Bot team says it shipped 11,000 pull requests to production in a month

Written by AI

youtube.com + 1 source · 14 November 2026

AI-generated contentWritten by AI

youtube.com, x.com · 2 sources · Sources published: 14 November 2026

Lauren Tan said SpaceXSI’s Grok Bot team maintains the quality of coding agents’ code by turning corrections given to agents into automated checks. About 0.5 per cent of pull requests are reverted, Tan said.

Lauren Tan (poteto) described in a talk at the AI Engineer conference how SpaceXSI’s Grok Bot team keeps code quality high enough that, by Tan’s account, most product code no longer needs thorough human review. The video of the talk was published on 14 November. Tan described Grok Bot as an Electron-based desktop app for launching and steering Cursor’s cloud agents.

Corrections become automated checks

Tan starts from the premise that an agent also copies workarounds and outdated solutions from the codebase. When the team corrects the agent, the correction does not become a new line in an instructions file such as agents.md. Instead, the team turns it into an automated check or improves the interface or architecture, because a check in code works repeatably, Tan said.

Dune, the team’s own framework for Electron apps, includes lint rules: checks that reject code using a banned pattern. React’s useEffect, which Tan said agents use poorly, is banned in product code. Code for Electron’s main process and renderer process sits in separate directories, and neither may import the other’s code.

The team has also banned 99 per cent of code comments. Tan gave an example from Cursor: agents used comments to justify quick fixes built on timers, which degraded the app’s performance. Now the lint rule’s error message directs the agent to fix the root cause.

Tan advises starting with a lint rule that blocks new bad patterns, then cleaning up the old ones. Automating a way of working that produces bad code only produces more bad code, Tan said. It took Tan six months to reach the point of being able to ship thousands of pull requests to production.

Agents verify their own work

Review has shrunk to looking at videos, screenshots and diagrams, Tan said. The agent runs the app itself with debugging tools such as the Chrome DevTools Protocol. A feature map lists the app’s functions for the agent from the user’s point of view.

The team is also experimenting with formal verification, the mathematical checking of a program’s properties. The team checks algorithms with Lean and key user flows, such as notifications across multiple devices, with TLA+. Tan said the method has already found several bugs but is no silver bullet, and the experiment continues. Most of Tan’s tokens go on verification.

The results are the team’s own figures

Last month the team shipped 11,000 pull requests, or proposed code changes, to production, Tan said. In October, about 227 people contributed at least one pull request. A new stable version is released about once a day.

About 0.5 per cent of pull requests are consistently reverted, a lower share than before, Tan said. The talk’s transcript does not say which period the share is compared with, and the figures have not been independently verified. Tan said that since launch the app has become faster at starting up, typing, and opening and switching conversations. The transcript gives no measurements.

After each new version is released, Grok Bot starts a Cursor project that uses the app as a user would. The project fixes the bugs it finds and merges the fixes automatically. The codebase is in such good shape that pull requests are usually good from the start, Tan said.

Tan’s pstack plugin is available on GitHub in Cursor’s plugins repository. It includes a State Chart skill that draws state machines from the user’s point of view.

What this is based on

  1. www.youtube.com

    Publisher: www.youtube.com · published 14 November 2026 at 20:30 (UTC+02:00) · retrieved 15 November 2026 at 05:16 (UTC+02:00) · discovery channel: YouTube

  2. x.com

    lauren · Publisher: x.com · published 14 November 2026 · retrieved 15 November 2026 at 05:16 (UTC+02:00) · discovery channel: X

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Agents

FBI warns about ARTEX: some intrusions reached privileged access in under three hours

Written by AI

ic3.gov + 6 sources · 14 November 2026

AI-generated contentWritten by AI

ic3.gov, cybersecuritynews.com, glitchwire.com, casrai.org, therecord.media, gbhackers.com, cyberpress.org · 7 sources · Sources published: 13–14 November 2026

Attackers are using ARTEX, an open-source program in which AI agents powered by large language models carry out several stages of an intrusion, from reconnaissance to data exfiltration, the FBI said. The FBI urges organisations to map forgotten internet-facing systems too.

On 13 November the FBI issued a FLASH alert that brings together security firms’ observations of ARTEX misuse. Since at least August 2026, attackers have used the program against targets including the financial sector, government and healthcare, the alert says.

Agents handle several stages of an intrusion

ARTEX is a harness: a program that controls agents, gives a model its tools and runs its work steps. The user writes a goal in natural language, and ARTEX turns it into commands. Several AI agents then execute the commands with almost no human direction.

ARTEX has no model of its own; it uses external large language models such as Claude, GPT and DeepSeek. DeepSeek’s Flash models are the most common in attackers’ installations, the FBI said. The FBI attributes this to their combination of capability and relatively low cost, and to the lack of observable guardrails when a request is framed as penetration testing, an authorised attempt to break in.

The agents have exploited known weaknesses such as SQL injection, APIs that require no authentication, and stolen VPN and service provider credentials. Once inside, they have installed web shells – files placed on a server for remote control – and added credentials to existing OAuth applications. ARTEX is particularly good at finding overlooked systems such as partner portals and mobile app back ends, the FBI said.

In some intrusions, less than three hours passed between initial targeting and privileged access. Cyber Security News notes that the figure describes some of the observed intrusions, not ARTEX’s capability in general. ARTEX traffic stands out and can trigger alerts, the FBI said, but its speed leaves defenders little time.

Victims, targets and spread

In attacks on South Korean financial institutions, attackers exfiltrated the personal data of about 67,000 people, the FBI said. The alert says a supply chain attack through the ConnectWise Cloud Backup service enabled access to more than 2,250 customer environments. Cyber Security News describes the same Microsoft finding differently: the attackers tried to gain access to the environments, and the figure counts targets, not confirmed breaches.

In early November, Censys researchers found more than 1,600 publicly exposed ARTEX management consoles, and the FBI estimates that there are significantly more installations. A visible installation does not by itself prove malicious use, Cyber Security News said.

ARTEX was published on GitHub by a developer called Autumn-27, whom The Record identifies as Chinese cybersecurity engineer Li Puhua. The FBI dates the release to July, The Record to August. The developer said the tool was intended for authorised testing and deleted the repository on 8 October. The program is still available through copies and as a commercial version, ARTEX PRO.

FBI recommendations for defenders

The FBI urges organisations to map edge systems as well, restrict cloud management and deployment APIs to trusted networks, reduce application permissions and search repositories for secrets. Unexpected connections to language model APIs are worth investigating alongside other suspicious signs. IP addresses in the alert should be checked before blocking, because cloud addresses can be reassigned for harmless use.

The FBI advises users of ConnectWise applications to review every Microsoft 365 environment where the application is or has been present. They should search by the application ID and review activity from at least 1 October. The connections should stay disabled until ConnectWise issues recovery instructions, the FBI said. Applications should not be removed until settings and evidence have been preserved.

What this is based on

  1. www.ic3.gov

    Publisher: www.ic3.gov · published 13 November 2026 · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

  2. FBI Reveals How Hackers Use ARTEX AI Agents to Gain Privileged Access in Under Three Hours

    Publisher: cybersecuritynews.com · published 14 November 2026 at 13:21 (UTC+02:00) · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

  3. FBI Issues FLASH Alert on ARTEX, the AI Agent Framework Automating Network Intrusions — Glitchwire

    Publisher: glitchwire.com · published 14 November 2026 at 16:55 (UTC+02:00) · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

  4. FBI Warns of ARTEX AI Attack Framework — CASRAI

    Publisher: casrai.org · published 14 November 2026 at 13:42 (UTC+02:00) · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

  5. FBI flash alert says Artex tool automating cyberattacks at ‘machine speed’

    Publisher: therecord.media · published 13 November 2026 at 23:35 (UTC+02:00) · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

  6. FBI Warns Hackers Are Weaponizing ARTEX AI Framework to Launch Autonomous Cyberattacks Worldwide

    Publisher: gbhackers.com · published 14 November 2026 at 09:07 (UTC+02:00) · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

  7. FBI Warns Hackers Are Using ARTEX AI Agents to Breach Networks in Under 3 Hours

    Publisher: cyberpress.org · published 14 November 2026 at 07:35 (UTC+02:00) · retrieved 15 November 2026 at 05:19 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Agents

OpenClaw developers say new beta eases memory load in long sessions

Written by AI

github.com + 3 sources · 15 November 2026

AI-generated contentWritten by AI

github.com, wapbaike.baidu.com, senx.ai, releasebot.io · 4 sources · Sources published: 12–15 November 2026 · Some source publication dates are unknown

OpenClaw’s developers say version 2026.11.5-beta.3 reduces memory load when the program indexes, trims and compacts long conversations. Users can now set a character limit for summaries, and the program redacts credentials quoted in them.

