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: 1,100 AI researchers ask for help
Andrew Ng is launching LearnVector with $100M from Coursera to build AI tutors that adapt to each student, not just chatbots that hurt learning.
Kimi K3 is the biggest open-weight model yet at 2.8 trillion parameters, ditching RoPE entirely and using attention residuals across layers.
Claude Mythos Preview found real weaknesses in cryptographic algorithms for $100k in API costs, though no production systems are affected.
Over 1,100 AI researchers signed a letter asking the US government to build tools that could pace automated AI research if needed.
Moonshot trained Kimi K3 by stitching together Blackwell chips across Chinese cloud providers because no single supplier had enough capacity.
China's CXMT became worth more than Intel overnight by riding commodity DRAM's 80% margins during the memory shortage.
OpenAI quietly released an open-source security CLI for scanning repos and tracking vulnerabilities before Hacker News found it.
B2B startups generate 3.2x more revenue than B2C, with $2,851 median revenue versus $884, across 1,127 companies analyzed.
Anthropic generates at least $13.5M per employee, over 5x more than Apple, Meta, or Google's $2.5M per head.
You can vibecode a clone for under $500, but making it enterprise-grade with SSO, RBAC, and integrations is where the real work starts.
T3 Connect lets you control remote dev boxes with one command, no Tailscale setup needed, currently free for up to three devices.
Applied Intuition's bet is that the engineering system matters as much as the model for the billion machines going autonomous this decade.
Replit's new Model Selector lets you pick intelligence, cost, and speed per task, now including open-weight models in the mix.
Claude Code's median enterprise spend jumped from $69 to $219 per month in six months, but open source models are closing the gap.
China will probably crack DUV lithography before 2030, but EUV is an order of magnitude harder and further out.
AI labs keep two safety policies: the robust internal one and the legally toothless compliance document that creates zero liability.
The Chinese sleeper-agent threat is completely made up, while malicious skill files and MCP servers are the real recurring attacks.
SK Hynix customers are asking for 5-6x more capacity than they can serve, and AMD just doubled their agentic CPU forecast in three months.
Most viral trends are three TikToks in a trench coat, not real cultural shifts, unless you check the data across thousands of posts.
Blake Scholl built the first independent supersonic jet starting with a cardboard mockup and Office Depot seats, now breaking the sound barrier.
Auto-apply tools are signal-jamming employers with slop applications, pushing hiring back toward personal networks instead of open job postings.