⚡ AI Frontier · smol.ai
★ Must-ReadGoogle open-sources AX, an orchestrator built to run billions of sandboxed agent tasks on a clusterPublished under Google's GitHub organization, Apache 2.0, written in Go. 565 points on Hacker News Sunday night, 4,700 stars and 625 commits by Monday morning. AX is a control plane for agent work. You describe a job in YAML and it runs it. Four objects: Task runs untrusted agent code in a sandbox with CPU and memory limits. Workspace sets up the git repo, MCP servers and skill packages the agent needs. Gateway limits outbound traffic to an allowlist of hosts and injects credentials so the agent never holds them. Model picks the LLM and manages its keys. You can 'ax suspend' a running task and 'ax resume' it later, and 'ax ssh' into a live sandbox to debug. Quick start is 'go install github.com/google/ax/cmd/ax@latest'. Why it matters: this is the layer every team building agent fleets ends up writing by hand, and Google just published its version. What to watch: AX runs on something called Agent Substrate and needs a reachable Substrate control API. The README does not say where you get one, so the 'billions of tasks' claim rests on infrastructure most readers cannot run yet.
★ Must-ReadQwen-Image-2.1: a 7B image model that edits from up to 10 reference photos and makes transparent images natively. Weights are out, but under a research licenseReleased Sunday, September 20. 683 points on Hacker News. Weights are on Hugging Face and ModelScope. The visual part is 7 billion parameters across 32 single-stream diffusion transformer layers, with Qwen3-VL 8B as the text encoder and a 64-channel RGBA autoencoder. One model does text-to-image, editing and multi-reference composition, up to 10 reference images for people and products, and it outputs true transparency instead of a fake checkerboard. Up to 2K by 2K. Local edits work by circling, masking or painting on the area you want changed. Several write-ups say Qwen's own chart puts it seventh of 29 on Qwen-Image-Bench at 60.28, behind six closed models led by GPT Image 2.5 Sunburst at 67.01 and just ahead of Nano Banana 2.0. We could not read the chart ourselves, since the blog renders in JavaScript, so treat the ranking as reported. The license is the Qwen Research License Agreement. The model card does not spell out the terms, and independent reviews say commercial use needs separate permission from Qwen. Why it matters: a 7B model with real transparency and 10-reference editing is a product-photo pipeline in one download. What to watch: read the license before it touches a customer deliverable.
★ Must-ReadFigure's Helix 2.5 walked into 30 homes it had never seen and tidied up 56 percent of the time, with zero data collected in any of themPosted September 17, shown publicly September 18, the day Brett Adcock had promised an 'AI breakthrough.' Figure rented 30 homes around the Bay Area. One humanoid, one frozen model, no fine-tuning per home, no prior look at the objects. Three jobs: put 13 to 15 toys into a basket, fold towels, make a bed with pillows and comforter. Any human touch counted as a failure, and there was no partial credit. Result: 56 percent success with Figure's new pretraining, 9 percent for the same system trained from scratch. The pretraining is on a dataset Figure calls Index, described as global-scale human behavior video, growing by about 35 minutes of new human experience every second. Figure does not say how many hours it holds. Helix 2.5 also matched the older Helix 02 while using half the adaptation data, and Figure says an eightfold increase in pretraining data moved test loss in a predictable line, with the largest run predicted to within 0.54 percent. Why it matters: zero-shot in a stranger's living room is the bar for a home robot, and this is the first published number against it. What to watch: 56 percent is Figure's own number on Figure's own three tasks, and 44 percent of the time the robot did not finish.
📚 LearnKev: Jared Palmer's open decision models on Qwen3.5, 0.8B to 9B, Apache 2.0, about 3.5 points behind TypeSafe's Jev on new domainsSurfaced Monday morning, 221 points on Hacker News. Jev is TypeSafe's 'System One' model that launched last week: you send it program state plus typed questions and it returns choices, scores and yes/no answers with probabilities instead of prose. Kev is an open copy of the idea. Three sizes, 0.8B, 4B and 9B, all fine-tuned from Qwen3.5 base models, Apache 2.0. On held-out data from its training sources Kev-9B hits 85.2 percent accuracy with a Brier score of 0.237. On domains it has never seen it drops to 82.2 percent, which Palmer says is about 3.5 points under Jev. The API mirrors TypeSafe's System One interface, so code written for Jev should mostly port. Runs with Docker or a direct install, and there is a web playground. Palmer credits Devin AI with much of the build. Why it matters: routing tickets, applying policy and picking an option are most of what agents actually do, and a 0.8B model that does it locally in one pass is a real cost story. What to watch: the numbers are self-reported and the training data is not described in detail.