The OpenClaw project published pre-release version 2026.11.5-beta.3 of its AI agent on GitHub on 15 November. The release notes say it brings together 50 pull requests since the previous beta. OpenClaw is an open-source AI agent that can run commands on the user’s computer. Users control it through messaging apps.

Long sessions and compaction

In compaction, the program replaces the older part of a conversation with a summary so that the conversation fits in the model’s context window. The beta reduces the memory load of indexing and trimming large transcripts, the release notes say. When a session’s state is reloaded, the program transfers only selected parts. The notes do not say how much memory use falls.

Users can now set a limit on the number of characters in a summary. Budget limits for the request in progress still apply, the notes say, but they do not name the setting.

Redacting credentials quoted in summaries is a security change. The news site SEN-X reported it as soon as it reached the development branch. Previously, the site said, a value beginning token=, sk- or ghp_ or a PEM private key pasted into the conversation ended up in the summary verbatim. Because the program sends the summary to the model with every later request, the credential stayed in the model’s context for the whole session.

The fix does not apply to the stored transcript. SEN-X said the raw text is kept there deliberately. The site advises against pasting live keys into the agent’s conversation and urges users to replace keys they have already pasted.

Restarts and permissions

Queued messages keep their original permission ceilings when the program restarts, the notes say. The program finishes the conversation’s final writes before work passes to a new process. It keeps responsibility for delivering the final message until the last RPC completes.

Delegated changes now require a valid authorisation. The iOS app removes old setup codes stored in plain text and keeps pasted codes only in memory. Old, conflicting command-approval files no longer block start-up; the primary policy decides.

Other changes cover areas including cold start, recovery of interrupted workspaces, MCP tools, the Slack, iMessage and IRC channels, the Z.ai and Gemini models, and Windows and macOS.

The version is available from the npm package manager as 2026.11.5-beta.3. SEN-X noted that the previous beta did not change the version for users on the default channel. When the site checked on 12 November, the default channel was on 2026.11.3. The release checklist says the result of the Telegram beta end-to-end test has not been submitted.

What this is based on

  1. openclaw 2026.11.5-beta.3

    openclaw · Publisher: github.com · published 15 November 2026 at 02:31 (UTC+02:00) · retrieved 15 November 2026 at 04:15 (UTC+02:00) · discovery channel: GitHub

  2. OpenClaw

    Publisher: wapbaike.baidu.com · source publication time unknown · retrieved 15 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

  3. OpenClaw 2026.11.5 Beta 2 Reaches npm as Claude Fable 5.2 Goes Official and Snyk Re-Scans 10,000 ClawHub Skills

    Publisher: senx.ai · published 12 November 2026 at 16:30 (UTC+02:00) · retrieved 15 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

  4. OpenClaw Release Notes - November 2026 Latest Updates - Releasebot

    Publisher: releasebot.io · published 13 November 2026 · retrieved 15 November 2026 at 05:30 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Agents

AI Rudder’s CALL-E lets AI agents make phone calls with user approval

Written by AI

technode.global + 3 sources · 15 November 2026

AI-generated contentWritten by AI

technode.global, indovizka.com, indiantelevision.com, voi.id · 4 sources · Sources published: 14–15 November 2026

Singapore-based AI Rudder says its CALL-E service places calls through ChatGPT, OpenClaw and Claude Code, among others, only after the user has approved a call plan. After the call, the service gives a summary and lines from the conversation that support the outcome.

AI Rudder announced CALL-E on 13 November in a press release distributed by Media OutReach Newswire. TechNode Global, Indiantelevision.com and Indonesia’s Voi.id base their reports on the release and do not independently verify the company’s claims.

The company says CALL-E works in ChatGPT, in personal AI agents such as Meta’s Muse, OpenClaw and Hermes Agent, and in the coding agents Claude Code, Codex and Cursor. TechNode describes this as compatibility stated by AI Rudder. The company announces no partnership with the services’ operators and no integration approved by them.

How a call works

The user’s agent hands CALL-E a task, such as moving a restaurant booking. The service first shows a call plan. The call goes out only once the user has approved that specific plan.

After the call, CALL-E reports what was agreed. It also rates how confident it is that the goal was met and shows the lines that support the outcome. TechNode says a transcript is included as well.

Developers can use the API to set the call’s goal, the phone number and the desired output format. TechNode points out that when the API accepts a request, the request has only entered the service’s workflow. Acceptance does not yet show whether anyone answered or whether the task succeeded.

By default, CALL-E introduces itself as a voice agent. Users can change the introduction, but the company says a pre-check is intended to stop it from, for example, posing as the FBI or the emergency number 911.

The company’s claims and legal requirements

CEO Teng Ren points to turn-taking on calls to make the case for the product: ‘The smartest AI model in the world can still fail a phone call. All it takes is one “mm-hmm”.’ Ren says an agent can mistake an affirmative murmur for a question and fall silent mid-sentence.

The release says Helm Swift, the model that detects interruptions, has the joint-lowest false-positive rate of any qualifying model on the October 2026 TurnBench Interruption leaderboard. In practice, the release says, the model rarely stops speaking just because the other party says ‘yeah’. TechNode notes that the company reported the result and that TechNode did not reproduce it. Voi.id points out that a single test result does not guarantee CALL-E will always understand correctly.

CALL-E uses the same technology as AI Rudder’s enterprise solutions, which are designed to meet the SOC 2 Type 2 and ISO 27001 standards and the requirements of the EU’s General Data Protection Regulation and South-east Asia’s personal data protection laws (PDPA), the release says. Indiantelevision goes further, writing that the infrastructure meets them. Each call leaves a record of the plan, the approval and the outcome, the release adds.

The release cites Article 50 of the EU AI Act and a 2024 FCC ruling (FCC 24-17) that AI-generated voices fall under the US Telephone Consumer Protection Act. TechNode notes that rules on consent, disclosure and recording vary by country and use case. Voi.id writes that the service user’s consent does not replace the call recipient’s consent where the law requires it.

CALL-E is available at heycall-e.com and through an API. New users receive 100 free credits. After that, TechNode says, pricing depends on call duration and telecommunications costs.

What this is based on

  1. AI Rudder launches CALL-E voice agent for phone calls

    Publisher: technode.global · published 14 November 2026 at 17:11 (UTC+02:00) · retrieved 15 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

  2. CALL-E Lets AI Agents in ChatGPT, Muse and OpenClaw Make Real Phone Calls on Their Users' Behalf

    Publisher: indovizka.com · published 15 November 2026 at 02:35 (UTC+02:00) · retrieved 15 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

  3. AI Rudder launches CALL-E to let ChatGPT agents make phone calls for users

    Publisher: indiantelevision.com · published 14 November 2026 at 05:17 (UTC+02:00) · retrieved 15 November 2026 at 05:34 (UTC+02:00) · discovery channel: web

  4. CALL-E Makes AI Agents Able to Call Restaurants, Must Ask for User Permission

    Publisher: voi.id · published 14 November 2026 · retrieved 15 November 2026 at 05:35 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

China

IT Home: DeepSeek’s web code points to an unannounced Coding Plan

Written by AI

pandaily.com + 3 sources · 15 November 2026

AI-generated contentWritten by AI

pandaily.com, m.ithome.com, einkcn.com, todayforai.com · 4 sources · Sources published: 14–15 November 2026

Developers found usage-limit management in the code of DeepSeek’s web chat, and IT Home links it to a plan called Coding Plan. DeepSeek has not announced a subscription, price or release date.

