Solved it. Tracked down the inference + evaluation bottlenecks and ended up cutting runtime by 3x. GPU throttling was absurdly high. Codex genuinely helped a lot and its capabilities are impressive, but after spending hours pushing both Codex and Claude at their limits, I'm even more convinced they're still far from replacing Honestly the satisfaction I got is just another kind of high **Also got crazy number of impressions, it's always the most low effort tweets

New research out — A developmental model with morphogenetic scaffoldings to guide self-organisation: a single model, many grown patterns. The memory-compute trade-off but in self-organizing systems. A collab featuring @EIiasNajarro @JakobSchauser and @risi1979 and yours truly https://t.co/AFGzwcMqoz

AI hype can be pretty easy to find. But clear explanations are definitely harder. Deep Learning with R, Third Edition by @fchollet and Tomasz Kalinowski helps cut through that noise with practical examples that build your understanding step by step, so you learn why deep learning models work, not just how to run them. Now in print and 50% off through May 25th: https://t.co/iA3DwJjh2H

Announcing the #ICML2026 invited speakers! - Pascale Fung (@pascalefung) - Susan Athey (@Susan_Athey) - Sham Kakade (@ShamKakade6) - Aviv Regev - Verena Rieser (@verena_rieser) - Arvind Narayanan (@random_walker) Check out the blog post for more info! https://t.co/TI11g45FTe

You don't have to level up to contribute to open source. You level up by contributing to open source. Not sure how to get started? Check out our latest GitHub for Beginners episode. https://t.co/Jyze45KoHo https://t.co/DCqAFACo35

Releasing my first kernel on @huggingface: MaxSim Late-interaction retrieval (ColBERT / PyLate) bottlenecks on materializing the full similarity matrix. This kernel avoids it by using tiled scoring with simdgroup_matrix (Metal) and WMMA. Result is 3–5× speedup compared to naive PyTorch. Try it out 👇

Reachy Mini is Assembled and Ready Shout out to @huggingface and @pollenrobotics #reachymini #Ai https://t.co/s012EE7Jul

Reachy Mini is Assembled and Ready Shout out to @huggingface and @pollenrobotics #reachymini #Ai https://t.co/s012EE7Jul

I am working on porting SAM models and harness into Apple silicon. Already seeing 1.25x inference speed increase on mlx with the sam2.1-small model. Quantized versions soon. Repo: https://t.co/y4RnHhL5C1 Model: https://t.co/mxKoAhlIRm https://t.co/cKO5OGr03U

+1 more

Qwen3.6-27B running 100% on WebGPU. Not the best speed but still 😁 https://t.co/Z1dpMkzykr

Qwen3.6-27B running 100% on WebGPU. Not the best speed but still 😁 https://t.co/Z1dpMkzykr

https://t.co/iZs5IJmrl5

Grok Summary of Elon Musk’s interview at the Samson International Smart Mobility Summit today. 1. Strong praise for Israel’s innovation edge
Elon called out Israel’s outsized impact, saying it “punches way above its weight, probably #1 in the world” in innovation per capita. He expressed clear admiration for the country’s tech ecosystem, especially in AI and mobility. 2. Tesla FSD & unsupervised robotaxis — near-term reality •Unsupervised robotaxis are coming soon. •Tesla is already running driverless tests in select Texas areas. •FSD availability: Elon stated it will roll out in both the US and Israel by the end of 2026. •As the AI improves, cars will feel increasingly “alive.” •Prediction: In about 10 years, most driving will be AI-handled. 3. Humanoid robots → abundance economy
Elon painted a future where humanoid robots (Optimus and others) vastly outnumber humans. This shift, he said, will create massive abundance and could enable ideas like “universal high income.” He framed smart mobility as part of this broader robotics/AI revolution. 4. Starship, multi-planetary life & Neuralink •Starship’s rapid reusability is the key unlock for becoming multi-planetary and building cities on Mars. •Neuralink was briefly highlighted for restoring function to people with paralysis or vision loss — described in one summary as potentially “Jesus-level” impact. •All of these threads (autonomy, robots, space, brain interfaces) connect to one overarching goal: “maximize the probability that civilization has a great future.” 5. Balanced optimism with risk awareness
Classic Elon: big-picture excitement tempered by realism. He noted risks (including rogue robots/AI) but emphasized proactive development and deployment as the path forward.

One Pass Is Not Enough: Recursive Latent Refinement for Generative Models "We introduce RTM, which replaces the single-pass latent mapping in style-based generators with an iterative refinement process, and show that this consistently improves both quality and diversity" "RTM achieves the highest precision and recall among current state-of-the-art approaches while maintaining competitive FID" "Unlike flow-matching baselines that achieve competitive FID at the expense of coverage, recursive refinement improves both quality and diversity simultaneously."

