Philpax

Nowhere. It's the same argument as upthread. A language that forces downstream consumers of an API to correctly deal with all possible outcomes would have prevented this.

Running inference on LLMs that can fit in 512GB. For example, you would be able to run either GLM-5.3-Flash or DeepSeek V4.1 Flash 1017 on this, both of which are very respectable models. It's up to you, of course, if ultimate privacy is worth $25k.

It was definitely AI-written. Has all of the usual cadence, "impact sentences", and pacing quirks. Pangram has it at 96% AI.

updated, ty

Post not up yet, but the models are: https://huggingface.co/collections/LiquidAI/lfm3

That's not necessarily incompatible. The raw output could come from RL rollouts where the model learnt from the process of solving these problems.

This work is/was a community endeavour, and if you're unpleasant, people will route around you. My sympathy ends because this isn't just an AI problem, it's an interpersonal problem.

The parent poster was talking about the amount of data transfer you could do on dialup over a month in 1996.

I wanted to be sympathetic to someone watching their hobby and source of income disappear, but I'm finding it quite difficult to do so when they're this smarmy and unpleasant. I would not want to involve this person in any of my endeavours, either.

One step closer to bringing the Rust experience to Python. Good.

Looks like this post's using "3T-class", which I think I've seen them use before.

Are we sure this is a new pretrain and not them rounding up 2.8T?

Eh, not necessarily, many of the RL vendors sell to Chinese labs, too: https://www.interconnects.ai/p/glm-53-how-chinese-labs-keep-...

> Bluesky is founded by the same founders of Twitter and is not open and has ridiculous policies

Huh? It is remarkably open and the policies are not really an issue.

I suspect that the additional friction in doing these things in Rust is still largely outweighed by everything else. I won't deny the friction - it's definitely there - but it has never been large enough for me to consider using anything else (especially reverting to C++!)

With that being said, I did write a compiler to specifically help me match C++ memory layouts in Rust [0], so I might be an outlier case :-)

[0]: https://github.com/ferrobrew/pyxis

For more detail: https://doc.rust-lang.org/reference/type-layout.html#r-layou...

I can confirm that this has been sufficient for all of my memory layout needs, including interfacing with existing C/C++ code in memory.

Is it? How so?

This is of interest to me, but I have no idea why you'd set `height: 720px`. I had to turn that off to make it readable.

And may we all be blessed with the compassion required to not say something like this about someone's potential lifework.

???

See also https://huggingface.co/nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A..., which fits in a 24GB card through a special 2-bit quantisation.

They weren't paying frontier rates. They had frontier-level staff to train their own bespoke, hyper-optimised[0] LLMs until they didn't.

[0]: https://blog.character.ai/optimizing-ai-inference-at-charact...

It's a bit silly to write an article like this without mentioning the company was hollowed out by Google in 2024.

It is difficult to make a company in a fast-growing sector work when all of your technical staff and leadership are working for Google, and your remaining staff can see the writing on the wall and are checking out.

Lack of focus can kill you, too. OpenAI learnt this the hard way: https://www.wired.com/story/openai-shuts-down-sora-ipo-ai-su...

the quote does not seem to match your statement here

People in sugar glass houses shouldn't throw stones :P

Given DHH's political opinions, this is an interesting comparison to make, even with

> I know how that era ended, and no - that is not the part of the analogy I am after.

to the extent where I'd almost say it's curious to draw upon this and not acknowledge the elephant in the room.

Agree with most of this, but:

> In mid-2026, China released Kimi K3, which was equivalent to the best released U.S. model at the time, Claude Fable.

I don't think this was, or is, true: https://artificialanalysis.ai/evaluations/artificial-analysi... - even from a vibe check alone, it was pretty clear that K3 was missing Fable's special sauce, especially for more hands-on problems.

The Chinese model lag depends on who you ask, but the current estimate seems to be about four months: https://stateofopensource.ai/

> In late 2026 OpenAI/Anthropic will release a tiny model that outperforms their previously-best 2026 model (Astra/Fable), that is priced 10x cheaper

We don't know the sizes of any of these, but GPT Luna and Claude Haiku 5.5 are tremendously competitive for their cost, and in some ways beat their larger-sized predecessors. I'd say there's enough doubt that I'd be willing to provide partial points to this claim. (No finetuning, yes.)

The point is not to have byte tokens: it's to have only byte tokens, so that the usual failures of tokenisation (e.g. the number of Rs in strawberry) can be avoided.