armcat

Maybe - though they refer to hyperscalers explicitly in other parts of the report, and then there is this:

"Hyperscalers have strong earnings from their core businesses. Yet, free cash flows are rapidly declining as hyperscalers try to sustain large AI investment needs, leaving debt financing increasingly necessary"

So I'm reading the actual IMF reports itself [1] and I come across this on the first page:

"First, although core firms in the AI ecosystem have been profitable so far—easing pressure on AI-related equity valuations—weaker or less durable investment and profitability in AI could trigger a correction"

Is that correct - that the AI firms have been "profitable"?

[1] https://www.imf.org/-/media/files/publications/gfsr/2026/oct...

No, they still operate on tokens, it's just the difference in how text is sampled. It's done through iterative denoising and in parallel, across multiple token positions.

There has been extensive research into token-free LLMs, but for some reason we are still operating in a token domain, so there is something to that.

Byte Latent Transformers (BLT): https://arxiv.org/abs/2412.09871

Charformer: https://arxiv.org/abs/2106.12672

No I think it's more accurate to call it WSL Containers, or WSLc, but the release is indeed WSL 3.0.*.

I love the comments in the <head></head> section (translated from Chinese). Never seen anything like it. Very explicit documentation of decisions, clear historical context, and lot of warnings and "donts" that seem to be directed towards AI agents building this site.

Great job in benchmarking! I wasn't even aware of the WSLc (WSL3) release. For others, the release info is available here: https://github.com/microsoft/WSL/releases/tag/3.0.1

EDIT: stumbled on the previous HN post here: https://news.ycombinator.com/item?id=49970507

The real story here is this wonderful exposition in applying diffusion models to a time series data that is neither discrete nor continuous. It’s always fascinating to see diffusion models applied in different scenarios, same with diffusion language models.

I’ve been trying to rebuild and modernise an old abandoned ms dos game, simply by pointing Codex (6.1 Sol) at the game directory, and it works insanely well. It’s able to understand the data formats, unpack graphics and sound assets, and reconstruct game logic.

This really speaks to me, and Terence Tao literally wrote something very similar few days ago [1] where he advocates for a more holistic approach to intellectual fields, and the value of community building.

On a personal level I had so many to-dos, like learning Rust "properly", but I feel incredibly demotivated now, because I can just get Opus or Sol to do it. But maybe I just need to block out one day a week to be my "no AI day", and just do it.

[1] https://mathstodon.xyz/@tao/117395269325940185

It's weird because two days after Jev was released there were a dozen decision models, a week later there are several dozen, mostly open source, OpenAI's own Decisions API [1] beats it, and you can easily finetune your own [2]. But as others have pointed out, this doesn't matter.

EDIT: As I wrote this Microsoft just released their own Decision-1 model [3].

[1] https://developers.openai.com/api/docs/guides/decisions

[2] https://unsloth.ai/docs/basics/train-your-own-decision-model...

[3] https://commandline.microsoft.com/microsoft-decision-1-model...

Was covered today already on HN:

https://news.ycombinator.com/item?id=50004790

https://quesma.com/blog/invisible-cities-one-shot/

Paywalled :-(

Their official website: https://www.atomicmachines.com/

I checked the prices of two schticks of 16GB DDR5, it went from ~ 100 USD to ~ 600 USD in 12 months.

There are bits and pieces of info scattered everywhere but no coherent picture. We know from a16z report [1] that only 2% of US households pay for AI subscriptions, so most of the seat based pricing comes from business and enterprise agreements. We know that OpenAI ads business has $1 billion USD in annualised revenue run rate [2]. We know from OpenRouter data [3] that in 2025 70% of the API token spend was across all the proprietary models (Anthropic did very well in 2025, while OpenAI clawed back market share in 2026).

It all starts to look like a very low margins business, and reminds me very much of telecom industry.

[1] https://www.a16z.news/p/state-of-markets-ii

[2] https://www.reuters.com/business/media-telecom/openais-ad-bu...

[3] https://openrouter.ai/state-of-ai

Those are insane benchmarks at this size. Well done!