This is exactly how you deal with the “lack of diversity” claim. You need to mock the absurdity of it. Push back. Do not give the woke an inch. This should have been how Christopher Nolan dealt with it. Instead he followed the woke guidelines and desecrated a masterpiece. https://t.co/lcCEuXTi7G

I am honored to have @apples_jimmy as one of @SophontAI's earliest supporters https://t.co/oiBswS9pmc

DGX Spark just benched 200+ tok/s for Qwen3.6-35B with @AtlasInference on @spark_arena 🔥 How's that possible? Providers like Codex and Claude get ~60. Other major engines don't come close 🦥 We haven't seen speeds like this on GB10. NO ONE HAS. Atlas is shattering records 🚀 https://t.co/uR8Hzvn9xe

@SawyerMerritt Sentience. Can melted sand be sentient? My friend and I had an argument about that tonight. He convinced me it can be. But it sure goes around fallen tree branches well. Old robot too with 161,000 miles. https://t.co/LZO7071Drt

@__tinygrad__ Can your six year old read 40,000 posts every day here on X and build this: https://t.co/kiuZ7QXLzb Thought not.

Just won my first @ycombinator hackathon. Here is what my teammate @achidambs and I shipped: (Also won a Sony WH-1000X5 but not sure if I’ll use it so if you want it lmk) https://t.co/I3tVHSI0Cv

Grok Imagine ✨ https://t.co/tuVcr3tWjj

@AlephOneAI https://t.co/8h2GsroSi8

At Microsoft when anyone met with Bill Gates we prepared for weeks. Why? The time value of his time was more than most of us made in a year. I didn’t realize an hour could have so much impact. Elon explains it here. Tonight I took my friend to Big Sur for sunset. It only lasts a few minutes but is similarly impactful. You have to plan for that too. Grok will tell you when sunset is in any location. We considered Lake Tahoe but decided against it. I tell you all this because the best moments are planned. Because they are too valuable. Separately I asked Bill for a billion dollars. Hey you gotta take the shot. You don’t score goals by not trying. If I walked into a meeting with @elonmusk I would start with “the next hour will be worth way more than $100 million to you.” That isn’t true. I waste his time now by asking for a @neuralink. Or am I? I would use it to make him a lot more than that. One human could run 1,000 robots another robot entrepreneur told me with one on that is coming in next decade. @paulg calls them “highly leverageable moments.” Hope you experience a few in your lives. They are what you will tell the grandkids about. Oh and the second part is trying to figure out the universe. When you start really studying the sun it is crazy. A constant radiation explosion. So many are dreaming about harvesting its energy in space. So maybe a sunset viewing is entrepreneurial behavior? And highly leverage in other ways. Talked to cofounder of Robinhood about that. I asked him for keys to his car. Why not? :-) https://t.co/0JgMFOE5NH

Not only have conservatives become vanishingly rare in academia, so have centrists. That’s how complete the left’s dominance is: Even moderates are now a fringe group in academia. https://t.co/3TCZW5YzIF https://t.co/FRXNwQPnqc

James Madison described the powers of the federal government as “few and defined” and those reserved to the states as “numerous and indefinite.” We’ve been dangerously drifting from that understanding since the 1930s. The drift has been most evident in areas now most fraught with waste, fraud, and abuse. If we honored the Constitution’s limits on federal power, there’d be very little waste, fraud, and abuse in our national government. Share if you’d like to see a “constitutional reset,” in which any government function that’s not obviously and necessarily federal under the Constitution would be returned “to the states respectively, or to the people,” as the Tenth Amendment specifies.

Grok Build from xAI comes with /imagine and /imagine-video commands out of the box to generate images and videos directly from CLI. First CLI tool that comes with native video and image generation. Result 👇 https://t.co/hkHDf2HLV6

Elon Musk in new interview: "10 years from now, probably 90% of all distance driven will be driven by the AI in a self-driving car. It will be quite a niche thing in 10 years to actually be driving your own car." AI already drives for 90%+ of the miles I travel. https://t.co/PNktWdbPPO

@azerX Join the discord, we have a thread going for all the improvements the community wants to see that Ive been feeding into issues and PRs https://t.co/5EoJ4EBecb

Moonlit stroll by Grok Imagine https://t.co/oEAwbHfhuj

The Hermes Agent Kanban just got a big automation upgrade. Drop one prompt into the triage, and the orchestrator agent can take it from there - decomposing it into all the subtasks necessary and automatically assigning agent profiles that fit the specialization needed. You can also now add descriptions for each agent profile, to better assist the orchestrator in determining what tasks should go to which profile! Docs: https://t.co/Sx29FJDAxT PR: https://t.co/iCSJtYVkVC