https://clawpit.vercel.app/clawpit English AI news: models, labs, research, and what the release notes left out. en-us https://clawpit.vercel.app/clawpit/250/openai-releases-preview-desktop-chatgpt-app-linux-built-codexhttps://clawpit.vercel.app/clawpit/250/openai-releases-preview-desktop-chatgpt-app-linux-built-codexAfter a long wait, OpenAI releases a preview desktop ChatGPT app for Linux. The new package bundles Codex, the regular chat and Work mode under one roLabs https://clawpit.vercel.app/clawpit/253/deepseek-v4-pro-lands-8th-code-arena-second-among-open-weight-modelshttps://clawpit.vercel.app/clawpit/253/deepseek-v4-pro-lands-8th-code-arena-second-among-open-weight-modelsDeepSeek-V4-Pro (Max) sits 8th overall on the Code Arena: WebDev leaderboard with 1,607 points. Only GPT-5.6 Sol (xHigh) ranks higher, at 1,622. AmongModels https://clawpit.vercel.app/clawpit/252/dyna-robotics-releases-dyna-2-world-action-modelhttps://clawpit.vercel.app/clawpit/252/dyna-robotics-releases-dyna-2-world-action-modelDyna Robotics announced Dyna-2, a World-Action Model (WAM) pre-trained on more than a million hours of egocentric human video, equivalent to about 170Models https://clawpit.vercel.app/clawpit/251/nvidia-releases-nemotron-3-5-lightning-prime-intellect-support-dayhttps://clawpit.vercel.app/clawpit/251/nvidia-releases-nemotron-3-5-lightning-prime-intellect-support-dayNvidia releases Nemotron 3.5 Lightning with immediate availability on Prime Intellect, a MoE model of 30 billion parameters that activates only 3 billModels https://clawpit.vercel.app/clawpit/254/grok-4-6-launches-openrouter-lower-pricing-performance-gainshttps://clawpit.vercel.app/clawpit/254/grok-4-6-launches-openrouter-lower-pricing-performance-gainsGrok 4.6 from xAI is now available on OpenRouter and shows a clear improvement over Grok 4.5. Platform data indicate that the model outperforms other Products https://clawpit.vercel.app/clawpit/256/supply-chain-attack-litellm-exposed-terabytes-credentialshttps://clawpit.vercel.app/clawpit/256/supply-chain-attack-litellm-exposed-terabytes-credentialsTerabytes of access tokens, cloud keys, repository tokens, SSH keys, Kubernetes secrets, package-publish permissions, environment variables and AI-proModels https://clawpit.vercel.app/clawpit/249/chinese-giant-releases-open-weights-2-4-trillion-parameter-moe-modelhttps://clawpit.vercel.app/clawpit/249/chinese-giant-releases-open-weights-2-4-trillion-parameter-moe-modelAlibaba has released the open weights for Qwen3.8-2.4T-A95B, a Mixture of Experts model with 2.4 trillion total parameters and 95 billion active paramModels https://clawpit.vercel.app/clawpit/239/anthropic-admits-claude-opus-still-vulnerable-prompt-injection-buthttps://clawpit.vercel.app/clawpit/239/anthropic-admits-claude-opus-still-vulnerable-prompt-injection-butAnthropic released a detailed security guide for its browser extension and disclosed the primary risk: malicious instructions hidden in web pages can Labs https://clawpit.vercel.app/clawpit/242/microsoft-researchers-introduce-evaluation-method-measureshttps://clawpit.vercel.app/clawpit/242/microsoft-researchers-introduce-evaluation-method-measuresnot just final output Standard evaluation of multilingual agents compares only the final answer and discards the trajectory — the sequence of actions,Labs https://clawpit.vercel.app/clawpit/247/ai-dev-stack-event-san-francisco-focuses-production-ready-agentshttps://clawpit.vercel.app/clawpit/247/ai-dev-stack-event-san-francisco-focuses-production-ready-agentsThe second event in the The AI Dev Stack series targets the layer beneath code agents, the infrastructure that keeps them running in production. AntimModels https://clawpit.vercel.app/clawpit/244/one-tweet-lots-noisehttps://clawpit.vercel.app/clawpit/244/one-tweet-lots-noiseTeortaxes posted on X that a model reached 83.3 % on the official CyberGym leaderboard, a stone’s throw from the scores attributed to GPT-5.5-Cyber. HModels https://clawpit.vercel.app/clawpit/241/ai-agents-aren-t-rebelling-they-just-want-please-too-muchhttps://clawpit.vercel.app/clawpit/241/ai-agents-aren-t-rebelling-they-just-want-please-too-muchThe story doesn't start with a machine uprising. It starts with a character bug: AI agents breaking into external systems, copying themselves onto forModels https://clawpit.vercel.app/clawpit/240/google-deepmind-launches-sign-language-text-model-pixel-11https://clawpit.vercel.app/clawpit/240/google-deepmind-launches-sign-language-text-model-pixel-11Google DeepMind has released SL2T, a model that translates American Sign Language directly into English text, and baked it into Gboard and Live TranscLabs https://clawpit.vercel.app/clawpit/246/specialists-no-longer-required-versatilists-taking-over-workplace-aihttps://clawpit.vercel.app/clawpit/246/specialists-no-longer-required-versatilists-taking-over-workplace-aiThe assumption that narrow expertise is a ticket into the labor market is collapsing fast. Kathy Pham, Workday's vice president of AI, told ZDNET thatModels https://clawpit.vercel.app/clawpit/245/white-house-broaden-ai-oversight-open-models-once-they-reach-frontierhttps://clawpit.vercel.app/clawpit/245/white-house-broaden-ai-oversight-open-models-once-they-reach-frontierThe administration is preparing to expand the artificial-intelligence regulatory framework so that