My favourite was Mulholland Drive DVD menu - it felt like you were watching LA in real time from the hills, while listening to eerie synth music. Pure art.

https://www.youtube.com/watch?v=zMztEUXXF00

This is probably one amazing positive these models have brought forward - the ability to break out of a single modality, and utilise others to help us learn and visualise. While the visuals here are amazing, my son for example prefers to learn by listening and talking, so we convert lot of his study materials into audio and real-time voice roleplay.

Also their GitHub repo: https://github.com/storytold/artcraft

It's still impressive that a hollywood actor/director can even mention a convolutional neural network, and he's right in how it works. His company InterPositive that makes AI tools for filmmakers did recently get acquired by Netflix [1], who in turn have stated the desire to create new series and films with the help of AI. So it's highly credible that he can whip up some CNN code in PyTorch.

[1] https://about.netflix.com/en/news/why-interpositive-is-joini...

I think this focus on a "holistic" approach applies to everything AI is touching now, not just math. On Twitter I see people one-shotting games, or reproducing games. If the goal is to just one-shot a game using AI, it's done. But if the goal is to produce immersive medium that people can truly enjoy, admire the story and the craftsmanship, and can find entire new ways of bringing a story to life, that's something else entirely.

Wonderful article, going from random walk to a Gaussian distribution. As an aside, I feel that normal distribution is way overused in practical contexts (apart from explaining Brownian motion, heights of people, and some other interesting observations). Heavy-tailed distributions that follow pareto or power-law, tend to be closer to actual reality, on many phenomena.

Interesting, thanks for the share!

I put Decisions API against classic statistics experiments: (1) loaded coin, and (2) marble selection from a jar, with replacement. I tested both predicate and choice questions. I ran thousand trials against each experiment, and I also did an experiment where I change the order of choices, to see if it matters. Summary:

* Using predicate questions gave nearly perfect/expected probability outcomes

* Asking it to choose an outcome behaved differently from drawing randomly - if the true probability of a red marble draw was 50%, using Decisions API produced 86%, i.e. it picked the right marble but gave a significantly more biased weight on its choice

* Changing the choice order changes the probabilities! Moving the red marble from first to last choice changed its probability estimate from 86% to 73%

Kind of like random projections. Pre ChatGPT there would every now and then come a paper that states something along the lines of: instantiate a large randomly populated matrix, multiply by the input, and win! It kinda makes sense because you "stretch out" the space and separation boundaries become easier, but I have never seen it fully utilised in production in any meaningful way. Anyone remember the DANs (Deep Averaging Networks)?

Regarding your world model statement. This is completely FALSE. Learning the visual statistics of a physical world is NOT the same thing as learning its causal dynamics. The difference is observational likelihood versus intervention-dependent dynamics. There have been great studies disproving video models as world models, like this ICML paper: https://proceedings.mlr.press/v267/kang25g.html. Unfortunately lot of people treat them as world models, mostly because of their ability to reproduce increasingly convincing physical behaviour without ever discovering the underlying physical laws. This is due to many things that I could write an essay about, but better conditioning, latent space represtnation, scaling etc, all make them look awesome.

I can still get absolutely insane results with MiniMax H3 - insane in the sense that it would not make sense at all and would make your head spin.

Sure and that's what I do, but a video can be seen as a causal generation on discreet sequence of images, each image conditioned on the one before it. It can also be seen as a series of image editing tasks. It would be cool to get this working in imagegen because of the amount of control you would get. Right now with video generation you can at best specify start and end frame and hope for the best.

Does anyone know if it can be used to generate accurate frame-by-frame sprite sequences? I found that no image model can do this well (with sufficient fidelity) - neither with one shot (full spritesheet), nor single frame conditioning. It would be great if an imagegen model could do this. What I do now (I use my own tool https://github.com/acatovic/ai-game-studio) is basically generate a reference image, then condition on that image to generate a very short video, then extract and prune frames. Then I get indie-level sprite fidelity about 90% of the time.

This takes me back, my first adventure in programming was as a 12 year old creating a Pong clone using Turbo Pascal on DOS. It came in handy later as a 20 year old when I was doing freelance as a Delphi dev. Amazing to see Pascal on HN.

I got into computers in the 90s and back then hackers like Kevin Mitnick and Kevin Poulsen were all the rage. They all faced the law and prison sentences. What's weird now is that we have something between gross negligence and malice, and nothing is being done, except maybe coordinated consolidation of AI power under the guise of "safety".