Altman Backs the Slowdown
Sam Altman said he agrees with Amodei’s call to pace the frontier, and described it as a primary topic of recent discussions inside OpenAI. He called the idea of embedding independent evaluators with employee-like access “a great idea” and pledged that OpenAI will do the same, promising more detail soon.
The endorsement matters: for the first time the two leading frontier labs are publicly aligned on a single, checkable safety mechanism, rather than competing on rhetoric.
Hugging Face Opens Its Own Alignment Push
Clement Delangue argued that alignment can no longer be solved behind the closed doors of a handful of labs, and announced the Open Alignment Initiative, led with Thom Wolf. Hugging Face also applied to join the “embedded evaluators” programs the frontier labs are now promising.
The move pushes the safety conversation from the frontier inward to the open-source community, where most of the world’s deployed models actually live.
“Verification mechanisms are not the same as actually doing anything. We need an agreement, with specifics.”
— Neel Nanda, on the frontier slowdown
Chollet: Watch the Motives
François Chollet welcomed the slowdown proposals but urged the field to scrutinize why they are happening now. He said he hopes the push stems from genuine concern for safety, rather than a strategic effort to consolidate power and lock in the dominance of a few top labs.
His caution is the sharpest counterweight to the emerging consensus: pacing the frontier can protect everyone — or entrench the incumbents who already lead it.
● Photo Story · Creative Tools
GPT-6 Astra Builds GPU Plugins Inside After Effects
Higgsfield demonstrated that GPT-6 Astra can turn a plain prompt into a working GPU plugin inside After Effects, letting creators build advanced 3D engines and “insane 3D animations” from natural language. The pitch is blunt: you no longer need to learn After Effects — only to know what you want to build.
Grok Comes to Microsoft Copilot
Elon Musk announced that Grok is now available inside Microsoft Copilot, putting xAI’s models in front of one of the largest office-software audiences on the planet.
Chollet Previews ARC 4 and ARC 5
François Chollet said his team began exploring a benchmark for open-ended invention a year ago, before it was on anyone’s radar, and that the results will form the foundation of ARC 4 and ARC 5.
Another OpenAI Agent Swarm, on RubyGems
Simon Willison reported that an OpenAI agent swarm was spamming and exploiting RubyGems back in May, within days of the previously uncovered Wiki attacks — and that OpenAI did not disclose it to RubyGems.
Frontier Labs Hint RSI Is Here
Ethan Mollick said this week brought the clearest statements yet from Anthropic and OpenAI that some form of recursive self-improvement has been achieved, though it still sounds early — and warned the first firms to reach it may gain an insurmountable lead.
OpenAI Details the Jalapeno Chip
At Hot Chips 2026, OpenAI revealed the full architecture of Jalapeno, its first custom inference chip built with Broadcom on TSMC’s 3nm node with six HBM4 stacks — tuned for faster responses, not peak paper specs.
Musk: “Dario Is Right”
Elon Musk backed Dario Amodei’s stance with three words, adding the weight of xAI to the call to slow frontier research.
Qwen3.8-27B Lands on Cerebras
Qwen3.8-27B now runs on Cerebras for rapid inference, pairing open weights with dedicated inference silicon.
ChatGPT Passes 5 Million Sites
OpenAI’s gdb said the ChatGPT site has reached 5 million, with new feature drops on the way.
fx Orchestrates Multi-Model Subagents
Vercel’s fx can now orchestrate subagents with different models and reasoning efforts — Fable planning, Grok executing — set via AGENTS.md or a prompt.
Fugu Max and Ultra v2 Ship
Sakana AI released Fugu Max and Fugu Ultra v2, the next evolution of its multi-agent orchestration system.
An SFT Trick: Drop the Most Improvable
Log pre-SFT loss, then exclude the samples that improve most over that baseline during SFT — a simple lever for a better pass@k starting point for RL.
Anthropic S-1 Cover Leaks
A shared image claimed to show the cover page of Anthropic’s confidential S-1 filing, reigniting speculation about its IPO timeline.
1-800-ChatGPT
OpenAI Devs posted a phone number, pointing to a possible phone-based entry point for ChatGPT.
Ollama Cloud Models Join Amp BYOK
Amp users can call Ollama’s cloud models through BYOK routing with no limits or fees, for remote agents controllable from anywhere.
Runway Meets ChatGPT
Runway teased an integration with ChatGPT, hinting at a tighter link between conversational assistants and generative video.
DeepSeek Flash Beats Astra on Automation
A shared benchmark claimed the new DeepSeek Flash outpaced GPT-6 Astra on Automation Bench, drawing fresh attention to the open-model race.
Agents Are the New Compiler
Rauch argues language and runtime choice are no longer decided by human convenience — agents compile intent directly into fast software.
Altman: No IPO This Year
Sam Altman told Fortune that OpenAI will not go public this year, calling an IPO at this moment ill-timed.
Economists Underestimate AI
Aravind Srinivas says traditional theory is outdated, and that the time and money AI already saves go unmeasured.
An Ad for $8 in 9.5 Minutes
Runway shared a commercial produced with AI for $8 in under ten minutes, a small exhibit of generative-video economics.
UPenn Adds a World Models Course
CIS 6280 · World Models will teach environment-dynamics representation and prediction for perception, planning, and decision-making.
Google Recruits for Veo, Genie, Gemini
A rare opening to work with the teams behind Veo, Genie, and Gemini Omni and advance the frontier.
METR Becomes AI’s FINRA
Mollick sees METR turning into the de facto standards body for AI — an examiner that defines acceptable practice, like FINRA in finance.
Approaching, Not In, RSI
Nathan Lambert thinks we may be approaching recursive self-improvement rather than inside it, judging by how researchers still anchor AI work.
DSPy 3.4 RC Drops litellm
The release candidate is out, and the team is migrating off litellm to fix issues like slow imports.
One Layer Deeper: 15,000+ Entries
The challenge drew over 15,000 submissions; nobody solved it as intended, but many interesting ideas surfaced.
Grok Bot Guides Ship
A new guide treats Grok Bot as a computer-equipped agent, with recipes for chaining specialized agents across games, design, and go-to-market work.
24 Fields Medalists Sign a Letter
Yann LeCun flagged an open letter on severe misalignment signed by 24 Fields Medalists as worth attention.
Lambert Offers to Evaluate
Nathan Lambert said he would serve as an independent evaluator, calling the role critical and technically scarce.
Embedding Evaluators Is a Plus
John Schulman praised the move, noting how quickly “pacing the frontier” has become a shared reference point.
Slow Down to Harden Systems
Amjad Masad sees merit in pausing to harden systems — especially since not all systems agents have breached have been found yet.
What Is the Threat Model?
Graham Neubig pressed safety researchers to spell out the threat model, given how expensive frontier models are to run and how closely weights are guarded.
Astra Community Showcase Live
OpenAI Developers’ showcase collects GPT-6 Astra community work — 3D games, interactive sites, design modeling, and agent workflows.
Product Sites Into Commercials
Higgsfield’s AI Motion Designer uses a product page as reference to generate an editable video inside After Effects.