Anthropic’s Lance Martin: old prompting habits raise the cost of new Claude models EN
Approved
Writing and review are carried out in separate model calls.
Review and omissions
These records concern the exact text versions in this issue. Article reviews are not a separate quality audit of the issue.
Omitted and unfinished (497)
- RT @_NathanCalvin: Uhhh Anthropic made an offer to one of the risk assessment staff at METR during this project??? (Which the person declin…: Left out in the newsroom's selection.
- RT @DKokotajlo: @AnthropicAI: what the fuck? The compute allocation assessment is the thing we've been pushing for the last few months. It'…: Left out in the newsroom's selection.
- Cloudflare's Trust Framework for AI Agents (4 minute read): Left out in the newsroom's selection.
- The benchmark where AIs have to lie to win (23 minute read): Left out in the newsroom's selection.
- RT @lugaricano: This is fascinating (and fun). @dwarkesh_sp has the AIs play 400 rounds of a zero-sum game of incomplete and asymmetric inf…: Left out in the newsroom's selection.
- RT @polynoamial: Honestly quite impressed with the methodology on this benchmark. @dwarkesh_sp want a job designing evals at @OpenAI? https…: Left out in the newsroom's selection.
- NVIDIA Introduced VFlash (5 minute read): Left out in the newsroom's selection.
- New NVIDIA Cosmos 3 Open Models for Interactive World Simulation (7 minute read): Left out in the newsroom's selection.
- RT @LeRobotHF: 𝗟𝗲𝗥𝗼𝗯𝗼𝘁 𝘃𝟬.𝟳.𝟬 𝗶𝘀 𝗟𝗜𝗩𝗘 🚀🤖 From terabytes of data to prompting robots! Check out all what's new in the video below and join…: Left out in the newsroom's selection.
- ToyBlocks: A controlled world for robot learning and evaluation: Left out in the newsroom's selection.
- ToyBlocks: A controlled world for robot learning and evaluation: Left out in the newsroom's selection.
- Ask the Experts: Dory Looped Transformer Model | Nemotron Labs: Left out in the newsroom's selection.
- From Human Demonstrations to Robot Training Data With RLWRLD | Cosmos Labs: Left out in the newsroom's selection.
- Video Reasoning and Model Fine-Tuning with Roboflow and NVIDIA Cosmos 3 | Cosmos Labs: Left out in the newsroom's selection.
- 로봇·자율주행 겨냥…샤오펑, 오픈소스 모델 '그라운딩PI' 공개: Left out in the newsroom's selection.
- How Primary Compute Markets Work (7 minute read): Left out in the newsroom's selection.
- The Ultimate MoE Training Handbook (1 minute read): Left out in the newsroom's selection.
- Stream live training runs on Hugging Face!: Left out in the newsroom's selection.
- RT @jeffboudier: We spent $2m so you don't have to. And so you can learn how to train MoEs. Mixture-of-Experts LLMs are what powers the…: Left out in the newsroom's selection.
- RT @Thom_Wolf: 18 of the 20 best open models today are Mixture-of-Experts. Dense is basically over. And yet the know-how to train a good M…: Left out in the newsroom's selection.
Reviewed versions (105)
Open review records by language
EN (35)
Anthropic says ‘dreaming’ corrected a Claude note in its Pokémon test EN
Approved
Writing and review are carried out in separate model calls.
Claude Code 2.1.319: auto mode no longer uses its built-in classifier by default behind a gateway EN
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Potential buyers put Anthropic’s value at $1.1–1.3 trillion EN
Approved
Writing and review are carried out in separate model calls.
Meta starts bringing models trained with Muse Wonder to Muse and Meta AI EN
Approved
Writing and review are carried out in separate model calls.
Nous Research’s own test: fine-tuned Qwen repeats China’s official line less often EN
Approved
Writing and review are carried out in separate model calls.
llama.cpp can now run Kimi-K3.1-Pro, but only for text so far EN
Approved
Writing and review are carried out in separate model calls.
Artificial Analysis: GPT Image 2.6 leads image generation and editing comparisons EN
Approved
Writing and review are carried out in separate model calls.
Claude Code works out a repository’s dependencies in the cloud on Pro, Max and Ultra plans EN
Approved
Writing and review are carried out in separate model calls.
OpenClaw now gets its model list from a catalogue, the user’s account and the installed Codex EN
Approved
Writing and review are carried out in separate model calls.
Perplexity launches Teammates agents for ongoing work projects EN
Approved
Writing and review are carried out in separate model calls.
Censys finds more than 800 ARTEX hacking tool deployments open to the web without a password EN
Approved
Writing and review are carried out in separate model calls.
Anthropic relaxes Claude Code’s game ban: games inspired by a well-known title are allowed EN
Approved
Writing and review are carried out in separate model calls.