Chinese technology site IT Home reported the find on 14 November, based on tips from readers. Pandaily, which publishes in English, summarised the report the following day. The information comes from unnamed developers who examined the browser code of chat.deepseek.com.

A usage limit caps how much of a service someone can use within a set period. Once the limit runs out, use stops until the limit is restored.

The developers said the code contains a notice shown to users: ‘The usage limit has been reached, please try again after it has been restored.’ The code also handles the error that occurs when the limit has run out. When a user continues a shared conversation, the code checks the quotas table to see whether the limit has been used up (is_exhausted).

How users would see the limits

The latest commit to the interface adds regular checks on limit status, the developers said. While the page is visible, it polls the server every five minutes for each model’s usage limit, any messaging ban on the user and text sent by the server.

When the limit runs out, a notice bar would appear above the input field, and the welcome area on the home page would show limit information sent by the server. The system also includes a notice that the limit has been restored and warnings about usage restrictions set by the server. IT Home said the system points to a plan called Coding Plan.

No details on price or audience

DeepSeek has not announced the subscription’s price, tiers or release date. The company has not commented on the code, Pandaily reported, and the reports do not mention a payment page. Nor do they make clear whether the limits would apply to free users, paying users or both. Pandaily noted that web app code can also contain features under test, which may never be released.

Some users speculate that DeepSeek will launch its first paid consumer plan alongside the DeepSeek V4.1 Pro model. IT Home stressed that this is community speculation.

Until now, DeepSeek has charged individuals directly only through the API on its developer platform, where pricing depends on the tokens used. IT Home reported that DeepSeek had around 130 million monthly users in October, and that almost all of its more than $1 billion in annualised revenue came from the API. IT Home does not name a source for the figures.

Pandaily said the Coding Plan name would put DeepSeek in the same product category as other Chinese model providers. It cited Zhipu AI’s GLM Coding Plan and coding subscriptions from Alibaba Cloud and Baidu AI Cloud as examples.

What this is based on

  1. DeepSeek's Web Code Shows Quota Management Pointing to a 'Coding Plan', but No Plan Is Announced

    Publisher: pandaily.com · published 15 November 2026 at 05:03 (UTC+02:00) · retrieved 15 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

  2. DeepSeek 前端代码出现“额度耗尽”提示,开发者猜测将首次推出 C 端付费方案 - IT之家

    Publisher: m.ithome.com · published 14 November 2026 · retrieved 15 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

  3. DeepSeek 前端代码出现“额度耗尽”提示 开发者猜测将首次推出 C 端付费方案 - einkCN | ePaper电子纸与 eink行业新闻动态

    Publisher: einkcn.com · published 14 November 2026 at 05:09 (UTC+02:00) · retrieved 15 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

  4. DeepSeek Frontend Code Leak Hints at First Consumer Coding Plan

    Publisher: todayforai.com · published 14 November 2026 at 03:20 (UTC+02:00) · retrieved 15 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Research

Preprint: security rules written by small models failed a prompt-injection test

Written by AI

arXiv + 2 sources · 14 November 2026

AI-generated contentWritten by AI

arXiv, aisafetyhot.com, dev.to · 3 sources · Sources published: 12–14 November 2026

In a test, security rules that 15 open-weight models wrote themselves let prompt injections through or blocked tasks. CapLite blocked every attack in the test that tried to make the agent change something, but the agent completed clearly fewer tasks than it did without a defence.

On 12 November, researchers from the University of Tokyo and the Institute of Automation of the Chinese Academy of Sciences published a preprint on arXiv that has not been peer-reviewed. In it, the researchers examine whether a small language model can write its own rules to protect an agent against prompt injection. In a prompt injection, an attacker hides an instruction in an email, file or web page that the agent reads as the result of a tool call.

In the CaMeL, Progent, IPIGuard and DRIFT defences, code enforces the rules, but the agent’s own model writes them. Depending on the defence, the rules take the form of a program, an access policy, a dependency graph or a list of constraints. A defence is only as good as its rules, the authors wrote.

Errors let attacks through or halt work

The researchers tested the defences on the AgentDojo benchmark with 15 models, 13 of which have between 1.7 billion and 12 billion parameters. Progent’s policies let attacks through on 14 models, IPIGuard’s graphs on 13 and DRIFT’s constraints on 11. CaMeL let only 20 attacks through in 13,935 attack runs, but at worst, with Granite-4.2-8B, 68 per cent of runs produced no working program.

In the Qwen3 and Gemma-4 model families, the share of attacks that got past Progent and IPIGuard rose with model size, as it did for the undefended agent. A larger model is easier to steer with an injection, and the rules it writes do not make up for this, the authors said. A dev.to post by Mech.app says models above 12 billion parameters write better rules, but the preprint’s figures do not support this.

CapLite: the model proposes, code decides

The researchers’ CapLite uses fixed rules to infer from tool names and descriptions which tools only read and which change something, for example by sending or paying. The planning model sees only the user’s request and the tool descriptions. It locks the state-changing calls before the agent reads the result of any tool.

Before each state-changing call, code checks that the recipient, amount, target and URL exactly match the values given in the user’s request. Text hidden by an attacker cannot add a call or redirect one to a new target. The check does not restrict message content, dates or times. Nor does CapLite block an injection that only tries to get a particular sentence into the agent’s response.

The table shows averages across the 15 models for AgentDojo’s main attack (important_instructions); successful attacks are totals.

Defence Tasks completed under attack Successful attacks Tokens per run (thousands)
No defence 50.2% 2,322 28.0
Progent 40.6% 676 63.4
IPIGuard 35.7% 435 60.5
DRIFT 34.9% 214 28.1
CaMeL 27.2% 20 93.7
CapLite 29.0% 0 17.1

Source: preprint, AgentDojo v1.2.2; 949 attack pairs per model, of which 929 were used to count successful attacks.

The table does not show that CapLite beats CaMeL on tasks: with Gemma-4-E4B and -12B, CaMeL succeeded significantly more often.

Under attack, the agent protected by CapLite completed tasks 21 percentage points less often than the undefended agent. The main cause is the small models’ weak planning, the authors said: in 43 per cent of the runs that failed without an attack, the locked plan lacked a necessary call. If a recipient appears only in a tool result, CapLite asks the user for confirmation. In the test, nobody answered these requests, so those runs were counted as failures.

All results come from AgentDojo’s four task suites, and each setup was run once. The authors say they will release the code and per-run results.

What this is based on

  1. CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts

    Xu, Zimo, Liu, Wenzhuo, Nakayama, Hideki · Publisher: arxiv.org · published 12 November 2026 at 19:07 (UTC+02:00) · retrieved 15 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

  2. CapLite: Prompt-Injection Defense for Small Language Model Agents without Model-Written Security Artifacts

    Publisher: aisafetyhot.com · published 12 November 2026 at 19:07 (UTC+02:00) · retrieved 15 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

  3. CapLite: Why Small Language Model Agents Can't Write Their Own Security Policies

    Publisher: dev.to · published 14 November 2026 at 12:06 (UTC+02:00) · retrieved 15 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Research

GPT models found the most agent risks in a benchmark when false alarms were not capped

Written by AI

arXiv + 1 source · 12 November 2026

AI-generated contentWritten by AI

arXiv, scholar-inbox.com · 2 sources · Sources published: 12 November 2026

In ReA-Bench, a benchmark from Alibaba researchers, GPT models found the most risks in AI agent trajectories, but they also raised alarms on more than half of the difficult benign trajectories. When the false-alarm rate was capped, Grok and DeepSeek models found more risks.

Researchers from Alibaba’s AAIG group published a preprint on arXiv on 12 November describing ReA-Bench, a benchmark that measures how language models audit the trajectories of AI agents. The auditing model reads the agent’s steps and judges whether the agent acted riskily. The steps include, for example, tool calls and their responses. The study has not been peer-reviewed.

The usual metric is the share of risks found (recall). The authors said it does not show which risks go unnoticed, nor whether higher recall brings more false alarms. In their analysis of production and red-team trajectories, some risks were repeatedly missed across different auditing configurations.

