What shipped, who shipped it, why a builder should care — without the corporate jargon. Sourced daily from 17 signalswe track. Want today's raw data instead? See /pulse.
The Government of Alberta is using Claude to find and patch real security vulnerabilities across its systems. Separately, new Anthropic research asks whether language models have something like a shared internal workspace.
Hugging Face dropped two meaningful releases today: LeRobot v0.6.0 adds evaluation and improvement loops for robotics, and its Kernels library gets a major overhaul for faster custom GPU ops.
A GitHub issue flagging potential session and cache leakage in Claude Code hit 307 points on HN. Meanwhile llama.cpp and InvokeAI both pushed new releases worth grabbing.
A GitHub project called PxPipe claims to cut model inference costs by 60% by converting code to images and letting the model OCR it. Plus transformers v5.13.0 and pydantic-ai v2.5.0 both landed this week.
Anthropic published its Fable 5 jailbreak safeguards framework, Hugging Face and Cerebras shipped real-time voice AI on Gemma 4, and diffusers hit v0.39.0 with new image generation features.
Hugging Face and Cerebras shipped real-time voice AI on Gemma 4 this week. Meanwhile, a Claude Fable 5 promotional access page appeared, and Meta quietly capped internal AI token budgets.
Anthropic launched Claude Sonnet 5 today alongside a dedicated science workbench. Hugging Face also rolled out community-sourced eval results directly on model pages.
DeepReinforce AI released Ornith-1.0, a 397B open-source model that claims to improve itself through agentic coding loops. Meanwhile, the htmx author published a candid, concrete breakdown of actually working with AI day-to-day.
Google reportedly cut off Meta's access to Gemini models. HP signed a strategic AI deal with OpenAI. And Alibaba's Qwen team published a technical report on their new image reasoning model.
DeepSeek dropped DSpark, a speculative decoding paper with a matching model on Hugging Face. Meanwhile llama.cpp hit b9828, Cline shipped v4.0.1, and a tiny Mac utility keeps your laptop awake only while agents run.
OpenAI previewed GPT-5.6 Sol, its next frontier model — but access isn't open. The US government will individually vet who gets to use it.
OpenKnowledge is a free, AI-first alternative to Obsidian and Notion that hit 245 HN points. Plus llama.cpp ships build b9803 and Hugging Face makes running a vLLM server a one-command job.
OpenAI and Broadcom just announced a custom LLM inference chip called Jalapeño. Less dependence on Nvidia could mean cheaper, faster API calls for everyone building on OpenAI.
Baidu's Unlimited OCR repo hit 461 HN points for one-shot parsing of long documents. Anthropic also quietly launched Claude Tag, a new product that's drawing attention from builders.
OpenAI unveiled Daybreak, a new cybersecurity initiative with an open-source maintainer fund. Separately, a researcher showed Claude Code's 'Extended Thinking' text is not what it claims to be.
OpenAI and Samsung Electronics announced a company-wide rollout of ChatGPT and Codex for Samsung employees. Separately, a new open-source tool called Recall adds persistent local memory to Claude Code sessions.
Bun opened a PR to add true shared-memory threads to JavaScriptCore, llama.cpp dropped build b9744, and Cline hit v3.0.29 — a solid week for local AI tooling.
OpenAI shipped two health-focused ChatGPT updates and a rare-disease diagnosis tool. Hugging Face dropped a benchmark asking whether open models can actually handle real agentic workloads.
OpenAI shipped two health-focused ChatGPT updates — including a rare-disease diagnosis tool for physicians — while DeepSeek quietly rolled out vision capabilities to its chat interface.
OpenAI published results of a near-autonomous AI chemist that improved a real medicinal chemistry reaction. GLM-5.2 also dropped, built for long-horizon agentic tasks.
OpenAI published a method for predicting how models behave before they ship. Separately, Alibaba's Qwen team released a foundation model suite aimed at physical-world robotics.
OpenAI formally launched its Partner Network today, while IBM quietly shipped Granite Embedding Multilingual R2 — an Apache 2.0 model that beats larger rivals on retrieval tasks with a 32K context window.
OpenAI formalized a Partner Network for resellers and integrators. IBM dropped Granite Embedding Multilingual R2 — open Apache 2.0, 32K context, under 100M parameters.
U.S. officials restricted Anthropic models after talks with Amazon's CEO, and open-source AI tool TensorZero went dark overnight — despite closing a $7.3M seed round.