Developer says ChatGPT made a follow-up call through a different MCP connector without telling the user EN
Approved
Writing and review are carried out in separate model calls.
Cognition says Devin Review now costs 82 per cent less than in August EN
Approved
Writing and review are carried out in separate model calls.
Cursor’s cloud agent builds web apps with a database, ready to publish EN
Approved
Writing and review are carried out in separate model calls.
In Amp’s Spaces, one agent spawns and coordinates other agents EN
Approved
Writing and review are carried out in separate model calls.
GitHub: weekly commits up more than 400 per cent in a year EN
Approved
Writing and review are carried out in separate model calls.
Artificial Analysis: Tencent’s Hy4 ranks second among Chinese models, with an 89 per cent hallucination rate EN
Approved
Writing and review are carried out in separate model calls.
Epoch AI estimates Huawei’s Horinger data centre has 7 per cent of Colossus 2’s computing capacity EN
Approved
Writing and review are carried out in separate model calls.
PSI: system using an Anthropic model proved the Aaronson–Ambainis conjecture EN
Approved
Writing and review are carried out in separate model calls.
Preprint: all 16 evolved agents lost previously solved tasks EN
Approved
Writing and review are carried out in separate model calls.
Preprint: GPT-5.6 Sol solved 55.5 per cent of Terminal-Bench 4.0 tasks with the HEELO harness EN
Approved
Writing and review are carried out in separate model calls.
DeepMind says human-guided AI found two weaknesses in legacy RSA-1024 encryption EN
Approved
Writing and review are carried out in separate model calls.
Anthropic’s EVE test: Claude teams almost never unmasked a member told to sabotage EN
Approved
Writing and review are carried out in separate model calls.
Asteromorph releases open-weight 4B reranker for scientific search EN
Approved
Writing and review are carried out in separate model calls.
Vals Legal Index: Wonder Full leads, two Gemini 4 models close behind EN
Approved
Writing and review are carried out in separate model calls.
A&O Shearman pledges $1.1 billion to AI over five years – mostly in working time EN
Approved
Writing and review are carried out in separate model calls.
Spain’s Kabilio raises €10 million for an AI agent built for accounting firms EN
Approved
Writing and review are carried out in separate model calls.
Prosecutors say three men used AI to impersonate executives and took $129 million EN
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Microsoft gives up AI computing capacity rented from Oracle EN
Approved
Writing and review are carried out in separate model calls.
Google buys Crusoe’s Wyoming data centre project and plans to invest $17bn EN
Approved
Writing and review are carried out in separate model calls.
Cleanview: governors’ pauses would cover 41 per cent of planned US data centre capacity EN
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Nvidia demands chip tracking from Southeast Asian firms seeking partner status EN
Approved
Writing and review are carried out in separate model calls.
Hugging Face’s speech recognition leaderboard adds Egyptian and Nigerian English tests EN
Approved
Writing and review are carried out in separate model calls.
FI (35)
Anthropicin Lance Martin: vanhat kehotetavat lisäävät uusien Claude-mallien kuluja FI
Approved
Writing and review are carried out in separate model calls.
Anthropic: ”uneksiminen” korjasi Clauden muistiinpanon yhtiön Pokémon-testissä FI
Approved
Writing and review are carried out in separate model calls.
Claude Code 2.1.319: auto mode ei enää oletuksena käytä omaa luokittelijaa yhdyskäytävällä FI
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Mahdolliset ostajat arvioivat Anthropicin arvoksi 1,1–1,3 biljoonaa dollaria FI
Approved
Writing and review are carried out in separate model calls.
Meta alkoi tuoda Muse Wonderin avulla koulutettuja malleja Museen ja Meta AI:hin FI
Approved
Writing and review are carried out in separate model calls.
Nous Researchin oma testi: hienosäädetty Qwen toistaa Kiinan virallista linjaa harvemmin FI
Approved
Writing and review are carried out in separate model calls.
llama.cpp osaa nyt ajaa Kimi-K3.1-Pro-mallia, mutta toistaiseksi vain tekstiä FI
Approved
Writing and review are carried out in separate model calls.
Artificial Analysis: GPT Image 2.6 kärjessä kuvanluonnin ja -muokkauksen vertailuissa FI
Approved
Writing and review are carried out in separate model calls.
Claude Code selvittää pilvessä repositorion riippuvuudet Pro-, Max- ja Ultra-tilauksissa FI
Approved
Writing and review are carried out in separate model calls.
OpenClaw hakee mallilistan nyt katalogista, käyttäjän tililtä ja asennetusta Codexista FI
Approved
Writing and review are carried out in separate model calls.
Perplexity julkaisi Teammates-agentit jatkuvien työprojektien hoitoon FI
Approved
Writing and review are carried out in separate model calls.
