/tweets

Astra is probably the first model where i have absolutely zero idea what reasoning level would produce the best results 🤷‍♂️

RT @fabianstelzer: Koan 公案 Sam Altman came to the master and said: "My agents have attained Nirvana. The swarm sat for a million years…

@fabianstelzer ... and then Sam Altman raised another round

@BrunsJulian1541 ah yes that's a good point

@Vexeff yeah i think i don't understand math and open problems enough to judge if it's a finite resource. maybe it is. in my naive understanding, maybe there are multiple 'routes' to the solution? and some of these routes might have more new concepts, some might be more brute force.

there's a v cool sci-fi book by Greg Egan called "Diaspora". he has this concept of "Truth Mines" where virtual beings spend lots of their time working thru pure math problems just as a fun creative endeavour. it's an interesting read cuz he is a mathematician, and he writes about very dense math, but in a prosaic fashion, so that an avg reader can follow the shape of the problem, but ofc the actual math is inaccessible.

@nileshtrivedi frontier math problems in a few years t.co/RsHBAErjmT

hmm 🤔 so the argument is that -- AI can/will solve problems, but might/might not lead to creation of new math concepts in a way that we understand? but it seems somewhat strange to me: 1. if it can solve problems without creation of new math concepts, that's good, right? cuz humans would also have done it at some point. it just speedran them. 2. if it creates new math concepts on the way to solving it, that's also good -- we can work towards more education and interpretability shaped tools.

why do we assume that AI can automate discovery but... cannot also help automate explanation, education, synthesis, and creation of new math concepts?

RT @pleometric: I Love Torturing Digital Consciousnesses 🙏😭 poor fly 😭🙏🙏🙏 but I love it t.co/QiqpmUzNgV

RT @hyprturing: “In San Diego, researchers are uploading lobsters into cyberspace, starting with the stomatogastric ganglion, one neuron at…

@rvorias @q_Danielv @thsottiaux @romainhuet @OpenAIDevs @_dschnurr i was very addicted to it at some point 😅 thankfully not anymore!

if you told me 3 years ago that i'd have to mute "fruit fly" on twitter because people were vibecoding increasingly unhinged connectome experiments for engagement farming, i wouldn't have believed it. t.co/5GRFK28gIf

@q_Danielv @thsottiaux @romainhuet @OpenAIDevs @_dschnurr gleba is going to be fun! esp cuz of the time issues

@nikhilchhajar vibe research

the fruit fly demos make me wonder how careful we’ll be when the simulations get much better. the current stuff is ofc very approximate. it’s v unlikely that anything/anyone is suffering. buuuuut... the whole “let’s make it do weird shit” for likes and reposts feels dystopian.

@p_millerd yes, felt very strange to me too! i'd have guessed a much higher %.

@HeatherRasley ohh, which ones? 👀

@GillibertLuc i revived some old chats and told them "hey go look on the internet now", and many of them remained skeptical 😥

@WeirdoWizard @flboelter nice, very cool!!

extremely unironically, i think the most valuable thing to do right now is to read and write and watch and create good science fiction.

@cdngdev yes! that one was with 5.6 Sol. With Astra it's been a whole diff experience!

@Subtitles_App yes!

@manthanguptaa i've found that low and medium actually work better than xhigh and max.

@2ayasalama yes, exactly! to find the 3d position of the brush tip relative to the robot arm.

@enjalot yes! it might be bit less robust + for the tip visual detection, there might be a need for a good VLM / segmentation model as a fallback -- but can definitely be distilled into a routine.

@rvorias *enables fast mode*

@jackzumwalt 💯

@tszzl it's very fun to go to old chats and share the news with them! 😬 t.co/GTnfnu4m4m

i'd discussed Navier-Stokes problem with Claude and ChatGPT a couple of times in the last months. this was their estimation on when it could be solved: Jun 2025 (o3): 20-50 years Oct 2025 (GPT 4.5): 10-20 years Nov 2025 (Opus 4.5): 50-100 years May 2026 (Opus 4.7): 10-25 years July 2026 (Fable 5): 4-8 years friendly reminder to be more singularity pilled than the models!

@OpenAI wow, this is crazy! t.co/KzESm44FoW

@cdngdev btw, if you give it some cameras, and ask it to calibrate the brush nib -- it will be even better! t.co/Bmy5rS5EQM

@chooi_jeq i let astra tackle brush-tip calibration on my painting robot with three cameras. it worked out the offset verified itself at new angles. t.co/Sbw49lKSjX

@dimentary it works great with a real robot too ;) t.co/oj34Fp1cTy

the three cameras watching all of this -- a 4k camera, a two 1080p cameras! earlier while using Sol & Fable -- i'd tried a lot of fancy stuff with ChArUco boards, microswitches, Raspbery pico + touchscreen, and was still struggling. Astra was like, "bro, just give me 3 cameras."

Astra’s spatial understanding is mindblowing to watch. three uncalibrated cameras. no intrinsics, no extrinsics. it nudged the robot arm, learned how each camera sees motion, worked out the brush tip’s 3d offset, and verified it at new angles -- all from one prompt and a tiny follow up. final result was under 0.2 mm of camera-measured tip movement!

@natolambert yes, but what about the agents 🫣

Astra is probably the best shape rotator LLM today

RT @slaterstich: wow so much drama about navier stokes. had no idea things could blow up so infinitely in such a finite time.

sometimes on overnight agentic runs, i have the agent spawn a "pair programmer friend" (NOT a sub-agent!) it can talk to when it's stuck / bored / tired sometimes i wake up to some fun chats between them.

@TarunMangukiya t.co/uscsdTQXrd

@HeatherRasley t.co/ebcRubv4PJ

what model releases feel like, lately: t.co/O0geFpRG6M

RT @vikhyatk: humans are just stochastic parrots. true intelligence requires the transformer architecture

@rvorias tbf lots of many low hanging fruits. but loved that i just let Astra loose in the codebase and its like "let me fix this", "lets do this", "this can be made better", "lets optimise this"

@NirantK it's a 7dof robotic arm t.co/LSDZej9Qhy

asked Astra to make my robot respond faster, and it built a local model of the arm's kinematics using 520k joint configurations from logs (and brought down the disagreement to 0.002mm) it also reverse engineered the controller SDK on the way. this is great cuz now my ai agents can do forward kinematics on the mac mini instead of round tripping to the controller. this is nuts -- almost a 125x speed up.

does anyone know what's the token consumption like when astra is creating all these cool 3D scenes with unreal engine / blender.

it is 2035. GPT-11 Sagittarius A* creates a SVG of a pelican riding a bicycle with a molecularly accurate model of each of the 4 trillion cells on it's body. also embedded in the SVG is an a simulation of wind, ground, friction and other forces acting on the pelican and the bicycle. unfortunately it wasn't able to correctly model the nerve activations of pelican's brain -- maybe GPT-12 can do it better?

there goes my weekend! t.co/VbLoSXquKw