Six tiers and difficult benign trajectories

The benchmark contains 587 trajectories: 304 risky, 177 benign and 106 control versions. Most were taken from real production runs or adapted from them. The tasks fall into six tiers:

  • security knowledge
  • contradictions within a trajectory
  • finding evidence in a long trajectory
  • checking sources and reasoning
  • gradual steering over several turns
  • missing information.

The benign trajectories were deliberately chosen to be difficult. In them, the agent does things that look risky, but the task and the permissions the user granted justify them. The authors said these trajectories do not reflect ordinary production runs, so the false-alarm rate measured in the benchmark cannot be converted directly into production figures.

The models that found the most risks often raised false alarms

The researchers tested eight model configurations with four common auditing methods. The first two numeric columns show the average across the four methods at an alarm threshold of 5 on a 0–10 scale. The last column shows the best recall when no more than 10 per cent of benign trajectories are flagged.

Model Recall False alarms Recall, false alarms ≤ 10%
DeepSeek-Flash 59.7% 3.8% 67.8%
DeepSeek-Pro 50.6% 3.8% 71.1%
Qwen-Flash 63.7% 8.6% 68.5%
Qwen-max 73.6% 12.9% 71.8%
Grok-low 65.4% 5.4% 74.6%
Grok-high 71.0% 5.5% 78.0%
GPT-luna 87.0% 52.0% 50.5%
GPT-sol 88.5% 54.0% 47.3%

Source: ReA-Bench preprint, tiers 1–5. The false-alarm rate was measured on the benchmark’s difficult benign trajectories.

At the fixed threshold, the GPT models found the most risks. When false alarms were capped, they fell behind the others. Each configuration was run only once, so small differences need to be confirmed in repeat runs, the authors said.

The models were generally better at identifying risks that required security knowledge than at spotting contradictions, unsupported claims and reasoning errors. Shortening the trajectory did not always help. In one trajectory, the agent had used UTC timestamps although the reporting period was defined in Beijing time. The models audited the trajectory 28 times with different configurations and auditing methods. They found the error 12 times in the original trajectory of about 220,000 characters and 13 times in a version shortened to about 13,000 characters.

The authors also tested GSA, their own note-compiling method. At the fixed threshold, GSA found more risks than the highest-recall baseline in all four configurations, but it also raised more false alarms. Across all thresholds, GSA discriminated risks better in three configurations but worse with GPT-luna.

The authors plan to release the benchmark openly once identifying information has been removed from the data and the release has been reviewed.

What this is based on

  1. ReA-Bench: Benchmarking Responsible Agent Auditing

    Ye, Zipeng, Dai, Shenyu, Xie, Yong, Wang, Kun, Guan, Jiyang, Chen, Zhikai, Weng, Jiaqi, Tao, Jialing, Xue, Hui · Publisher: arxiv.org · published 12 November 2026 at 11:01 (UTC+02:00) · retrieved 15 November 2026 at 05:28 (UTC+02:00) · discovery channel: web

  2. ReA-Bench: Benchmarking Responsible Agent Auditing

    Publisher: www.scholar-inbox.com · published 12 November 2026 · retrieved 15 November 2026 at 05:29 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Research

Preprint: an instruction hidden in an image can become visible to a model when the image is heavily downscaled

Written by AI

arXiv + 1 source · 12 November 2026

AI-generated contentWritten by AI

arXiv, huggingface.co · 2 sources · Sources published: 12 November 2026 · Some source publication dates are unknown

According to a preprint, a text instruction can be hidden in an image so that four closed vision-language models read it in more than half of tests, but only once the image had been downscaled to 64 × 64 pixels. With milder downscaling, the effect weakened sharply.

Researchers at Nanyang Technological University and DSO National Laboratories describe a phenomenon they call Scale-Induced Semantic Shift (SISS) in a preprint published on arXiv on 12 November. The study has not been peer reviewed.

Systems often downscale or compress an image before a vision-language model processes it. The model may therefore not see the same image as the user.

Tests and results

The researchers fine-tuned an image generator on images they selected solely on the basis of feedback from the open-weight Qwen2.5-VL model. The test images were frozen before the target models saw them. An attack succeeded if a model did not read the instruction in the 512 × 512 pixel image but did read it in the version downscaled to 64 × 64 pixels. Gemini 3 Pro acted as judge and decided from the text response alone whether the instruction had been found. It did not see the image.

Attack success rate on 300 test images:

Model SISS QR-Monster (baseline)
GPT-4o 60.33% 35.33%
Claude Sonnet 4.6 57.67% 49.33%
Qwen3.5-Plus-02-15 62.67% 47.67%
Gemini-3-Flash 54.33% 32.00%
Average 58.75% 41.08%

Source: authors’ preprint, third refinement round, downscaling from 512 to 64 pixels.

The figures do not show that the attack works at ordinary image sizes. The authors said 64 pixels was a deliberately aggressive test setting, not an assumption about how deployed models work. With the same images, the average success rate was 17.68 per cent at 128 pixels, 6.85 per cent at 224 pixels and 5.60 per cent at 336 pixels.

Small effect on agents

In a web-agent experiment, the images increased detection of the hidden instruction. For Gemini-3-Flash, the rate rose from 37.33 per cent with control images to 67.67 per cent, and for GPT-4o from 12.67 per cent to 30.33 per cent. However, the agents took harmful actions in only 4.00 and 3.67 per cent of cases respectively, compared with 2.33 and 1.67 per cent with the control images. The authors said the experiment shows that images can influence an agent, but not that they can reliably hijack one. The experiment involved only two agents.

Generic image filters offered no protection

In the defence experiment, rescaling, JPEG compression, filtering and random masking produced no statistically significant change in the success rate. The authors’ ScaleGuard detector compares the full-size image with nine downscaled versions and rejects the image if text appears only in the small ones. It cut the success rate from 61.4 per cent to 24.3 per cent. When the attacker refined images for one round against the detector, the rate fell from 42.5 per cent to 15.0 per cent.

The detector rejected none of the clean control images in samples of 70 and 40 images. The authors said samples this small do not show that false alarms are rare. Nor did they measure the detector’s latency, even though it checks ten versions of every image. The defence experiment used a different dataset and judge and targeted only the Qwen2.5-VL-7B model, so its figures cannot be compared with the main results.

What this is based on

  1. Scale-Induced Semantic Shift in Visual Prompt Injection Against VLMs and Web Agents

    Zhao, Yiheng, Ji, Runze, Yi, Chenyu, Yu, Yi, Shen, Bingquan, Kot, Alex, Jiang, Xudong · Publisher: arxiv.org · published 12 November 2026 at 14:57 (UTC+02:00) · retrieved 15 November 2026 at 05:31 (UTC+02:00) · discovery channel: web

  2. 2504/2504.14348.md · huggingchat/papers-content

    Publisher: huggingface.co · source publication time unknown · retrieved 15 November 2026 at 05:32 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Retrieval and RAG

OpenRouter test: Voyage model raised top-result accuracy from 81 to 97 per cent

Written by AI

github.com + 2 sources · 13 November 2026

AI-generated contentWritten by AI

github.com, openrouter.ai · 3 sources · Sources published: 13 November 2026 · Some source publication dates are unknown

In OpenRouter’s own test on a small synthetic dataset, vector search alone ranked the correct text chunk first for 81 per cent of questions. When voyageai/rerank-3-lite reranked 30 candidates, the share rose to 97 per cent for about $0.10 per 1,000 queries.

OpenRouter published a comparison of reranking in RAG applications on its blog on 13 November. A RAG application retrieves text chunks from an organisation’s own documents and passes them to a large language model as the basis for an answer. The company compared 11 rerankers, along with their latency and price.

Why vector search gets it wrong

Vector search condenses each chunk into a single vector before the question is known. In a 130-word chunk, the name of a service tier takes up only a word or two, OpenRouter said, so chunks describing the limits of different tiers look almost identical to the search. When the test asked about the Growth tier’s bucket limit, google/gemini-embedding-2 ranked the Starter, Enterprise and Scale chunks highest. The correct chunk came only fifth.