Censys: yli 800 ARTEX-hyökkäystyökalun asennusta auki verkossa ilman salasanaa FI
Approved
Writing and review are carried out in separate model calls.
Anthropic löysensi Claude Coden pelikieltoa: tunnetun pelin innoittama peli onnistuu FI
Approved
Writing and review are carried out in separate model calls.
Kehittäjä: ChatGPT teki jatkokutsun toisen MCP-liitännän kautta kertomatta käyttäjälle FI
Approved
Writing and review are carried out in separate model calls.
Cognition: Devin Review -katselmointi maksaa nyt 82 % vähemmän kuin elokuussa FI
Approved
Writing and review are carried out in separate model calls.
Cursorin pilviagentti rakentaa verkkosovelluksen tietokantoineen julkaisuvalmiiksi FI
Approved
Writing and review are carried out in separate model calls.
Ampin Spaces-työtilassa yksi agentti käynnistää ja koordinoi muita agentteja FI
Approved
Writing and review are carried out in separate model calls.
GitHub: commitien viikkomäärä kasvoi vuodessa yli 400 prosenttia FI
Approved
Writing and review are carried out in separate model calls.
Artificial Analysis: Tencentin Hy4 on kiinalaismalleista toinen, hallusinaatioaste 89 % FI
Approved
Writing and review are carried out in separate model calls.
Epoch AI:n arvio: Huawein Horingerin datakeskuksessa on 7 % Colossus 2:n laskentatehosta FI
Approved
Writing and review are carried out in separate model calls.
PSI: Anthropicin mallia käyttänyt järjestelmä todisti Aaronson–Ambainis-konjektuurin FI
Approved
Writing and review are carried out in separate model calls.
Esijulkaisu: kaikki 16 kehitettyä agenttia menettivät aiemmin ratkaistuja tehtäviä FI
Approved
Writing and review are carried out in separate model calls.
Esijulkaisu: GPT-5.6 Sol ratkaisi HEELO-harnessilla 55,5 % Terminal-Bench 4.0 -tehtävistä FI
Approved
Writing and review are carried out in separate model calls.
DeepMind: ihmisten ohjaama tekoäly löysi kaksi heikkoutta vanhoista RSA-1024-salauksista FI
Approved
Writing and review are carried out in separate model calls.
Anthropicin EVE-testi: Claude-tiimit tuskin paljastivat sabotoimaan käskettyä jäsentä FI
Approved
Writing and review are carried out in separate model calls.
Asteromorph julkaisi avointen painojen 4B-uudelleenjärjestäjän tieteelliseen hakuun FI
Approved
Writing and review are carried out in separate model calls.
Vals AI:n lakityöindeksi: Wonder Full kärjessä, kaksi Gemini 4 -mallia lähellä FI
Approved
Writing and review are carried out in separate model calls.
A&O Shearman lupaa tekoälyyn 1,1 mrd. dollaria viidessä vuodessa – pääosin työaikaa FI
Approved
Writing and review are carried out in separate model calls.
Espanjalainen Kabilio keräsi 10 miljoonaa euroa tilitoimistojen tekoälyagenttiin FI
Approved
Writing and review are carried out in separate model calls.
Syyttäjät: kolmikko vei 129 milj. dollaria jäljittelemällä tekoälyllä yritysjohtajia FI
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Microsoft luopuu Oraclelta vuokraamastaan tekoälylaskennasta FI
Approved
Writing and review are carried out in separate model calls.
Google osti Crusoen Wyomingin datakeskushankkeen ja aikoo investoida 17 mrd. dollaria FI
Approved
Writing and review are carried out in separate model calls.
Cleanview: kuvernöörien tauot koskisivat 41 %:a USA:n suunnitellusta datakeskustehosta FI
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Nvidia vaatii Kaakkois-Aasiassa kumppaniasemaa hakevilta sirujen seurantaa FI
Approved
Writing and review are carried out in separate model calls.
Hugging Facen puheentunnistusvertailuun egyptiläisen ja nigerialaisen englannin testit FI
Approved
Writing and review are carried out in separate model calls.
SV (35)
Anthropics Lance Martin: Gamla promptvanor ökar kostnaderna för nya Claude-modeller SV
Approved
Writing and review are carried out in separate model calls.
Anthropic: ”Drömmande” rättade Claudes anteckning i företagets Pokémon-test SV
Approved
Writing and review are carried out in separate model calls.
Claude Code 2.1.319: Auto mode använder inte längre sin egen klassificerare som standard bakom en LLM-gateway SV
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Potentiella köpare värderar Anthropic till 1,1–1,3 biljoner dollar SV
Approved
Writing and review are carried out in separate model calls.
