·
Finding signal on Twitter is more difficult than it used to be. We curate the best tweets on topics like AI, startups, and product development every weekday so you can focus on what matters.
PLUS: Gemini 4 Argon's expert ranking
Epoch AI estimates chips shipped through 2027 could run hundreds of millions of concurrent agents, matching 140–720 million full-time workers.
Arena's CEO says Gemini 4 Argon ranks #1 on text but drops to #7 with experts and #10 on agents.
Applying DeepSeek V4 Pro serving benchmarks to that chip supply yields about 1.9 billion concurrent agents, equal to 8 billion workers.
OpenAI used DevDay to unveil Dots, its fifth attempt at an AI agent product, alongside faster developer tools and an app marketplace.
GitHub's September ship log expands Project HydraFusion to more Copilot surfaces, a research preview that orchestrates multiple models rather than one.
xAI released an experimental TypeScript SDK that puts text, voice, image and video in one package with the latest Grok models.
Perplexity released seven open source projects, including a multimodal decision model scoring 85.7% average across 11 benchmarks.
Underdog, a private personal AI assistant from @0xSigil, launched with backing from a16z, Khosla Ventures, and Hummingbird.
Meta announced Muse Gadgets with open-source SDKs and a Muse Home Link device.
a16z introduced Discourse, not a group chat, not a broadcast social network, built off the Cosign graph.
llama.cpp can now run decision models locally through a new /v1/systemone endpoint, according to Georgi Gerganov.
Dan Shipper says Astra's habit of adding too many bells and whistles makes it less reliable than Fable for long delegated tasks.
Dan Shipper says Astra has good design taste but added labels and buttons in places that didn't make sense on a login page.
Theo's T3 Code PR adds Pi support, scheduled tasks, thread forking, and agents that start child agents on any provider.
SpaceXAI plans to bring roughly 420,000 Nvidia GPUs online in November to meet billions of dollars in new compute commitments.
Google launched a Project Suncatcher prototype satellite to test how TPUs handle spaceflight, with up to 8x more solar power in orbit.
McDonald's uses an AI pricing engine across 14,000 U.S. restaurants to estimate each area's willingness to pay, per Reuters.
Hugging Face found the same model scores 62% in one agent harness and 33% in another.
Google Research's Cogentic pairs provers with advisor and verifier agents and produced new results on five open problems, checked by domain experts.
Purdue's owners took out $10.7 billion in the decade after a 2007 guilty plea and $635 million fine.
Freckle's kids phone charges the same $25 a month whether a child uses it for five minutes or five hours.
Greg Isenberg lays out a 13-step playbook for vibe manufacturing, from AI-written CAD files to made-to-order sales.
Jack Friks left Claude Opus 5.5 alone for two hours on a vague redesign task and returned to 27 commits.
Levelsio says Fable 5.1 on high effort fixed a 4-byte struct mismatch that Opus 5.5 missed, getting Urban Terror running on the web.
Overheard in San Francisco: "Is Sholto like Anthropic's Roon?"