vkaku

Maybe people don't have a reason to write Zig code when rust or even C might do, and this could be that reason.

I've been very productive getting some Zig code to work with 0.17.0 and I really don't find a reason to contribute anything back given this policy.

The productive ones will just fork and accelerate faster. I understand the need for the mental muscle, this framing just sounds terrible at this point. Something something human scrutiny is not even an option the people want to consider.

They've got no hardware moat and they somehow think people will pay more money for slow software?! Hmmm ....

Everyone is making better software than NVIDIA today with AI, to the extent they are un-nerfing hardware and making stuff run 2x faster. Why would anyone in their sane mind pay NVIDIA anything when they can vibe engineer a faster library?

With jailbreaks coming in everywhere, gamers have a better future with piracy than what any of these console companies are going to "give" them.

Me: Let's make LLMs use that programming language.

I need this for my M6 Mini.

Good enough small models are the way.

If you're doing some FPGA and related tooling and CI/CD work, let's chat!

Insert stick in bicycle meme.

It was amazing back in the day. Wonder why Google really shut it down. I do thank Orkut for Orkut. It's where I met my wife.

L take.

Honestly, the folks writing inference drivers for Llama.cpp / GGML would sort of benefit from better compiler work like this.

In general, Valve's work has been exceptional and supplementing AMD's own ROCM/OpenCL and Vulkan team, they've gotten a lot of defaults right and people should work together with Valve to improve support for their chips.

I really think the authors in FT are not considering:

- The innate value of the models come from not training chips nor hosting, but the actual quality of thinking traces, inference costs and usage. Anthropic shines here, but not for long.

- The variety of skills and evals on new tiny open models proves Claude is expensive and not extremely better than these models today, and math benchmarks alone do not cut it.

- Enterprises are realizing that they can just buy some regular Macs and do their own inference far cheaper than invest in Claude Subscriptions, and Claude given tools are losing their edge to harnesses such as Pi and Opencode.

- Nobody needs Hacking Tools unless they are red-teaming their own code. The niche of people wanting this could write their own tools and workflows to test the effectiveness of many LLMs on these workflows and see how far they can push it.

- The amount of collusions, hypes and fear mongering do not sit well with many people (including myself) - stick to the science and make usage safe, like everything else we do with software in production.

- The IPOs often get priced down with market discovery and anyone who's buying it ONLY is interested in the long run, not inflated valuations not hype not any of these lies used to sell this stock. I am bearish about the IPO at this rate.

- For me, reasons for SpaceX money investments are NOT due to AI *That part is a true waste of money* but good enough due to manufacturing, supply chains and actually building useful Satellite infrastructure; I still am bearish about them launching NVL72 nodes in space, not because of the space part - I just don't trust those chips enough for it to sustain that long.

Of course it's a sponsored post.

LLMs were a bad idea. MCPs were the band-aid to get them closer to the real world and do something for recursive self improvement. Then LLMs became better suited for everyday use.

I'm tired of these kind of posts. This author is on my black pill list.

It's great they do C23. I'd love to get three-four projects worth of things merged together for a good reason:

- TinyCC, SDCC, Musl and Cosmopolitan Libc

Reason is that I require a quick fast portable compiler with a decent runtime that can also be embedded into a larger runtime, like the way Bun does - while keeping runtime small and portable.

SDCC does not need the same amount of bloat - having decent small C runtime is a job in itself. Fixing it once could allow a better compiler that isn't quite as bloated as LLVM nor as small like TinyCC.

That gap isn't solved.

Those Razr phone updates need to start landing sooner ....

Many developers will likely be developing at Android 16 APIs only. Will rely on Aptoide+AOSP exclusively. This is likely the timeline we'll see a decline in Play Store and Pixel usage.

We all know these aren't purely technical reasons, it's to break the monopoly of one Internet Company's decoder format portfolio. You can't let one Fruit Company's format be the thing either, so it's basically a L for engineering but a W for pleasing everyone TBF.

Market correction will happen. Banks go down and close during recessions. Hope these people are wise enough to see through these effects.

Don't buy this argument. It's like saying, we lords who hold this capital will build all this economically destructive stuff anyway, and wait for the world to not react to our stupid ways of enriching yourselves.

People are not going to slow down because this was brought to them on less than endearing terms. They don't see any of these stated noble intentions.

Claude was already used to cause economic, political and social destruction. As Anthropic is seen doing it, others aren't going to just sit down, read the blog and say, oh, I'll stop developing models because Dario, you touched my heart with your true words.

Here. You are not alone. Just run it and host models yourself.

https://github.com/guilt/xet-server

Here: https://github.com/guilt/xet-server

This is great. All algorithmic efficiencies are amazing!

One thing I'd remind all scientists and the wonderful people here is this wonderful meme/line from Jurassic Park: "Your scientists were so preoccupied with whether they could they didn't stop to think if they should."

What is the actual amount of data that needs to be pre-trained and what is not? Nobody has come up with great answers to this question, and I'm already seeing amazing 0.5b-2b parameter models working very well with n-Gram corpuses of data. So, how many parameters do you really need for a given workload?

Next Quarter? :) That's too long

LOL ... Accurate ;)

A lot of time accessibility is about that extra effort one can't get with tight development budgets. I apply generous amounts of butter while I keep the thing dry and let it be otherwise accessible at the same time.

A lot of the AI progress is stalled because these chips do not handle thermal contraction/expansion well at all, and end up permanently destroying these chips. HBM has been a failure at this scale with chips getting destroyed every 4 months or so during operations.

I'd like to see some actual science saying, here was the problem, here's how we solved it, here's the AFR data, here's this running after X cycles etc. Nobody has done this reliably yet. That entire industry is hiding the bodies.

I'd like to next the article which says almost no need for training models. That's really what I'm interested in at this point