All AI news

AIBoris Cherny says he used Opus 5.5 with Lean to formally verify the Claude Agent SDK, with a couple of short prompts producing 16 PRs fixing bugs and race conditions. He also reports that TLA+ works well, sometimes combined with Lean to find data flow, concurrency, and state management issues. The post links to his actual prompts as another example.

AIAnthropic's Boris Cherny says Claude works best when prompted like a coworker, with a clear goal and no heavy scaffolding. He says prompts now matter most for stating the desired task, the effort level, and how Claude should verify its result, whereas prompts mattered more in the Sonnet 3.5 era.

AIElvis Saravia argues that building good evals on top of AI systems can put a practitioner at the frontier of their domain or task quickly. He advises readers to learn eval construction, calling it worth the investment. The post responds to Garry Tan's point that agents writing markdown skills on cron jobs can handle most knowledge work.

AICline says users can now run security-related tasks in Cline that were previously impossible with GPT or Claude. The post gives install commands for the CLI via npm and a link to the desktop app.

AINVIDIA's technical blog describes bitwise determinism for large-scale pretraining with Megatron Core, which makes training runs easier to debug, validate, and resume reproducibly. The source says these benefits matter most for models with trillions of parameters trained across thousands of GPUs, where multiple parallelism dimensions, low-precision computation, and distributed checkpointing complicate failure reproduction and fix validation.

AIGoogle DeepMind has released EmbeddingGemma 2, an open 740 million parameter model that maps text, images, audio, and video into one embedding space. It is built on the Gemma 4 architecture under an Apache 2.0 license and supports an 8K token context window. The company reports a code benchmark gain from 68.76 to 78.68 on MTEB Code and says the model can run on-device with about 567MB of active RAM for the full multimodal version on a Google Pixel 11 Pro.

Why it matters: The release shows how a 740M-parameter embedding model can cover text, code, images, audio, and video on local hardware, with memory and storage figures to compare against other on-device options.

AIScale AI has released AgentEnv, an open source framework for building high-fidelity reinforcement learning environments. Modal says its Sandboxes can deploy these environments in seconds at the scale RL rollouts require.

AISome U.S. energy companies are trying to generate more electricity from existing nuclear reactors rather than wait for new plants, which take a long time to build. The source does not provide specific figures, timelines, or company names.

AIAn Arizona appellate court ruled that an AI-generated video of manslaughter victim Christopher Pelkey carried undue emotional weight and ordered the defendant resentenced. The video, scripted by Pelkey's sister Stacey Wales, was shown at sentencing, where the judge said he loved it and imposed the maximum 10.5-year term. The court found that the AI video, unlike photographs in State v. Rose, does not reflect actual events and rendered the sentencing procedure fundamentally unfair.

AINathan Lambert says he is increasingly convinced the US needs to legalize distillation, since banning it is impossible. He warns that otherwise strategic competition will become untenable within a few years. The quoted context notes that Chinese labs are now outpacing Western counterparts such as Reflection, Nvidia, and Thinking Machines on model strength.

AIOpenAI says it removed some friction for developers building on its API, while the main post gives no specific changes. Background from @OpenAIDevs says the five paid usage tiers become three—Build, Launch, and Grow—and Grow, the new highest tier, requires $500 in total API payments, down from $1,000 for the previous top tier.

AINVIDIA's DOCA GPUNetIO lets GPU applications control networking and data movement directly, rather than routing each transaction through the CPU. The source says host-driven network handling adds latency on the critical path and limits how quickly distributed applications can respond in real time. The provided text is truncated, so details of the unified software stack are not available.

AIOn a livestream, Arthur Soroken interviews Amir and Ali Sadeghian, co-founders of Astrocade, about how generative AI and Gemini multimodal models are reshaping game development.

AIGoogle Gemini's Guided Vision supports everyday tasks such as reading fine print, finding objects, describing items, and exploring one's immediate surroundings. The feature is also presented as practical visual assistance for older adults, people with low literacy, and anyone reading in low light.

AIWith Guided Vision turned on, Gemini Live verbally prompts users to pan, tilt, or step back when their camera framing is too high, too close, or off to the side. These prompts help Gemini get the clear visual context it needs to answer questions.

AIGuided Vision in Gemini Live is now available on Android devices running Android 9 and above, where Gemini Live is supported. Users need the latest Gemini app and Android system software to get started.

AIA 9B model trained with SFT and prompt repetition at about $25 of compute reaches roughly 79% of Jev's Decision Index score. It ranks second among all models at 10B parameters or below and 14th of 70 overall.

AIAnthropic is expanding its Cyber Verification Program to give verified security professionals broader access to its most capable models. Through the program, they can use Claude Mythos 5.1, Opus 5.5, and Sonnet 5.5 with safeguards designed for defensive work. New tiers will also allow authorized offensive work such as penetration testing and red-teaming.

AIClaude Code v2.1.292 adds a --marketplace option to claude plugin install, which adds the marketplace if needed and then installs the plugin from it. The release also adds an effort parameter to the Agent tool and fixes several security issues, including permission prompts bypassed for network (UNC) file reads and a sandboxed read path that could return files outside approved access.

AILangChain is promoting a LangChain Academy course that teaches developers how to build long-running agents for complex workflows using Deep Agents. The post links to the course page and invites readers to enroll today.

AIAmjad Masad shared a link to a full conversation, with no further details provided in the post itself. The linked video is YouTube content, and the post offers no further specifics about its subject or claims.

AIOpenRouter now offers google/gemini-nano-banana-2.1, which accepts up to 14 reference images and supports Grounding with Google Search. The model produces cleaner 1:4, 4:1, 1:8 and 8:1 panoramas at 2K and 4K. Pricing is $0.0336 per image at 1K, $0.0504 at 2K and $0.1134 at 4K.

AIAnthropic's Lydia Hallie says users can still claim a one-time bonus credit for Claude Code cloud sessions until October 7 by running /claim-credit. Cloud sessions run Claude Code on a fresh VM per task, letting several run at once even after the laptop is closed.

AIAnthropic's ClaudeDevs account says Pro or Max plan subscribers on September 23 can still claim a one-time bonus credit for cloud sessions by running /claim-credit in Claude Code by October 7 at 11:59pm PT. Cloud sessions draw on this credit first before counting toward plan limits, and the credit expires November 4.

AIAnthropic's ClaudeDevs says Claude Code cloud sessions run each task on a fresh VM, so users can start several at once. The sessions keep running after the user closes their laptop. A field guide covers seven suitable workflows and how to connect GitHub.

AIPika announces that users can try Nano Banana 2.1 through its create.pika.art app page. The post provides only a link and no details on features, pricing, or capabilities.

AIA Hugging Face Space lets users try EmbeddingGemma 2 directly in their browser using WebGPU. The post shares only a demo link, with no further details on model size, benchmarks, or pricing.

AIGitHub has released ReviewBench, a new open benchmark for evaluating AI code review agents, shaped by analysis of 103.9M GitHub pull requests. It contains 219 PRs across 19 languages, and developers can bring their own code review agent, evaluate it, and submit their results.

AIGoogle releases EmbeddingGemma 2, its first native multimodal embedding model, built on Gemma 4 under Apache 2.0. It embeds over 100 languages, code, images, audio, and video into one vector, with an 8,192-token context and four sizes from 270M to 740M parameters. Matryoshka output dimensions of 768, 512, 256, or 128 are supported, and the model is available in Sentence Transformers and LiteRT-LM, with a reported 14% gain on MTEB Code.

Why it matters: The release extends an embedding model to text, code, images, audio, and video in one vector, a useful option for retrieval systems that mix media types.