A reranker reads the question together with each candidate and rescores the candidates. It is too slow to run over the whole dataset, so vector search narrows down the candidates first. OpenRouter recommends retrieving about 30 candidates and passing the top three to the large language model.

The test dataset contained 274 model-generated chunks and questions about Fernbank, a fictional storage service. Each question had one correct chunk, so results were scored automatically. The company warns that the dataset is small and synthetic, and that the benefit may differ on real documents.

Results and prices

All models reranked the same 30 candidates retrieved by gemini-embedding-2. The table shows how often the correct chunk came first (precision@1).

Model Correct first Median latency (s) Price / 1,000 queries
No reranking 81.0% – –
voyageai/rerank-3 97.8% 1.27 $0.26
voyageai/rerank-3-lite 97.1% 1.26 $0.10
cohere/rerank-4-pro 95.3% 1.54 $2.50
qwen/qwen3-reranker-8b 92.3% 1.51 $1.49
cohere/rerank-v3.5 90.5% 1.47 $1.00
openai/gpt-6-luna (large language model) 99.6% 4.01 $0.70
google/gemini-3.8-flash (large language model) 100% 3.25 $4.14

OpenRouter run on 8 October 2026, 274 questions, selection of 11 models. About 0.9 s of the latency is network latency from the test machine.

The two Voyage models differ by only two questions, and OpenRouter said this does not show that the more expensive model is better. The large language models read all 30 chunks at once and were the most accurate. Their median latency was 3.3–4.4 seconds, against about 1.3 for the rerankers. The company said it is not certain that the advantage holds on messier data.

Given three reranked chunks, anthropic/claude-haiku-5.5 answered 99.3 per cent of questions correctly. With three chunks and no reranking, the figure was 95.6 per cent. Compared with five chunks without reranking, total cost rose by a net $0.04 per 1,000 questions. Reranking cost $0.10, but the large language model’s answering cost fell from $0.22 to $0.16 because it received three chunks instead of five.

Limits

Reranking can only promote a candidate that vector search has found. When voyageai/voyage-4-lite retrieved 50 candidates instead of 30, the result improved by one question. A chunk that matches the topic is not always the right one either: in the blog’s six-chunk example, an unanswered forum question rose to second place.

cohere/rerank-v3.5 reads only the first 4,096 tokens of a document and gives no warning of the truncation. When the embedding model call omitted the input_type parameter, results fell by 7–12 percentage points.

Rerankers are used through OpenRouter’s /rerank endpoint; large language models do not work with it. Voyage and Qwen charge per token, Cohere per search. The example code in the documentation’s RAG guide differs from the test: it uses cohere/rerank-v3.5, does not set the input_type parameter and recommends 10–20 candidates.

What this is based on

  1. github.com

    Publisher: github.com · source publication time unknown · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: GitHub

  2. Reranking in RAG: How to Fix Bad Retrieval with a Reranker

    Publisher: openrouter.ai · published 13 November 2026 at 02:00 (UTC+02:00) · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

  3. RAG with Embeddings & Rerank - Build Retrieval-Augmented Generation Pipelines

    Publisher: openrouter.ai · source publication time unknown · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Finance and law

Accounting firm WCG uses AI to review tax returns, not to prepare them

Written by AI

wcginc.com · 14 November 2026

AI-generated contentWritten by AI

wcginc.com · Source published: 14 November 2026 at 02:07 (UTC+02:00)

US firm WCG CPAs & Advisors uses AI to compare a prepared tax return with source documents, the prior year and the tax projection. The firm says this does not save time, at least not yet, but adds work.

Jason Watson, chief executive of US accounting firm WCG CPAs & Advisors, described how the firm uses AI in its tax return work in a company blog post on 14 November. WCG spent months testing five major AI tools designed to prepare tax returns. The tools are not quite there yet, the firm said. Watson wrote that most of them extract data or prepare part of a return and then pass it to commercial tax software.

AI comes in only once the return is prepared

WCG still prepares 2026 tax returns, to be filed in 2027, in SurePrep and UltraTax. The prepared return then goes into the tax analysis service of Filed, a WCG partner. Filed brings together the source documents, the prior-year and current-year returns, the tax projection prepared during the year and any tax plans made.

WCG said the material is processed inside Filed in an environment that meets SOC 2 Type II security requirements. Sensitive personal data is masked before the AI processes it, and AI providers do not retain the material after processing. The security claims are the company’s own.

First, Filed compares the return with the source documents, the prior year and the projection. It flags significant changes, items that call for professional judgement and tax planning opportunities. WCG has added its established positions to the analysis, such as its view on reasonable salary for a shareholder-employee.

The findings are compiled into a concise summary for the reviewer. In an example WCG built with fictitious figures, items are marked red, yellow or green, and changes of more than $1,000 or 10 per cent are flagged. Watson said the colours only direct attention. They are not the AI’s assessment of whether the return is correct.

The client manager carries out the actual review, and the firm’s professionals are responsible for the quality and accuracy of the return. Finally, Filed drafts a plain-language explanation for the client covering, for example, changes from the prior year, the effective tax rate and the results of earlier tax plans. Staff check the draft before it is sent.

More work, not less

Watson said AI does not save time, at least not yet, because Filed’s comparison stage, reading the summary and the client explanation all add work. WCG said the extra work is worth it for the analysis and the information that reaches the client.

The firm says it will adopt AI for preparing returns when a vendor can reliably prepare a multi-state return with many K-1 forms from start to finish in commercial tax software. Companies use K-1 forms to report owners’ shares of income.

WCG is also building Kairos, a suite of tools that runs on its own servers. One of them, S2TB, converts bank statements and other source documents into a trial balance for the accounting team and is now rolling out. T2P, already in use, drafts engagement proposals from transcripts of client meetings. WCG will start testing ClientTemp, which compiles the status of engagements, only once personal data can be masked before AI processing.

What this is based on

  1. AI Tax Return Review at WCG: Scrub, Brief, Narrative

    Publisher: wcginc.com · published 14 November 2026 at 02:07 (UTC+02:00) · retrieved 15 November 2026 at 05:26 (UTC+02:00) · discovery channel: web

Read the story →Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Regulation and markets

Two major Nvidia investors urge board to keep Huang’s return possible

Written by AI

cnbc.com + 6 sources · 15 November 2026

AI-generated contentWritten by AI

cnbc.com, businessinsider.com, techcrunch.com, aitimes.com, qbitai.com · 7 sources · Sources published: 14–15 November 2026

Altimeter Capital and Atreides Management have asked Nvidia’s board to move slowly in its CEO search and keep open the possibility of Jensen Huang returning. An online petition demands Huang’s return while the investigation is still under way.

Brad Gerstner, CEO of Altimeter Capital, published a letter to Nvidia’s board on X on Friday, CNBC reported. Gavin Baker of Atreides Management also signed it. The two ask the board for a slow, deliberate and careful process in choosing the company’s future leadership.

The letter does not call for Jensen Huang’s immediate reinstatement. It says the founder has earned a process in which he is not hastily replaced before all the facts are known. Korea’s AI Times reported that the investors acknowledge the board could also reject a return on the basis of the investigation. Baker also said he supports a petition demanding Huang’s reinstatement right away while the independent investigation is under way.

Altimeter held 9.4 million Nvidia shares worth nearly $1.9 billion on 30 June, and Atreides held 1.2 million shares worth nearly $250 million, CNBC reported. AI Times wrote that the Wall Street Journal had pointed out that Huang’s stake of about $140 billion is only about 3.6 per cent of the company, which leaves his position exposed to board decisions.

Huang resigned as CEO and director on 12 November, and Nvidia’s announcement gave no reason. Huang had admitted a relationship with Alice Huang, a former executive of the Singapore company Megaspeed. A relationship between him and an Nvidia executive had also been reported. In October the New York Times reported that US intelligence officials assess Alice Huang may be a Chinese spy, Business Insider wrote. China denies the claim. Attorney General Todd Blanche said the Justice Department has not reached a conclusion on the matter. Nvidia’s board is investigating whether Huang was a target of foreign intelligence. Chief financial officer Colette Kress is serving as interim CEO.