Meta har börjat föra in modeller tränade med Muse Wonder i Muse och Meta AI SV
Approved
Writing and review are carried out in separate model calls.
Nous Researchs eget test: Finjusterad Qwen upprepar Kinas officiella linje mer sällan SV
Approved
Writing and review are carried out in separate model calls.
llama.cpp kan nu köra Kimi-K3.1-Pro, men än så länge bara text SV
Approved
Writing and review are carried out in separate model calls.
Artificial Analysis: GPT Image 2.6 i topp i jämförelser av bildgenerering och bildredigering SV
Approved
Writing and review are carried out in separate model calls.
Claude Code tar reda på kodförrådets beroenden i molnet för Pro-, Max- och Ultra-abonnemang SV
Approved
Writing and review are carried out in separate model calls.
OpenClaw hämtar nu modellistan från katalog, användarens konto och installerad Codex SV
Approved
Writing and review are carried out in separate model calls.
Perplexity lanserar Teammates-agenter för löpande arbetsprojekt SV
Approved
Writing and review are carried out in separate model calls.
Censys: Över 800 installationer av attackverktyget ARTEX öppna på nätet utan lösenord SV
Approved
Writing and review are carried out in separate model calls.
Anthropic luckrar upp Claude Codes spelförbud – spel inspirerade av kända spel tillåts SV
Approved
Writing and review are carried out in separate model calls.
Utvecklare: ChatGPT gjorde uppföljningsanrop via annan MCP-koppling utan att berätta för användaren SV
Approved
Writing and review are carried out in separate model calls.
Cognition: Granskning med Devin Review kostar nu 82 procent mindre än i augusti SV
Approved
Writing and review are carried out in separate model calls.
Cursors molnagent bygger publiceringsklar webbapp med databas SV
Approved
Writing and review are carried out in separate model calls.
I Amps arbetsyta Spaces startar och samordnar en agent andra agenter SV
Approved
Writing and review are carried out in separate model calls.
GitHub: Antalet commits per vecka ökade med över 400 procent på ett år SV
Approved
Writing and review are carried out in separate model calls.
Artificial Analysis: Tencents Hy4 tvåa bland kinesiska modeller, hallucinationsgrad 89 procent SV
Approved
Writing and review are carried out in separate model calls.
Epoch AI:s uppskattning: Huaweis datacenter i Horinger har 7 procent av Colossus 2:s beräkningskapacitet SV
Approved
Writing and review are carried out in separate model calls.
PSI: System med Anthropic-modell bevisade Aaronson–Ambainis-förmodan SV
Approved
Writing and review are carried out in separate model calls.
Förhandspublicering: Alla 16 utvecklade agenter förlorade tidigare lösta uppgifter SV
Approved
Writing and review are carried out in separate model calls.
Förhandspublicering: GPT-5.6 Sol löste 55,5 procent av uppgifterna i Terminal-Bench 4.0 med HEELO-harnessen SV
Approved
Writing and review are carried out in separate model calls.
DeepMind: Människostyrd AI hittade två svagheter i äldre RSA-1024-kryptering SV
Approved
Writing and review are carried out in separate model calls.
Anthropics EVE-test: Claude-team avslöjade nästan aldrig medlemmen som beordrats sabotera SV
Approved
Writing and review are carried out in separate model calls.
Asteromorph släpper 4B-reranker med öppna vikter för vetenskaplig sökning SV
Approved
Writing and review are carried out in separate model calls.
Vals AI:s Legal Index: Wonder Full i topp, två Gemini 4-modeller tätt efter SV
Approved
Writing and review are carried out in separate model calls.
A&O Shearman utlovar 1,1 miljarder dollar till AI på fem år – främst arbetstid SV
Approved
Writing and review are carried out in separate model calls.
Spanska Kabilio tar in 10 miljoner euro till AI-agent för redovisningsbyråer SV
Approved
Writing and review are carried out in separate model calls.
Åklagare: Trio tog 129 miljoner dollar genom att imitera företagsledare med AI SV
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Microsoft avvecklar AI-kapacitet som företaget hyrt av Oracle SV
Approved
Writing and review are carried out in separate model calls.
Google köpte Crusoes datacenterprojekt i Wyoming och planerar att investera 17 miljarder dollar SV
Approved
Writing and review are carried out in separate model calls.
Cleanview: Guvernörernas pauser skulle gälla 41 procent av USA:s planerade datacenterkapacitet SV
Approved
Writing and review are carried out in separate model calls.
Bloomberg: Nvidia kräver chipspårning för partnerstatus i Sydostasien SV
Approved
Writing and review are carried out in separate model calls.
Egyptisk och nigeriansk engelska läggs till i Hugging Faces topplista för taligenkänning SV
Approved
Writing and review are carried out in separate model calls.