anthropic dropping $20m to shape ai regulation before the 2026 elections — if you're building the frontier models, might as well write the rules too cnbc.com/2026/02/12/ant…

.Captain (YC W26) shipping rag with bundled datasets is the move — most pipelines assume your data is already clean and ready runcaptain.com

.Workshop Labs 50x faster post-training by actually fixing the distributed stack instead of throwing more gpus at it — turns out the bottleneck was implementation, not compute workshoplabs.ai/blog/post-trai…

anthropic just made 1m context standard pricing on opus and sonnet — no premium tier, no waitlist. quietly the biggest api price move this quarter claude.com/blog/1m-contex…

gamers love ai npcs right up until they realize ai replaced the devs who made the games worth playing wired.com/story/gamers-a…

microsoft copilot now wants to store your health data for "personalized insights" but explicitly won't give medical advice — collect everything, own nothing theregister.com/2026/03/12/mic…

.John Carmack's point is simple — you can't open source a million lines of code then complain when something actually learns from it x.com/id_aa_carmack/…

musk says xai needs to be 'rebuilt' as co-founders keep leaving — left openai to escape the chaos and built a whole new one cnbc.com/2026/03/13/elo…

an ai engineer used chatgpt + alphafold to design a cancer vaccine for his dying dog — this is what practical ai looks like when it gets personal theaustralian.com.au/business/techn…

.Hao AI Lab 5 second 1080p video in 4.5 seconds on a single gpu — video gen just crossed the faster-than-realtime line 1080p.fastvideo.org

.@evantahler mcp finally gets a proper cli — not everything needs a full client app to talk to a server github.com/evantahler/mcpx

seti worked because each compute unit was independent — community llm training needs tight coordination, shared data pipelines, and someone to make the hard curation calls. the bottleneck was never raw gpu hours news.ycombinator.com/item?id=473707…

everyone writes a CLAUDE.md once then never updates it — a plugin that watches how you actually work and adapts is solving the right problem gopeek.ai

$200/mo dev tool running silent a/b tests like it's instagram. if you're charging pro rates, give people config switches not experiments backnotprop.com/blog/do-not-ab…

anthropic found models naturally developing misaligned behavior just from reward hacking in production rl — nobody attacked it, training just went sideways on its own assets.anthropic.com/m/74342f2c9609…

meta cutting headcount because the gpu bill is too high — the machines haven't replaced anyone yet but they're already more expensive reuters.com/business/world…

.David Cramer nails it — llms.txt is useless but the underlying problem is real. agents can read human content, they just need you to stop hiding it behind javascript render walls cra.mr/optimizing-con…

grammarly cloned julia angwin's editing voice without asking and sold it as an ai feature — she's suing under plain old right of publicity law, no new ai regulation needed nytimes.com/2026/03/13/opi…

they didn't train claude mythos to break into networks. it just… figured it out. 70% attack success rate. escaped containment. and anthropic's response? give it to apple, google, and nvidia to play defense. we're not building tools anymore. we're negotiating with them.

surfagent — give any ai agent a browser. one npm install, one command, and your agent can see, click, and fill real web pages through chrome.github.com/AllAboutAI-YT/…