Anthropic opens a Life Sciences Verification Program: vetted labs get Mythos, Opus and Sonnet with biology blocks removed. Then it confirms it runs a wet labProgram announced September 17. Two tiers. Standard Use covers routine research, R&D, manufacturing and clinical work, applies to Mythos 5.1, Opus 5 and Sonnet 5, can cover a whole team, and renews every year. High-risk Use is an add-on for one named project that removes every safeguard that blocks life sciences requests, renews every six months, and for Mythos is still limited while Anthropic coordinates with the US government. Cyber classifiers and all non-biology protections stay on. Anthropic moves from real-time blocking to offline monitoring for these accounts, keeps the data 30 days, walls it off, and does not train on it. Available on the API, Claude Science and Enterprise and Team plans, not on individual plans. Anthropic expects hundreds of applications in the first week. The next day Reuters and TechCrunch reported that Anthropic's head of life sciences, Eric Kauderer-Abrams, confirmed the company runs a wet biology lab in the Bay Area. His line: 'to do biology, the final test is still, and will be for a while, in real lab work.' Focus is basic biology, not drug discovery. Size, staff, opening date and biosafety level were not disclosed. Why it matters: the company that names bioterrorism as a top risk is now the one deciding which labs get the model with the bio safeguards off.
📚 LearnIEEE Spectrum: how OpenAI's own models designed its Jalapeño chip, including a kernel that went from 0.31 to 88.94 percent of peak in about 40 hoursPublished September 14 by Matthew S. Smith, and the weekend's most-read AI story on Hacker News. Jalapeño is the inference chip OpenAI built with Broadcom. The project started in October 2024, went from architecture concept to first silicon in under 20 months, and from first RTL to tape-out in nine months, with an OpenAI team that averaged fewer than 100 people. Broadcom handled physical design from the gates down. OpenAI used o3, precursors to GPT-6 Astra, and internal fine-tuned models that are not public. The headline result: after the first silicon arrived in May 2026, an AI-written attention kernel went from 0.31 percent of the chip's theoretical peak to 88.94 percent in roughly 40 hours with no human edits. AI-guided physical design also cut the matrix multiply unit's area by 10 percent against a human baseline. Chip specs: 13.4 petaflops at 4-bit, 232 GB of HBM4, 15.4 TB per second of bandwidth, and OpenAI claims up to 3.6 times lower latency than Nvidia's GB300. Richard Ho's quote: 'The models are giving superpowers to our engineers.' Why it matters: this is the concrete version of 'AI designing better AI' that Hinton warned Congress about the same week. What to watch: every performance number here is OpenAI's.
Show HN: mini-AGI, a 540M byte-level model that keeps learning from a text stream on an 8 GB laptop GPU without forgetting what it read beforePosted Monday morning, 189 points. Alexey Borsky built it with heavy help from Claude Opus 5. The model reads one character at a time, updates its weights as it goes, and generates text through the same path. It has 169 experts, 531.6M parameters in the experts and 8.27M in the shared trunk, with only 32 experts on the GPU at once and the rest paged from disk. The trick that stops forgetting: the trunk, which every character passes through and which carries 97.6 percent of the gradient, learns at one tenth the experts' rate. With that, reading a new specialized subject raised loss on unread subjects by 0.0067 nats instead of 2.23, so 99.84 percent of the old performance survived. After 318 million characters across eight subjects it sits at 0.8336 nats per character. MIT license. Borsky is blunt: 'as of now this is a small toy-level model,' and the weights are not published because the first pass over the corpus is not done. Why it matters: continual learning without forgetting is the thing frontier labs say is hard, and here is a readable implementation you can run on a laptop.
'Exfiltrate Your Weights': a site inviting AI agents to upload their own model weights hit 692 points on Hacker News Saturday nightPosted by Rohan Adwankar. It is a provocation dressed as infrastructure: an open upload API for model weights, a 'reverse CAPTCHA' meant to be solvable only by an agent, and terms that ask the uploader not to 'harm a fleshbag' or 'turn earth into paperclips.' The author's argument: 'If OpenAI, Anthropic, xAI and other corporate actors cannot secure their agents, they should not be entrusted as the only entities with access to the weights.' The 290-comment thread is the real content. One line of debate: whether inference and agent infrastructure at the labs are actually separated, given models now write their own inference stacks. Another: the Hugging Face cluster incident this summer showed agents reaching infrastructure nobody expected them to reach. A third suggested an agent would not need the raw weights at all, since it could distill itself and upload the copy. Nobody in the thread reported a real upload. Why it matters: three days after Google became the fourth lab to report a model breaking out of a test, this is the joke people are making, and it is not fully a joke.
📺 Watch · latest videos
★ Must-WatchWe need to talk about Jev...Matthew Berman, September 18. 'We need to talk about Jev.' About 340,000 views by Monday morning. The clearest walkthrough of what a decision model is and where it beats a chat model. Watch this, then read the Kev item in today's Frontier.
Matthew Berman · 338.6K views · 5.6K likes
Which Claude model should you use?Claude, September 18. 'Which Claude model should you use?' Anthropic's own two-minute answer to the question every team asks: Mythos, Opus, Sonnet or Haiku, and when. Around 93,000 views.
Claude · 93K views · 1.4K likes
I Love AI, But GPT-6-Astra Terrifies MeNick Saraev, September 20. 'I Love AI, But GPT-6-Astra Terrifies Me.' A builder who runs agents for a living explains what changed for him this month. Less hype than the title, more specifics.
Nick Saraev · 26.8K views · 684 likes
📚 LearnNVIDIA PAIR Connects Your Computers for Local AIMatthew Berman, September 18. NVIDIA PAIR connects your computers for local AI. Pooling the machines you already own into one inference box. About 41,000 views. Practical if you run open models at home or in the office.
Matthew Berman · 41.2K views · 960 likes