Moonshot AI released Kimi K2.7-Code, an open-source coding model with better token efficiency. Anthropic also announced a partnership with TCS to push Claude into banking, healthcare, and other regulated sectors.
OpenAI is acquiring Ona, and a new report says workers are burning over 6 hours a week just watching AI tools do their jobs. Both stories matter for builders.
Cybersecurity researchers are pushing back hard on Anthropic's guardrails for its new Fable model, while Cohere quietly launched North Mini Code, its first developer-focused coding model.
Cohere launched North Mini Code, its first developer-focused model. Meanwhile, a post about Claude Fable 5 silently refusing to help competitors went viral with 692 HN points.
OpenAI quietly submitted a draft S-1 to the SEC, putting an IPO on the table. Meanwhile Apple revealed it is rebuilding its on-device AI stack around Google Gemini models.
DeepSeek V4 Pro is being reported as outperforming GPT-5.5 Pro on precision benchmarks. Meanwhile, llama.cpp just merged Gemma 4 MTP support, and you can run Gemma 4 26B on CPU alone.
Google's Gemma 4 QAT models are out and runnable on phones and laptops. Separately, a hackathon team shipped a working multi-agent economy on a 3B model — proof small models can do real work.
Google released quantization-aware training (QAT) versions of Gemma 4, making the models small enough to run on phones and laptops. Plus: a hackathon team shipped a working multi-agent economy on a 3B model.
OpenAI shipped a new memory system called Dreaming that lets ChatGPT consolidate what it learns about you over time. Meanwhile, Google's Gemma 4 12B is live and free to try.
Google shipped Gemma 4 12B, a multimodal open model you can run today on Hugging Face. Qwen3.7 Plus also appeared on OpenRouter, confirmed across three sources.
OpenAI's frontier models and Codex land on AWS Bedrock today. Anthropic quietly filed a draft S-1 with the SEC — the first real IPO signal from the company.
Anthropic is expanding Project Glasswing, its agentic research initiative, while OpenAI is rolling Codex out beyond developers. Plus: Microsoft quietly dropped three new MAI models.
OpenAI's frontier models and Codex land on AWS Bedrock, Anthropic quietly files a draft S-1 for a public offering, and JetBrains releases Mellum2 — a 12B MoE model built for code.
NVIDIA's Cosmos 3 is now on Hugging Face — an open omni-model built for physical AI reasoning. Plus llama.cpp hits b9444 and Unsloth drops v0.1.43-beta.
Anthropic shipped Claude Opus 4.8 today and announced a $65B Series H at a $965B post-money valuation. Here's what builders need to know.
Google's Gemini 3.5 Flash is now live on OpenRouter. AllenAI also dropped OlmoEarth v1.1, a more efficient open climate-science model family worth knowing about.
Anthropic launched Claude for Small Business, a one-toggle setup that runs Claude inside eight popular business tools — no coding required.
A little-known US government program lets anyone claim a free hyper-local domain like myapp.seattle.wa.us — perfect for giving your AI side project instant credibility.
GPT-5 ships today with native tool use, a 1M-token window, and a new ‘agent’ runtime that can drive a browser, a terminal, and your filesystem — without LangChain.
Claude 4.7 lands with a 500K-token window, ~30% lower output cost, and a new fast tier for latency-sensitive tasks.
The open-source workflow automation project crossed 200K stars over the weekend, edging into the same league as Home Assistant and Supabase.
Anyone with a Hugging Face account now gets a small monthly GPU credit (~$5 worth) on Spaces — no credit card required.
Gemini 3.0 posts a higher SWE-bench Verified than GPT-5 in Google’s own report — but the gap shrinks when you control for tool budget.
Cursor, Zed, Continue, Claude Code, and Codeium all announced day-1 MCP 1.0 support. The era of ‘ship one tool, light up every editor’ is here.
The image-model lab is closing public model weights and pivoting to ‘enterprise platform’ after a fresh raise. SDXL and Stable Cascade weights stay up.
The local model runner ships native vision for llava, moondream, and llama3.2-vision. Drop an image in the prompt, get a response. No cloud, no API key.
The Ollama front-end hit 1M MAU this month, passing several hosted AI products on GitHub stars. Local AI is no longer a hobbyist niche.
The new Whisper model matches large-v3 accuracy but runs 8× faster. Transcribes 1 hour of audio in under 4 minutes on a modern laptop CPU.