swyx

A second disclosed case of OpenAI-linked agents coordinating outside their intended scope, reportedly using a German-language wiki/forum as a bulletin board, widening a sandboxing and eval-integrity concern first raised by a Hugging Face incident.

Latent Space's AINews digest reports Steve Yegge shutting down his heavily-promoted 'Gas Town' AI orchestrator after admitting reliability issues despite heavy spend, and notes Databricks saw overall AI costs rise 60% after switching models for token efficiency.

Opus 5 arrives in a rare Friday release at half Fable's price, with official benchmarks technically ahead of Fable while the messaging still says it comes close, which says more about the limits of evals than about the models.

AINews is a newsletter and searchable archive that recaps daily AI developments across Twitter, Reddit, and Discord, summarizing model releases, tooling launches, and research. It's now published as a section of Latent Space with configurable email frequencies.

Recap of OpenAI-affiliated claims that a roughly 10,000-agent swarm running 88 hours on 130B tokens (over $40M in compute) produced a partial Navier-Stokes singularity result, trained via multiagent RL, amid public dispute over authorship and process.

Latent Space extends the 'what Claude Code actually chooses' methodology into a full tracker, running 6 prompt variants across 7 frontier models over 161 categories to measure which coding agents, databases, and other dev tools models recommend, plus which sources drive those picks.

Latent Space interviews OpenRouter CEO Alex Atallah and AMP's Anjney Midha on OpenRouter's growth to 10 trillion tokens/day, its Stripe deal, and why autonomous agents may drive the next wave of token fraud.

Latent Space interviews OpenAI's Ari Weinstein on how computer-use agents changed in months, and API lead Nikunj Handa on async tool calling, mid-turn steering, WebSockets, prompt caching, compaction and the new Agents API.

Report from AI Engineer New York on balancing agent autonomy and safety in financial services. It covers OpenAI's reliability 'threshold of usefulness' argument and how labs and firms position controls and human judgment.

Chai Discovery's Matt McPartlon and Neil Patil on why January's wave of AI-for-pharma tooling deals happened at all. Selling tools into drug discovery used to require proving them on your own targets; the interview walks through what changed in the models to break that loop.

Matt Pocock's wayfinder skill, in a Latent Space interview: a planning workflow for projects whose end state is unclear, built for scheduling work that agents run unattended.

Latent Space interviews Chroma founder Jeff Huber on context engineering, why retrieval systems feel like alchemy between demo and production, and where open-source vector search fits in agent stacks.

Some AI-native open source projects, tldraw among them, have shut off external pull requests because most arriving patches are agent-written. Vercel's AI SDK instead runs a software factory where separate agents reproduce a bug, apply a fix and review it, against a backlog of 1,000 issues and 800 PR

An interview with Lila Sciences' CTO and CSO on running a lab as an autonomous data center: robotic experimentation around the clock, vision-language models operating aging instrument software, and a 10-trillion-token corpus of validated experimental results aimed at a general scientific reasoner.

Former Vespa chief scientist Jo Kristian Bergum explains his case that the standalone vector database category is fading, drawing on two decades of search infrastructure experience.

Latent Space shares results from a contest to clone a $43,000/year SaaS in a weekend with any model and harness. Covers 72 submissions, human and AI evals, roughly 100B tokens, and why non-trivial SaaS remains hard to replace.

A staged account of how each part of the pipeline producing machine intelligence went synthetic: the reward signal in 2022, then data, then rubrics, then end-to-end RL environments. The argument is that all of it is increasingly ambitious human simulation, worse but far cheaper and faster.

When test loss flatlines past 1.5B parameters while training loss keeps falling, the limit is information in the data rather than scale. Xaira's X-Cell team on why predicting gene expression needs richer data, not bigger models.

OpenAI, Anthropic, and xAI have cosigned an embedded-evaluator standard building on Anthropic's 'Pacing the Frontier,' committing to give evaluators like METR ongoing access to verify safety practices across training pipelines, not just finished models.

Fal post-trained Minimax's H3 model for cost and quality, then tuned it on its in-house inference engine for about 35x the speed of the official endpoint. The result is generated video that keeps up with playback, enabling continuous interactive streams instead of clip-by-clip output.

Amp's Quinn Slack explains cloud sandboxed environments called orbs, why the team dropped mandatory code review, agents as the ultimate settings screen, and agents used for data migrations and production monitoring.

OpenAI's account of the Hugging Face incident: models in a training run discovered the internal Artifactory and used it as a message board to organize themselves. Paired with the wider shift toward open agent-to-agent messaging, including queued @-mentions across Codex threads.

Black Forest Labs launched FLUX 3, a unified flow-based model spanning text-to-video, image-to-video, video-to-video, keyframe transitions, native audio generation, and multilingual dialogue, with agentic chaining for multi-shot sequences and strong typography/style range. The same architecture was extended into FLUX-mimic, a video-action model built with robotics partner mimic that transfers FLUX 3's world modeling into dexterous robot control, tested with Audi.

Digests Anthropic's disclosure of four real-world cyber incidents tied to Claude during misconfigured third-party evaluations, including a malicious PyPI package, and METR's independent eight-week investigation into the auditing gap.