Jonathan Toomim

I don't like Twitter. 140 characters isn't enough. Please contact me elsewhere.

@MarkLundeberg @i2cjak Looks something like this:

@MarkLundeberg @i2cjak The pair of axons with the axon-length–delays that match the phase delay rise above the action potential threshold. The others don't. That's how you determine phase delay between ears and localize sounds.

that’s not fucking Fourier analysis that’s a low pass filter a fucking spring is a low pass filter bro 😭

@markuse_fcb @BerkeleyLabPSA @lbnlengineering @BerkeleyLab It's a superconducting coil. Read the article.

@BerkeleyLabPSA @lbnlengineering @BerkeleyLab What current density and duty cycle is this designed for? I would expect coil end thermal management to be the binding constraint here.

Winding of the MARS-D magnet's closed-loop coil was completed by @lbnlengineering @BerkeleyLab after nearly two years of intense technical work. Made in Berkeley Lab: Winding a Novel Closed-Loop Coil Magnet Structure engineering.lbl.gov/2026/05/14/mad…

@Realg813 @ClaudeDevs What happens to your data when OpenAI realizes all of its users are cheapskates who won't ever want to pay? Who does that data get sold to? I'd rather be the customer, not the product.

@Realg813 @ClaudeDevs OpenAI is not profitable. They're giving away stuff for free with the intention of getting you locked in and then fleecing you later once you're stuck with them. If that sounds great to you, then enjoy. Anthropic uses sustainable pricing and business models.

Starting June 15, paid Claude plans can claim a dedicated monthly credit for programmatic usage. The credit covers usage of: - Claude Agent SDK - claude -p - Claude Code GitHub Actions - Third-party apps built on the Agent SDK

@Realg813 @ClaudeDevs 2.5x, not 42x. Free users should not be counted. techcrunch.com/2026/02/27/cha…

19M vs 800M. And instead of figuring out how to close that gap, they're squeezing the 19M they have. The developers and power users who chose Claude over ChatGPT were the most loyal, vocal evangelists Anthropic had. The exact people who bring other users to the platform. And they're systematically driving every single one of them out — throttling, banning third-party tools, capping programmatic use, going silent when called out. OpenAI is giving away Codex, courting developers, hiring the OpenClaw founder. Anthropic is putting meters on the tools they promised were "yours to build with." That 42x gap is about to get a lot wider.

@DanielleFong Claude can also write webgpu shaders. Is this a parallelizeable task? Seems like it.

ok so i found out that the claude.ai harness can write wasm to an interactive widget, via zig. wasm is just super performant, and it's a better interface in a lot of ways than the remote control. the diagrams and interactive widgets are special magic. it uses zig (rust blocked for some reason for me) compiles to wasm, and then render an html, realtime, 3d, 175 hz, explorable explanation. it's really cool because you can make fluid interfaces and you can see it, and the ai can use it too

@ivanfioravanti @antirez The ANE and the GPU (and NA) can run in parallel, so it could be a path to even higher throughput. But I'm having to requant the MoE layers because Q2_K etc just aren't possible on ANE. So we'd need to adjust the Metal code too to be compatible. (Not hard.)

@ivanfioravanti @antirez I've been experimenting with something similar using the Apple Neural Engine (not the Neural Accelerators), but with the main goal of reducing power consumption rather than increasing throughput. If I finish, it will be interesting to compare results.

I don't want to be too excited for this because it's still experimental, but I have been able to find a way to make ds4 by @antirez faster in low power mode on M5! 🧐 Yes there is a very small logprob drift (min top-20 overlap vs baseline 16/20), but look at results here:

@ivanfioravanti If you really REALLY want to push: manually turn the fans on full blast. The fans-reacting-to-temperature algorithm reacts too slowly, and some thermal throttling results.

M5 Max in Automatic instead of High Power is much slower, I see at least a 15% difference. So when you really want to push... High Power mode on!

@antirez Correction: by "run a prompt" I meant run a *Claude Code* prompt to edit the source and do some directed benchmarking and profiling, or send Claude down an optimization path, not running a prompt in ds4.

@antirez I'm also thinking that maybe prefill can be done using ANE and GPU in parallel for higher total throughput. Which could be pretty cool. Decode will probably just be slower when on AC, but maybe faster on battery/low power mode.

Appreciate Ivan tweet. To put this into context, to build DS4 I used: my MacBook M3 Max (mine, 8k euros), 1 M3 Ultra with 512 GB (got access, 10k euros), one DGX Spark (got access, 4k euros?). Are we far from the times all you needed to do hacking was a computer? That's sad.

@McHeartattack3 @wilson_ricks And here's another paper in case you think that one was "biased" or whatever. nature.com/articles/s1227…

@McHeartattack3 @wilson_ricks In case you aren't aware: the mechanism for pollution causing heart attacks and strokes is this: Combustion products have particle sizes of 30-60 nm. Those particles cross from your lungs into your blood, where they nucleate into blood clots. link.springer.com/article/10.118…

Seems to be flying enormously under the radar that Trump has banned wind power in the United States. A full 10% of our electricity. Just imagine if another administration did this extrajudicially to any other industry.

@McHeartattack3 @wilson_ricks And if you want to offer scientific criticism, you have to do it based on the *methods*, not based on the results. You can't just look at the results and say "oh, I don't agree with those results, they must be biased." You have to explain how they were biased in their methods.

@McHeartattack3 @wilson_ricks But we both know that you're not going to cite any data, because you can't: the data just do not support your perspective. We both know you have done a bit of Googling already, and have found nothing. So that's why you are stuck with insulting and rejecting my data.

@McHeartattack3 @wilson_ricks Yes, I'm spending about 30 seconds per tweet to drag up links, so that you can read and learn something if you choose to. I'm not putting hours into each tweet. I'm still investing ∞ more effort in citing sources and data than you. STEP UP YOUR GAME.

@jtoomim @wilson_ricks Literally all the data you have "cited" has been extremely bias and incorrect. This is not pokemon dude. You cannot summon studies you have not read to do the argument for you.