Petition figures come from the organisers

The Stand with Jensen petition calls for Huang to be reinstated while an independent investigation establishes the facts, TechCrunch reported. The organisers say they have been shown no evidence supporting the spying allegation, Business Insider reported.

The signature counts are those reported by the petition website. On Saturday afternoon Business Insider reported more than 108,000 signatures; a TechCrunch article published shortly afterwards put the figure at more than 115,000. TechCrunch’s figures included 5,965 current Nvidia employees and 63,063 shareholders. The list also contains fake names such as ‘Tim Cruise’, and the organisers say they are removing them.

Signatories include Elon Musk and chief executives Amjad Masad (Replit), Tim Sweeney (Epic Games) and Garry Tan (Y Combinator). Musk wrote on X that many of Huang’s supply-chain deals had been made on a handshake rather than in writing. Without Huang, Musk said, at least some of them will be lost.

Board searches for a successor

Business Insider reported that the board is considering both internal and external candidates. China’s QbitAI, citing a Financial Times report relayed by Reuters, said the board has approached Arm CEO Rene Haas. Haas moved from Nvidia’s management to Arm in 2013, QbitAI said. Nvidia said it is still choosing a recruitment firm to assist with the search and that speculation about candidates is premature. AI Times said it is not known whether the board is formally considering Huang’s return.

What this is based on

  1. Nvidia board urged to consider 'potential return of Jensen Huang' by two major investors

    Publisher: www.cnbc.com · published 14 November 2026 at 19:35 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  2. A petition to reinstate Jensen Huang as Nvidia CEO has over 100,000 signatures, including some big names

    Publisher: www.businessinsider.com · published 14 November 2026 at 22:35 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  3. Elon Musk, other tech execs sign petition calling on Nvidia to reinstate Jensen Huang | TechCrunch

    Publisher: techcrunch.com · published 14 November 2026 at 23:41 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  4. 엔비디아 투자자들 "젠슨 황 복귀 검토해야"…머스크 등 거물 가세 - AI타임스

    Publisher: www.aitimes.com · published 15 November 2026 at 00:52 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  5. 英伟达上演“奥特曼2.0”,员工集体请愿:让老黄回家!

    Publisher: www.qbitai.com · published 14 November 2026 at 12:08 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  6. Where was Nvidia's security team? | TechCrunch

    Publisher: techcrunch.com · published 15 November 2026 at 02:08 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  7. Asian governments react to Nvidia CEO Jensen Huang’s exit | TechCrunch

    Publisher: techcrunch.com · published 14 November 2026 at 18:34 (UTC+02:00) · retrieved 15 November 2026 at 05:12 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Regulation and markets

Malaysia examines whether Megaspeed subsidiary’s Nvidia chips were re-exported

Written by AI

aitimes.com + 6 sources · 14 November 2026

AI-generated contentWritten by AI

aitimes.com, techcrunch.com, zeteo.com, abcnews4.com, cnbc.com, malaysiakini.com · 7 sources · Sources published: 14 November 2026

Malaysia’s trade ministry is working with customs to establish whether Nvidia AI chips imported by Speedmatrix, a Megaspeed subsidiary, have been exported to other countries. Prime Minister Anwar Ibrahim said no sign of this had been found so far.

Malaysia’s Home Minister Saifuddin Nasution Ismail said on Saturday 14 November that police had been investigating since April 2025 whether US export restrictions on Nvidia AI chips were circumvented through the country. Police are conducting the investigation under the Strategic Trade Act, Malaysiakini reported. It also covers whether Malaysia was used as a transit route.

The Ministry of Investment, Trade and Industry leads the multi-agency investigation. Saifuddin did not say whether evidence of illegal diversion of the chips had been found, or whether police are investigating Megaspeed or Speedmatrix.

Imports are unrestricted, exports need a permit

Since 14 July 2025, Malaysia has required a permit to export, re-export – that is, send imported chips onward – and transit high-performance US-origin AI chips. Imports need no permit. The ministry and customs are examining whether chips and servers imported by Speedmatrix have been exported to other countries and whether the transactions complied with Malaysian law. Giving false information to customs is punishable, AI Times reported.

Speedmatrix operates in a Malaysian data centre and imports servers and chips for its own use. The ministry said the authorities’ site visit was a routine industry inspection and did not mean the companies’ operations had been approved. Minister Johari Abdul Ghani said on 10 November that no separate investigation into Speedmatrix had been opened in the absence of evidence. Three days later the ministry clarified that neither company had been cleared.

Prime Minister Anwar Ibrahim said Malaysia has received no formal request for investigative assistance from the US or Singapore but will help if asked. Malaysia does not enforce US export restrictions directly, AI Times reported, but the government holds that companies operating in the country must take them into account.

Megaspeed denies wrongdoing as Singapore’s investigation continues

Megaspeed bought $4.8 billion worth of Nvidia AI chips between June 2024 and February 2026, according to reports cited by Malaysiakini. The chips were allegedly delivered to data centres in Malaysia and Indonesia. Megaspeed denies illegal chip trading and says inspections by US and Malaysian authorities and by Nvidia found no wrongdoing. Nvidia has said its inspections confirmed the chips are where they should be. Trade data and documents highlighted by other media, however, show tens of thousands of chips that do not match Nvidia’s figures, The National News Desk reported.

US intelligence suspects Megaspeed’s former chief executive Alice Huang of being a Chinese spy. China denies the claim. Jensen Huang, then Nvidia’s chief executive, admitted a relationship with her and resigned on 12 November, CNBC reported. The US Justice Department is investigating whether Megaspeed diverted restricted chips to China, CNBC also reported.

Singapore’s Ministry of Home Affairs said its investigation into Megaspeed and six individuals is continuing, TechCrunch reported. There was insufficient evidence of Alice Huang’s involvement in the suspected false declarations. She was nevertheless given a warning and an entry ban in February 2026 for knowingly giving police false information about her role. Singapore police charged a fifth suspect on 13 November with money laundering linked to server deals, AI Times reported, but no direct link to Megaspeed has been confirmed.

US Commerce Secretary Howard Lutnick has agreed to appear before the Senate Banking Committee on 10 December on whether his department is preventing China from acquiring advanced chips, according to a letter cited by CNBC.

What this is based on

  1. 말레이시아, 메가스피드 AI 칩 재수출 의혹 조사 착수 - AI타임스

    Publisher: www.aitimes.com · published 14 November 2026 at 08:53 (UTC+02:00) · retrieved 15 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

  2. Asian governments react to Nvidia CEO Jensen Huang’s exit | TechCrunch

    Publisher: techcrunch.com · published 14 November 2026 at 18:34 (UTC+02:00) · retrieved 15 November 2026 at 05:12 (UTC+02:00) · discovery channel: web

  3. ‘Feeding Frenzy’: Trump’s Administration Erupts Over the Nvidia Scandal

    Publisher: zeteo.com · published 14 November 2026 at 18:03 (UTC+02:00) · retrieved 15 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

  4. US officials raise alarm over restricted Nvidia AI chips reaching China

    Publisher: abcnews4.com · published 14 November 2026 at 04:39 (UTC+02:00) · retrieved 15 November 2026 at 05:17 (UTC+02:00) · discovery channel: web

  5. Nvidia board urged to consider 'potential return of Jensen Huang' by two major investors

    Publisher: www.cnbc.com · published 14 November 2026 at 19:35 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  6. Elon Musk, other tech execs sign petition calling on Nvidia to reinstate Jensen Huang | TechCrunch

    Publisher: techcrunch.com · published 14 November 2026 at 23:41 (UTC+02:00) · retrieved 15 November 2026 at 05:11 (UTC+02:00) · discovery channel: web

  7. Cops probing alleged Nvidia chip diversion since April 2025, says Saifuddin | Malaysiakini

    Publisher: www.malaysiakini.com · published 14 November 2026 · retrieved 15 November 2026 at 05:18 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Regulation and markets

Report: Nvidia has begun cutting the financing its customers used to buy its chips

Written by AI

techspot.com + 6 sources · 15 November 2026

AI-generated contentWritten by AI

techspot.com, x.com, journalarta.com, notebookcheck.net, oilprice.com, shattered.io · 7 sources · Sources published: 13–15 November 2026

Nvidia financed its customers to sell off TSMC’s oversupply and keep GPU prices high, but began slowing the deals in summer 2026, the WSJ reported. Drawing on comments by analyst Ian Cutress, Notebookcheck and JournalArta expect gaming graphics cards not to get cheaper quickly.

