Ben Gras

High-level languages let you pack more meaning into each line than e.g. C, so you can express complex ideas quickly, at the cost of runtime performance. Sometimes a good trade-off. But if LLMs make low-level code cheap to write, the calculus might change.

@NicoInberg @erikmauritz Ik herinner me zeker WOL en Nina Brink's champagne dansje nog. Ik zit nu even de Wikipedia van het de WOL beursgang te lezen en vind - door een moderne bril misschien - eigenlijk de terechtheid waarmee NB zwart werd gemaakt achteraf misschien niet zo'n uitgemaakte zaak. Wat jij?

Herinnert u zich deze nog? Vandaag 26 jr geleden kwam World Online naar de beurs. @erikmauritz ziet analogie met OpenAI straks. Of toch niet? Lees hier meer: deaandeelhouder.nl/columns/wordt-…

Following up on yesterday's announcement by @amsatdl about the lack of signals received from Tianwen-1 since 2025-12-23, I have published a short post that contains the orbital state based on telemetry received up to 2025-12-22.

The ‘number station’ sending mystery messages to Iran ft.com/content/86c4a4… via @ft

tinygrad's built in 321 line LLM server is now fast. This is on one 9070XT in a tinybox red. Try it. It has 0 deps, and includes the tokenizer and web page in the file.

If you ever feel like you're late to the game, consider that in the 1890s many scientists thought physics as a field was completely solved (quote below is from Albert Michelson in 1894). On the front of intelligence science, it feels more like the 1870s. For the first time we have something that is starting to really work (however primitive it may be), which we can use as a springboard for the next few decades of discoveries.

I always lost performance when I tried to use silu/gelu activations in my RL value networks, and I finally understand why. If the pre-activation values are small, the smooth curve through zero is basically a linear activation, destroying the representation power of the network. You need a batch/layer/rms norm on the preactivations to put them in the range the smooth activations are designed for. Internal norms generally hurt performance on our RL tasks, but combining them with a smooth activation at least works basically as well as a raw relu (but slower). So, not actually a win, but the lightbulb of understanding was good!

@MuhammadAdilIn1 @PrismNdss26 @akulgoyal00 TL;DR -- @PrismNdss26 is basically a gift-wrapped starter kit for conducting effective systems research on real-world security operations, an area that frankly needs more attention and talent. Join us!

@NicoInberg Ik ben blij dat deze reactie het eerste is dat ik ervan zie, ipv de gebruikelijke pearl clutching. Fijn Femke!

Humor is tegenwoordig bijna strafbaar.

There are two categories of people: those who quickly figure out that chatbots give you the answer you expect when you ask questions in a biased way, and the ascended polymaths currently out-thinking every expert on Earth

@gynvael Wow. That looks incredible. Pledged 2 hardbacks, can't wait.

Just got this link on my discord - kickstarter.com/projects/bitma… - passing it along because this book looks fun!

@JoshDWalrath The plateauing at 90% is a bit telling - 10% of us are uggos. Not you though buddy.

I know what I look and sound like, so I'm that exception...

@JoshDWalrath Looks like just a superb wednesday.

An embarrassment of riches.

@NicoInberg Subliem Nico

je zou zeggen dat is de nr 1 maar het is de runner-up

I now have the perfect number of followers. No more followers allowed. Thanks!

Ben je een journalist of activist en vrees je, waar ter wereld ook, de veiligheidsdiensten: Lockdown Mode (‘isolatiemodus’ voor iPhone) is your friend 404media.co/fbi-couldnt-ge…

Marty at Id did ask for a video from me, but there was a miscommunication about when he needed it, so I didn’t get it done by the actual anniversary…

New SPT-3G constraints on CMB secondary anisotropies, including the thermal and kinematic SZ effect! "At ℓ=3000, we find thermal SZ power at 143 GHz of DtSZ=4.91±0.37μK2 and kinematic SZ power of DkSZ=1.75±0.86μK2" @NSF #southpole #Antarctica arxiv.org/abs/2601.20551

A conventional narrative you might come across is that AI is too far along for a new, research-focused startup to outcompete and outexecute the incumbents of AI. This is exactly the sentiment I listened to often when OpenAI started ("how could the few of you possibly compete with Google?") and 1) it was very wrong, and then 2) it was very wrong again with a whole another round of startups who are now challenging OpenAI in turn, and imo it still continues to be wrong today. Scaling and locally improving what works will continue to create incredible advances, but with so much progress unlocked so quickly, with so much dust thrown up in the air in the process, and with still a large gap between frontier LLMs and the example proof of the magic of a mind running on 20 watts, the probability of research breakthroughs that yield closer to 10X improvements (instead of 10%) imo still feels very high - plenty high to continue to bet on and look for. The tricky part ofc is creating the conditions where such breakthroughs may be discovered. I think such an environment comes together rarely, but @bfspector & @amspector100 are brilliant, with (rare) full-stack understanding of LLMs top (math/algorithms) to bottom (megakernels/related), they have a great eye for talent and I think will be able to build something very special. Congrats on the launch and I look forward to what you come up with!