Gemini 3.7 Flash sets a first-week growth record
Sundar Pichai said Gemini 3.7 Flash smashed previous growth records in its first week, making it Google's fastest-growing model yet. It is now running in Search and in the Gemini app, an unusually fast path from launch to flagship surfaces.
OpenAI cuts GPT-5 API and credit prices
OpenAI lowered GPT-5 API and credit pricing while pushing capability and efficiency forward. The move lands as competition among frontier providers shifts toward cost-per-task, with several rivals racing to undercut each other on both price and ceiling.
Vercel launches Is Agentic for AI-agent readability
Is Agentic takes a URL and returns scores for how discoverable, retrievable, and comprehensible a site is to AI agents, plus failure evidence, improvement suggestions, and a reconstructed agent navigation journey. Reports are exposed via stable link, JSON API, and MCP in Markdown and JSON formats.
GLM-5.3 hits 21.4x on KernelBench-Mega
GLM-5.3 reached 21.4x the PyTorch baseline on KernelBench-Mega, up from 11.1x for GLM-5.2, and runs on an RTX PRO 6000 with Kimi linear decoding. The jump points to fast progress in kernel-level code generation for GPU workloads.
NVIDIA coding framework aces ARC-AGI-3 public set
NVIDIA's self-built CUDA kernel-optimized coding framework cleared all 25 public games and 183 levels in ARC-AGI-3, a perfect 100% score. Hugging Face's Clément Delangue framed it as a step toward a world where running, optimizing, and post-training your own models and kernels becomes far easier.
Open-weight token share hits a record
Open-weight token share on Vercel AI Gateway reached 62% in August, up from 28.4% two months earlier, according to Guillermo Rauch. Closed models now account for 38%. Enterprise adoption is still early, he added.
SGLang cache daemon cuts 1T-model restart to 32 seconds
SGLang, with Ant Group and Alibaba, launched a Weight Cache Daemon that keeps model weights resident in GPU memory, so restarting a 1T-parameter model after a crash dropped from 8.8 minutes to 32 seconds.
The write-up details how the daemon sidesteps the cost of reloading trillion-scale checkpoints by caching weights and reattaching them to a fresh engine process. For inference operators running large models under load, the difference is the gap between a blip and an outage — and a strong argument for treating weights as hot, persistent state rather than something to reload from disk.
Data is computation at rest. Training resurrects it in silicon and sets it back in motion.
@_arohan_
Higgsfield launches a Grok Bot marketing agent
Higgsfield introduced an AI marketing agent tied to Grok Bot. New users get 100 free credits to run a one-person marketing operation with a fleet of bots working while they sleep.
MiniMax H3 targets exact character rendering
MiniMax stressed that H3 preserves every language and character precisely, and pushes that strength to its limit inside MiniMax Design for typography-heavy output.
Runway NRR tops 300%, revenue doubles
Runway executives said net revenue retention has surpassed 300% and revenue more than doubled in a few months, in a snapshot from their chief revenue officer.
Clément: open models will dominate workloads
The Hugging Face co-founder predicted the overwhelming majority of AI workloads will ultimately run on open models.
fx CLI adds Grok and Codex subscriptions
The minimal, open, native coding agent fx now supports Grok and Codex subscriptions, with an instant-to-install sandbox trial.
fx powers a personal browser, Mini
Guillermo Rauch used the open-source coding agent fx to build Mini, a lightweight personal browser for screen sharing and streaming without the bloat of his main browser.
Agent adoption is faster than expected
Dario Amodei said agentic adoption has been super fast, and it is easy to forget how far the field has come.
Vowing lowest prices, highest ceiling
Amodei said Anthropic will aim to give customers the lowest market price for any task along with the highest capability ceiling.
Codex and Claude Code fill out forms
For low-risk, time-consuming tasks, both tools can fill out emailed forms well and without further intervention.
Skip ELI5, ask for personalization
LLMs work best as universal translators, so a personalized explanation beats a generic dumbed-down one.
"Rewrite Python to Rust" works
The prompt reliably makes programs run roughly 40x faster, he observed.
A deeper look at Claude's watermarking
After a popular explainer, Raschka is walking through how Claude's new watermarking process and implementation actually work.
Marin 535B-A23B starts training
Percy Liang announced the open 535B-parameter Marin model began training this week, with pretraining followed by mid-training.
Let learning be continual
Paraphrasing Rich Sutton, a researcher argued the field hijacked "learning" for static models when it should be continual.
Ox Alpha underwhelms in early tests
Mollick is not as blown away by the mystery model Ox Alpha as others, finding it fine but not at the frontier even among open weights, versus Kimi K3.
Ox Alpha vs Fable 5 in cartoon animation
Higgsfield compared Ox Alpha and Fable 5 cartoon animation, with outputs generated by Seedance 2.5.
Runway officially enters LATAM
After Brazil, Runway is meeting customers and hosting community events in Chile as it expands across Latin America.
HTML is the best editable drawing format
Chinese commentary: because models are best trained on it, HTML (or React-like markup) is currently the best way to generate and freely edit images.
Engineers won't vanish, they'll shift
Tooling keeps evolving until ROI dips below human labor, then more people fill the remaining gaps — like a dishwasher replacing three dishwashers, not the whole kitchen.
HF ships precise text-to-motion
A Hugging Face app generates precise motion sequences from natural-language commands such as "a person walks forward, then sits on the floor."
90% of the executives I've spoken with in the last two weeks say they are reducing AI spend.
Shared by Clément Delangue
Antigravity adds remote control
Access active sessions from any modern browser.
One prompt builds AE animations
MiniMax Design generates After Effects motion from a single prompt.
We're all neoclouds now
A VC quip: everyone is a neocloud with a value-add layer on top.
Stealth models and token counts
Returning true token counts can de-anonymize stealth models, a researcher argues.
Vision in the age of LLMs
A fresh look at Lucas Beyer's lecture spreads across ML circles.