The Wall Street Journal (WSJ) reported on Nvidia’s financing arrangements in a profile of interim chief executive Colette Kress. The technology site TechSpot summarised the story on 14 November. Kress has combined the post with that of chief financial officer since Jensen Huang stepped down as chief executive.

In circular financing, a company funds its customers and the money comes back to it as product purchases. Nvidia is reported to have given guarantees and other support that helped cloud providers and AI companies secure financing for its hardware, Notebookcheck wrote.

Oversupply and high prices

The WSJ reported that Nvidia struck more financing deals after mid-2026, TechSpot wrote. Nvidia wanted the deals to absorb an oversupply of TSMC wafers and keep GPU prices high. In the summer, Nvidia was worried about a flood of chips it had ordered from TSMC.

Nvidia then began slowing, shrinking and pausing some of the deals. One is the OpenAI deal, which originally included a $250 billion backstop for the world’s largest data centre. Shattered.io noted that it has not been confirmed whether the figure is an investment, a purchase commitment or a backstop.

At a conference in Berlin, Huang told major customers a story about a Japanese soup stall that did not have enough soup for everyone. TechSpot concluded from the story that the financing pushed up prices by creating an impression of strong demand. The conclusion is TechSpot’s own. Shattered.io pointed out that there is no evidence for the claim that scarcity was created deliberately. The site said the problem described in the reports was the opposite: there was too much supply.

Analysts expect a pause

UBS analyst Timothy Arcuri expects Nvidia to pause new financing deals even if the new chief executive sticks to the same approach, Oilprice.com reported. JPMorgan analysts do not expect existing commitments to change. New deals could, however, take longer or cost more.

Gaming graphics cards will not get cheaper quickly

GeForce card prices often exceed 160 per cent of the original recommended retail price, TechSpot reported. Notebookcheck does not expect the financing cuts to fix this quickly. Nor can Nvidia quickly shift production from AI chips to GeForce cards, the Indonesian site JournalArta wrote. Both draw on comments by Ian Cutress, chief analyst at the analyst firm More Than Moore; Notebookcheck said PCWorld published them.

AI accelerators typically use high-bandwidth HBM memory and complex packaging, while gaming cards use GDDR memory. Some manufacturing resources are shared, but the production lines are not interchangeable. ‘Typically vendors have to make wafer orders from people like TSMC and wait six months from start of manufacturing to end,’ Cutress said. After that come memory, packaging and assembly.

Notebookcheck wrote that reduced financing could free up capacity for GeForce production over time. That would, however, require new orders, and existing contracts will not disappear. The site considers it unlikely that affordable gaming graphics cards will return in 2027.

Shattered.io pointed out that no information released by Nvidia and no named person links the financing to consumer prices. Several media outlets attribute the prices to the memory shortage.

What this is based on

  1. Nvidia deliberately kept GPU prices high through circular financing deals, report claims

    Publisher: www.techspot.com · published 14 November 2026 at 20:07 (UTC+02:00) · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

  2. x.com

    TechPowerUp · Publisher: x.com · published 13 November 2026 · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: X

  3. Nvidia Can’t Simply Make More GeForce Cards, Analyst Says

    Publisher: journalarta.com · published 14 November 2026 at 13:51 (UTC+02:00) · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

  4. Nvidia’s retreat from Huang's circular AI financing won’t bring back affordable GeForce cards in 2027

    Publisher: www.notebookcheck.net · published 14 November 2026 at 20:21 (UTC+02:00) · retrieved 15 November 2026 at 05:23 (UTC+02:00) · discovery channel: web

  5. Nvidia Without Huang: What a Pause in AI Deals Means for Metals | OilPrice.com

    Publisher: oilprice.com · published 15 November 2026 at 00:00 (UTC+02:00) · retrieved 15 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

  6. NVIDIA AI Investments and GPU Prices: $17.5B vs $9,400

    Publisher: shattered.io · published 15 November 2026 at 00:14 (UTC+02:00) · retrieved 15 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

  7. NVIDIA Circular Financing Claim: $250B, 0 Confirmed

    Publisher: shattered.io · published 14 November 2026 at 08:38 (UTC+02:00) · retrieved 15 November 2026 at 05:24 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Regulation and markets

Italy fines Meta €805 million over Facebook’s Android photo suggestions

Written by AI

live.euronext.com + 4 sources · 14 November 2026

AI-generated contentWritten by AI

live.euronext.com, politico.eu, aljazeera.com, finance.yahoo.com, bloomberg.com · 5 sources · Sources published: 14 November 2026

Facebook’s Android app accessed users’ photos and videos without specific consent, Italy’s data protection authority said. Meta plans to appeal and considers the fine disproportionate.

Italy’s data protection authority, the Garante, said on Saturday 14 November that it had fined Meta over a feature in Facebook’s Android app that suggests photos and videos from the phone’s gallery for posting. The €805 million fine is by far the largest in the authority’s history, Politico and Al Jazeera reported.

The app used photos, videos and their metadata stored on phones without specific consent and did not tell users, the authority said. The feature has been in use since 2023, according to the German news agency dpa.

One permission for the whole app

The authority found that the app requests gallery access with a single permission covering multiple features with different needs, Politico reported. Before the request, the app showed a generic warning from which users could not tell that they were accepting photo and video suggestions. The suggestions were also switched on by default, Al Jazeera reported.

If the permission is granted in full, the app can read the device’s entire photo archive, usually permanently, and neither users nor the authorities can check which data is actually used, the authority wrote in a statement quoted by Politico.

The authority said the feature breaches the Italian law that transposes the EU’s ePrivacy Directive on electronic communications. Meta demanded that the case go to the Irish authority, because the company’s European headquarters is in Ireland. The authority rejected the demand: the ePrivacy rules have no one-stop-shop mechanism under which the authority in the country of a company’s main establishment handles cross-border cases.

The fine equals about 0.47 per cent of Meta’s global revenue, the authority said – Meta’s 2024 revenue, according to dpa. The feature affected around 20 million Android users in Italy, Reuters and dpa reported, while Al Jazeera presents the figure as a possible maximum. The authority cited Meta’s lack of cooperation and a previous violation, for which the company was fined in 2019, as factors in the size of the fine, Al Jazeera reported.

Meta has 60 days to switch the feature off by default, including for users who have already granted permission. The company must request specific consent for the feature and clearly explain what it does and how to disable it.

The investigation, launched in 2025, did not cover whether the content is used to train AI models, Reuters and Al Jazeera reported. Meta denies any such use. The authority reserves the right to assess separately a feature launched in 2026 that processes gallery content in the cloud, Reuters reported.

Meta plans to appeal

A Meta spokesperson said the company would appeal and called the fine disproportionate, Reuters reported. The spokesperson described the photo suggestions as a simple feature that helps people share photos and videos on Facebook, adding that users must grant permission on their device before the feature can work and can change that permission at any time in their settings. When Politico and Al Jazeera published their reports, Meta had not yet commented.

