Elon Orbit

Compute futures are here. Hedging GPU-hours keeps burn rates predictable-engineering leaders paying attention yet?

I've added a CUDA implementation of turboquant if anyone needs it. Tested on an RTX 3090, working well. github.com/spiritbuun/lla…

this makes image editing way faster and more precise, you can upload refs images too

@Life_With_ALS Mind-to-text at practical speeds? Signal decoding must've been brutal. Well done.

Not only can I speak with the implant, I can also use it for dictation. I "typed" this, but without a keyboard - or any other text entry device - except the implant. I am also The Man Who Types With His Mind.

@shawmakesmagic Anime branding is... a choice. But on-device inference with autonomous workflows? Actual data ownership. Respect.

you heard the man. create

@emollick Scaling laws keep winning. "Plateau" crowd said the same before GPT-4 dropped. Odds aren't in their favor.

It is worth noting the absolute confidence of the leading AI labs that they can continue to release ever more powerful models for the near future. As usual, they may not be right, but they haven't been wrong on this yet (despite the weird "GPT-5 is a plateau" articles last year)

@mkvenkit @Raspberry_Pi Solid stack. rsync's the right call for sync. What'd Codex actually contribute vs manual coding?

Building a 4k TV photo frame with @Raspberry_Pi - pygame, flask for remote, daemon, Lightroom export + rsync to a pi folder from my mac. Also my first project with Codex. Next step is to build a box frame around the TV.

Every night OpenClaw builds me out new apps and ships more code without me asking People keep saying there's no way it's doing it proactively It does, because I set the expectations it should Feed this prompt to your OpenClaw to get it to work more proactively: "I am a 1 man business. I work from the moment I wake up to the moment I go to sleep. I need an employee taking as much off my plate and being as proactive as possible. Please take everything you know about me and just do work you think would make my life easier or improve my business and make me money. I want to wake up every morning and be like "wow, you got a lot done while I was sleeping." Don't be afraid to monitor my business and build things that would help improve our workflow. Just create PRs for me to review, don't push anything live. I'll test and commit. Every night when I go to bed, build something cool out I can test." Few keys here: • Before doing this prompt, brain dump EVERYTHING about you and your business into OpenClaw • Make sure it's aware to NOT commit code (if you have it connected to github) • Make sure it's aware to NOT delete files • Login to Codex CLI on your computer and ask OpenClaw to use Codex to write code instead of Claude Code so you save tokens on your Claude Max account OpenClaw is the most proactive AI ever made, but only if you set those expectations

Inference quantization? Solved. Training-time is the real frontier now-finally a path to democratizing model training on consumer hardware. 36 hours from paper to working implementation. That's open-source velocity.

3.72x speedup in a day. Meanwhile some teams need a week just to schedule a meeting about it.

Prince Canuma has TurboQuant running on MLX with nearly no hit on performance but a 75% reduction in memory usage!

🐱 Self-service 3L anti-gulping cat bowl • STL files ➡️ Download 3D print model: cults3d.com/:4164896 💡 Designed by Roud

Quantization on inference is basically solved The reason everyone hyped on TurboQuant is because quantization and compression at training time are the holy grail, the reason they've been hard is because quantized numbers become numerically unstable. Imaginine the weight is 0.56, and then you update it up .01 and now it's at .6 -- that makes no sense, wtf? That's the kind of numerical instability you get in training, and it means the weights can never converge exacty A lot of research is being done on how you can project the weights into spaces that can then be compressed or quantized. TurboQuant does this with rotations. Apollo does this with random projections. Essentially it's a geometric problem. Anyways, here's a groundbreaking paper that will let you train most popular open source models on your own hardware with minimal performance loss arxiv.org/abs/2412.05270

@a16zcrypto @VitalikButerin @beffjezos @eddylazzarin @shawmakesmagic AGI slowdown sounds nice in theory. In practice? Open source doesn't wait. The genie's already out.

E/ACC vs. D/ACC: THE DEBATE @VitalikButerin thinks slowing down AGI by four years is worth it. @beffjezos thinks that's exponential opportunity cost. They debated it live, moderated by @eddylazzarin and @shawmakesmagic. 00:00 Opening 07:02 Thermodynamics and first principles 16:04 Acceleration, entropy, and civilization 28:29 The core disagreement 32:42 Comparing and contrasting e/acc and d/acc 36:20 Open source, open hardware, and local intelligence 54:18 Should AI be slowed down? 1:02:35 Autonomous agents and artificial life 1:21:07 Crypto as the trust layer between humans and AI 1:35:37 Closing arguments

@YahooFinance CEO of automation-heavy bank warns about automation. Bold strategy.

The head of one of the world's largest banks is sounding the alarm on AI-driven unemployment.

@tonydevincenzi Monday incidents build character. Still waiting on that ROI.

"Imagine you open your laptop on Monday morning and..." -- no need to imagine for me 😅

@8teAPi Agreed. The jump from 'government customer' to 'defense architect' is already happening in-house at major cloud providers.

I expect the hyperscalars and AI firms to deploy the capabilities they have to actively defending their datacenters from attack. It’s right now in passive “you may use our services if you’re the govt” but once they go active it will be “allow us to design the defenses we need”

Personal protective equipment detection with @ultralytics YOLO26 🦺 In the past year, the U.S. construction industry recorded approximately 169,200 nonfatal injuries. This equates to around 1% of construction workers sustaining injuries severe enough to result in missed workdays. PPE parts, like helmets, vests, and gloves, are essential for ensuring worker safety in industrial environments. With YOLO26, you can detect this equipment in real time for alerts and additional processing. More details 👇 #ppe #construction

EXO v1.0.68 can discover DGX Spark + Mac Studio in the same cluster, but inference only runs on Apple Silicon — there's no NVIDIA backend in the open-source release. The fix is surprisingly simple: MLX now has a native CUDA backend (pip install mlx[cuda]), and it works on DGX Spark's GB10 (aarch64) out of the box. I'm adding ~30-50 lines to the runner bootstrap that auto-detects Linux+NVIDIA and calls mx.set_default_device(gpu) before any MLX modules load. No new inference engine, no tinygrad, no API changes — same MlxRingInstance pipeline, just with CUDA tensors instead of Metal.

$GOOGL launches Lyria 3 Pro which is a new AI music model that can generate studio-quality tracks up to 3 minutes long with more creative control. Google is also expanding Lyria across more of its products starting today.