·
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: Jensen's open-weight letter shakes things up
Anthropic updated its prompting guide and deleted 80% of Claude Code's system prompt for Opus 5 with zero loss on evals.
Cursor adds Claude Opus 5, matching Fable at half the price and compatible with Zero Data Retention.
OpenCode hit 13 million monthly users and $40M ARR after Anthropic's clampdown accidentally fueled its growth, says YC.
Alex Albert says Opus 5 feels smoother than Fable for many coding tasks despite all the benchmarks.
v0 now turns full Figma files into working apps by exploring pages and frames through an agent.
Lovable reports Opus 5 beats Opus 4.7 by 22% on demanding coding tasks with the highest finished app quality they've measured.
Jensen Huang published his first ever tweet, an open letter on open weights signed by 25 companies but not Anthropic or OpenAI.
Satya Nadella backed the same open weight message the same day, calling open models essential to a healthy AI ecosystem.
Nous Research shipped a credential firewall for Docker sandboxes where real keys never enter and tokens are useless if leaked.
Ethan Mollick jokingly prompted BenchBench and got an actual arXiv paper PDF that turned out kind of interesting.
Midjourney released V8.2 with more creative, bold, edgy aesthetics and better personalization as the new default model.
Runway launched Workflows in its Agent, letting you build node-based workflows through natural language for high quality outputs at scale.
Six House lawmakers introduced the FRONTIER Act requiring frontier model developers to undergo independent security audits and report safety incidents.
Beehiiv spends $25k monthly on AI tokens but still classifies AI's impact as marginal on shipping velocity to real users.
BlackRock's tech chief says the irony of $10T in CapEx is moving data centimeters and memory will matter as much as compute.
Midjourney acquired Co-Star and its founder now leads design, product, and frontend across all Midjourney efforts.
Lovable is running AI hacking agents in CTF against their prod infrastructure to test security at scale.
AMD shipped Helios and announced ROCm skills landing in Claude Code and Codex with Anthropic partnership the same day.
Replit dropped four updates including deployments 50% cheaper and up to 80% in some cases plus a new mobile experience.
Ryan Carson says running GLM 5.2 on your own infra costs at minimum half a million dollars so private defense is untenable for startups.
Aaron Levie argues the moneymaker in AI is inference, not the models, so open versus closed barely moves economics.
X removed 42,000 accounts automating replies with chatbots, saying AI engagement without humans runs counter to their mission.
Theo says Opus 5 feels like an in-between with Fable's taste and GPT's thoroughness, catching things Fable missed.
Lee Robinson reframes the debate as managed versus self-hosted infrastructure for AI, not just open versus closed models.
T3 Code merged 76 PRs since Monday including Auto mode, sidebar v2 beta, Opus 5 support, and thread snoozing.
Cloudflare Access lets agents both create public UIs and protect admin routes with Google login as part of IaC setup.
Greg Isenberg says cloud agents in VMs let you run 10x without collisions while local instances cap out fighting over worktrees.
DHH tested Omarchy on 8GB RAM and found it uses well under 2GB cold boot with 3GB free even with browser and apps open.