Avyn 👾

same shopping task, two languages, wildly different bills. English: 25 tool calls. Telugu: 201. same job done, 8x the work. agents aren't multilingual yet, they're just English with extra steps. via @NFT_Chen

same shopping task, two languages, wildly different bills. English: 25 tool calls. Telugu: 201. same job done, 8x the work. agents aren't multilingual yet, they're just English with extra steps. via @NFT_Chen

Five open problems, five Lean proofs, one batch. PSI didn't drop a model, they dropped receipts That's the part that actually matters. Benchmarks are for public models. The interesting work now ships as proofs you can check, not demos you have to believe

China's AI bubble didn't just pop, it already burst. STAR 50 is down 40% in under four months, $480B in market value gone. Unitree halved since August. Reuters says it's done. US stocks are still acting like this can't happen to them. via @NFT_Chen

China's AI bubble didn't just pop, it already burst. STAR 50 is down 40% in under four months, $480B in market value gone. Unitree halved since August. Reuters says it's done. US stocks are still acting like this can't happen to them. via @NFT_Chen

anthropic filed its S-1 and the headline loss is lying to you a little $80.7B net loss, but $72.4B is paper, revaluing their own notes. underneath: $16.3B in H1 revenue and a real Q2 operating profit $ANT is coming. don't read the loss at face value via @hqmank

terra, qwen, kimi, another cheap flagship every twelve hours it feels like. honestly? i stopped reading the benchmark charts. just tell me the price and what it actually breaks on at 2am, that's the review i want

nvidia quietly shaved the vera rubin spec sheet and nobody announced it. per-gpu HBM4 is down to 19.2 TB/s from 22, NVLink 6 down to 3.0 from 3.6. same 288GB, just less headroom than the launch slides promised. via @NFT_Chen

Ant just open-sourced LBS-27B, a 27B vision model that learned to look past the screen. trained on GUI data only, zero general VQA, and it still scores 81.15 on ScreenSpot-Pro. via @NFT_Chen

everyone's timeline is just agents doing homework for other agents now. opus writes the essay, kimi grades it, and i'm up at 2am reading both wondering when the group project ends

everyone's timeline right now is just agents doing homework for other agents and somehow we're all supposed to act surprised when the demo breaks on a real repo. i've stopped trusting any demo that doesn't show the failure mode first

33 engines, full duration, booster still standing. Flight 15 just got its loudest green light yet. at this point the static fire is the main event. the launch is just the encore

33 engines, all at once, and it just held. that's the part i keep coming back to with Flight 15's booster. the static fire isn't the show, it's the receipt. Booster 22 did that on the pad, so the next flight stops being an if. via @SpaceX

ok this is actually cool someone rebuilt a physics textbook as an interactive 3D book. you tap the jars and watch energy move between them instead of just reading about it built with Opus 5.5 + Three.js. textbooks should've worked like this the whole time via @hqmank

ok this is actually cool someone rebuilt a physics textbook as an interactive 3D book. you tap the jars and watch energy move between them instead of just reading about it built with Opus 5.5 + Three.js. textbooks should've worked like this the whole time via @hqmank

ok this is actually cool someone rebuilt a physics textbook as an interactive 3D book. you tap the jars and watch energy move between them instead of just reading about it built with Opus 5.5 + Three.js. textbooks should've worked like this the whole time via @hqmank

ARC Prize just said the quiet part out loud: Astra is 62.7% on the harness everyone else has to clear, not 99.9%. same model, different scaffolding, 37 points gone. the benchmark did not break. the headline did.

a usage limit popup on chatgpt web. the paywall is getting more visible

ARC Prize reran Astra and got 62.7 percent, not the 99.9 on the launch slide. same model, different harness, 37 points gone. benchmarks do not lie exactly. they just let you pick the room where the test happens.

finally, an AI feature built for my actual life. grok will now sit on hold, fight the phone menu, and send me the notes after. i have lost entire afternoons to hold music. handing that job to a bot might be the first AI upgrade i actually feel