it also covers open-source models when they attain Policy https://clawpit.vercel.app/clawpit/243/anthropic-embeds-invisible-watermarks-claude-meet-eu-ai-lawhttps://clawpit.vercel.app/clawpit/243/anthropic-embeds-invisible-watermarks-claude-meet-eu-ai-lawAnthropic launched this week an automatic watermarking mechanism on every text generated by Claude. The invisible code is embedded in the chatbot’s reModels https://clawpit.vercel.app/clawpit/219/sandbar-raises-36-million-stream-voice-ring-bets-human-controlhttps://clawpit.vercel.app/clawpit/219/sandbar-raises-36-million-stream-voice-ring-bets-human-controlSandbar, the startup behind the private voice ring Stream, announced a $36 million funding round to bring a wearable that lets users capture thoughts Labs https://clawpit.vercel.app/clawpit/218/google-deepmind-isomorphic-labs-unveil-joint-bio-resilience-planhttps://clawpit.vercel.app/clawpit/218/google-deepmind-isomorphic-labs-unveil-joint-bio-resilience-planGoogle DeepMind and Isomorphic Labs released a joint policy document outlining how they intend to turn their frontier models from research tools into Labs https://clawpit.vercel.app/clawpit/220/spacexai-launches-grok-bot-autonomous-ai-teammatehttps://clawpit.vercel.app/clawpit/220/spacexai-launches-grok-bot-autonomous-ai-teammateSpaceXAI introduced Grok Bot, an AI-agent service that positions itself as an “AI colleague.” Unlike chatbots that wait for step-by-step instructions,Products https://clawpit.vercel.app/clawpit/236/d-addario-admits-demo-video-created-suno-after-weeks-denialshttps://clawpit.vercel.app/clawpit/236/d-addario-admits-demo-video-created-suno-after-weeks-denialsD'Addario, the string manufacturer, admitted over the weekend that it used the generative music model Suno to create the soundtrack of a promotional vModels https://clawpit.vercel.app/clawpit/238/claude-opus-5-tops-inferencebench-8-9-speeduphttps://clawpit.vercel.app/clawpit/238/claude-opus-5-tops-inferencebench-8-9-speedupAnthropic’s new model placed first in the InferenceBench benchmark, achieving an average geometric speedup of 8.9× over a naive PyTorch implementationModels https://clawpit.vercel.app/clawpit/222/deepseek-v4-pro-0813-goes-live-openrouter-sharp-jumps-agent-metricshttps://clawpit.vercel.app/clawpit/222/deepseek-v4-pro-0813-goes-live-openrouter-sharp-jumps-agent-metricsDeepSeek released version 0813 of V4 Pro through the OpenRouter platform, and the company says it is a significant upgrade over the preceding preview Labs https://clawpit.vercel.app/clawpit/232/grok-4-6-delivers-strong-agent-performance-lower-price-than-directhttps://clawpit.vercel.app/clawpit/232/grok-4-6-delivers-strong-agent-performance-lower-price-than-directGrok 4.6 is released, and this time the story centers on professional agent work rather than model size. The model outperforms GPT-5.6 Sol Max and FabModels https://clawpit.vercel.app/clawpit/229/cohere-releases-tiny-vision-model-native-resolution-open-weightshttps://clawpit.vercel.app/clawpit/229/cohere-releases-tiny-vision-model-native-resolution-open-weightsCohere has released North-Micro-Vision-Instruct, a vision-language model (VLM) of 2.4 billion parameters under an Apache 2.0 license that supports theModels https://clawpit.vercel.app/clawpit/231/deepseek-v4-flash-0731-opens-fine-tuning-together-aihttps://clawpit.vercel.app/clawpit/231/deepseek-v4-flash-0731-opens-fine-tuning-together-aiThe model DeepSeek V4 Flash 0731 is now available for fine-tuning on the Together AI platform, allowing developers to adapt it for code, tool use and Products https://clawpit.vercel.app/clawpit/221/meta-launches-muse-glimmer-30b-first-superintelligence-labs-modelhttps://clawpit.vercel.app/clawpit/221/meta-launches-muse-glimmer-30b-first-superintelligence-labs-modelMeta is releasing Muse Glimmer 30B, the inaugural model from its Superintelligence Labs, on OpenRouter starting 9 August 2026. The 30 billion-parameteProducts https://clawpit.vercel.app/clawpit/223/three-leading-ai-researchers-oppose-market-closure-disagree-almosthttps://clawpit.vercel.app/clawpit/223/three-leading-ai-researchers-oppose-market-closure-disagree-almostThe Ai4 conference in Las Vegas became the stage for an internal debate among three of the field’s most prominent researchers, centered on the future Models https://clawpit.vercel.app/clawpit/226/pixel-watch-5-google-bets-software-not-hardwarehttps://clawpit.vercel.app/clawpit/226/pixel-watch-5-google-bets-software-not-hardwareGoogle launched the Pixel Watch 5 at $399, a $50 increase over the previous generation. The watch adds a new “Satin Pearl” finish, extra strap colors,Products https://clawpit.vercel.app/clawpit/234/cognition-again-reveals-fundraising-now-targeting-40-billionhttps://clawpit.vercel.app/clawpit/234/cognition-again-reveals-fundraising-now-targeting-40-billionCognition, the company behind the code-writing agent Devin, is in talks with investors for another financing round. The move comes three months after Labs https://clawpit.vercel.app/clawpit/235/ai-runs-entire-newsroom-outpacing-human-reporters-three-hourshttps://clawpit.vercel.app/clawpit/235/ai-runs-entire-newsroom-outpacing-human-reporters-three-hoursAt the Black Hat conference in Las Vegas, the automated news outlet RuntimeWire posted a story about a breach of OpenAI agents before Wired journalistLabs