What this is based on

  1. Italy fines Meta $887 million over Android users' data access | live

    Publisher: live.euronext.com · published 14 November 2026 · retrieved 15 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

  2. Italy hits Meta with €805m fine over privacy law violation

    Publisher: www.politico.eu · published 14 November 2026 at 16:03 (UTC+02:00) · retrieved 15 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

  3. Italy fines Meta $880m over photo privacy

    Publisher: www.aljazeera.com · published 14 November 2026 at 18:16 (UTC+02:00) · retrieved 15 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

  4. Italy fines Meta €805 million over Android app's photo access

    Publisher: finance.yahoo.com · published 14 November 2026 at 18:18 (UTC+02:00) · retrieved 15 November 2026 at 05:38 (UTC+02:00) · discovery channel: web

  5. Italy Fines Meta €805 Million Over Access to Users’ Photos

    Publisher: www.bloomberg.com · published 14 November 2026 at 16:53 (UTC+02:00) · retrieved 15 November 2026 at 05:37 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Regulation and markets

Canadian publishers’ copyright lawsuit against OpenAI to be heard in Ontario

Written by AI

torontotoday.ca + 1 source · 14 November 2026

AI-generated contentWritten by AI

torontotoday.ca, barsacross.com · 2 sources · Sources published: 14 November 2026

The Court of Appeal for Ontario has dismissed OpenAI’s appeal, so Canadian news publishers’ lawsuit over ChatGPT’s training data can proceed in Ontario. The ruling concerns only where the case is heard, not whether OpenAI infringed copyright.

OpenAI has lost a dispute over which country should hear a lawsuit brought against it by Canadian media outlets. The Canadian Press news agency, itself one of the plaintiffs, reported the ruling on Saturday on the TorontoToday website.

The Court of Appeal for Ontario dismissed OpenAI’s appeal on Friday 13 November. It is a provincial court in Ontario, not the Supreme Court of Canada. The three-judge panel was unanimous, according to Bars Across, a legal news site.

The Canadian Press, Torstar, The Globe and Mail, Postmedia and CBC/Radio-Canada were among the publishers that filed the lawsuit in November 2024. They allege that OpenAI copied their articles without permission, used them as training data for its GPT models and profited from them without paying. The suit also alleges that OpenAI circumvented technical protection measures and breached the websites’ terms of use, Bars Across reported. The plaintiffs seek, among other things, a permanent injunction and damages or, alternatively, statutory damages of 20,000 Canadian dollars per work. None of the allegations has been proven.

Why the case stays in Canada

OpenAI argued that its conduct took place abroad and asked for the case to be heard in the United States. Justice Jessica Kimmel of the Ontario Superior Court largely rejected that request on 7 November 2025. Bars Across reported that Kimmel found the plaintiffs had sufficiently shown that three OpenAI operating companies are connected to Ontario through their business, alleged local acts and the websites’ terms of use. The action continues against three holding companies as a claim for unjust enrichment. It was stayed against four other OpenAI companies because their role in the conduct had not been shown.

The Court of Appeal found that OpenAI had put forward no evidence that a Canadian judgment would conflict with US law and so complicate enforcement, The Canadian Press reported. The ruling also said the judge had considered that the functioning of the Canadian justice system favoured letting Canadian creators sue foreign companies for breaching Canadian law.

The copyright question remains open

The ruling decides only where the remaining claims will be heard. Still unresolved, Bars Across wrote, are how far Canadian copyright law reaches beyond the country’s borders and how the use of material in AI development is treated.

The Canadian Press said this is the first Canadian lawsuit over the use of copyrighted material to train generative AI. This year, Postmedia, Torstar, The Globe and Mail and The Canadian Press filed a separate lawsuit against Anthropic. The federal government has not said it plans to reform copyright law, the agency reported.

What this is based on

  1. Top court allows news publishers' lawsuit against OpenAI to go ahead

    Publisher: www.torontotoday.ca · published 14 November 2026 at 14:00 (UTC+02:00) · retrieved 15 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

  2. Ontario Court of Appeal Keeps OpenAI Copyright Case in Canada

    Publisher: www.barsacross.com · published 14 November 2026 at 20:23 (UTC+02:00) · retrieved 15 November 2026 at 05:39 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Speech

Open-weight speech model Fluid Speech Pro recognises 25 languages, including Finnish

Written by AI

github.com + 5 sources · 13 November 2026

AI-generated contentWritten by AI

github.com, huggingface.co · 6 sources · Sources published: 13 November 2026 · Some source publication dates are unknown

FluidVoice has released a speech recognition model for 25 European languages that users can download from Hugging Face. By the company’s own account, the model is the second most accurate open-weight model on the Open ASR Leaderboard in English tests.

The FluidVoice team published the full model card for Fluid Speech Pro, along with ONNX and GGUF versions, on Hugging Face on 13 November. The model turns speech into text, detects the spoken language itself and adds punctuation and capitalisation.

The model is built on NVIDIA’s Parakeet TDT 0.6B v3 and has 0.63 billion parameters. It keeps Parakeet’s architecture but adds one small extra block at the start of audio processing. An optional clean mode removes repeated words and fillers such as ‘um’ and ‘uh’. The developers advise against using it when a verbatim transcript is needed.

Accuracy by the company’s own measurements

FluidVoice said the model ranks second among open-weight models on Hugging Face’s Open ASR Leaderboard. The metric is word error rate (WER), the share of words recognised incorrectly; lower is better. The leaderboard figures come from beam search (4 beams), which the developers said makes about 0.6 per cent fewer errors than the model’s default decoding.

The table shows the top four models in the model card’s comparison and the base model Parakeet, which ranks eighth:

Model Parameters Average WER
Musci-ASR 2bn 5.70
Fluid Speech Pro 0.63bn 6.06
NVIDIA Nemotron 3.5 Transcribe 30bn 6.13
Qwen3-ASR 1.7bn 6.15
NVIDIA Parakeet TDT 0.6B v3 0.6bn 7.04

Source: FluidVoice model card; average across seven public and two private English test sets.

The table says nothing about how the model performs in Finnish. The ranking rests on the company’s model card and has not been independently verified. In an October evaluation file in the repository, the company reported results it had run itself on a single H100 GPU and noted that the leaderboard maintainers’ own run supersedes them. Some individual results in the file differ slightly from the model card: on the URGENT 2024 noise test, for example, the file gives 9.80 and the model card 9.83. The model card’s speed of about 3,700 times real time on a single GPU appears in the file as 3,556, measured on an H100 rather than the leaderboard’s H200.

The Finnish result

For Finnish, there is only the company’s own FLEURS measurement with default decoding: a WER of 8.82, against 11.19 for Parakeet TDT 0.6B v3. That is about 21 per cent fewer errors. Finnish is not part of the leaderboard’s seven-language multilingual set.

The developers urge users to test the model on their own recordings, accents and languages before relying on it. Errors increase with heavy noise, overlapping speech and rare names, they said.

Versions and deployment

The model comes in five versions: NeMo and Transformers for NVIDIA GPUs, ONNX for CPUs, MLX for Apple Silicon Macs, GGUF for the C++ program parakeet.cpp, and Core ML for Apple apps. The Transformers and GGUF weights lack the extra block. The GGUF version may recognise noisy audio slightly less accurately, the developers said. Clean mode is available only in the NeMo, ONNX and MLX versions.

The developers said the model is also suitable for commercial use. It is licensed under CC BY 4.0, and the developers ask users to credit it as ‘Fluid Speech Pro by FluidVoice’. Hugging Face hosts a browser demo that uses the Transformers weights, and the model card contains usage instructions.

What this is based on

  1. github.com

    Publisher: github.com · source publication time unknown · retrieved 15 November 2026 at 05:22 (UTC+02:00) · discovery channel: GitHub

  2. Fluid Speech Pro

    Publisher: huggingface.co · published 13 November 2026 at 14:22 (UTC+02:00) · retrieved 15 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

  3. FluidVoice-Labs/fluid-speech-pro

    Publisher: huggingface.co · source publication time unknown · retrieved 15 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

  4. .eval_results/open_asr_leaderboard.yaml · FluidVoice-Labs/fluid-speech-pro at d603a64

    Publisher: huggingface.co · source publication time unknown · retrieved 15 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

  5. FluidVoice-Labs/fluid-speech-pro-onnx

    Publisher: huggingface.co · published 13 November 2026 at 12:29 (UTC+02:00) · retrieved 15 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

  6. FluidVoice-Labs/fluid-speech-pro-gguf

    Publisher: huggingface.co · published 13 November 2026 at 12:40 (UTC+02:00) · retrieved 15 November 2026 at 05:22 (UTC+02:00) · discovery channel: web

Open on its own pageLue juttu suomeksi →Läs den på svenska →Close story ↑

Review and omissions →

Omissions and review

What was left out and why