Hacker News Reader

Comment by roenxi | Hacker News Reader

This is really exciting! They're laying out an architecture that may mean even small players with cheap GPUs can compete with the majors. The idea implies that eventually crowd-sourcing an open AI is probably technically feasible and we've got the Chinese actively researching how to do it to a high standard that competes with the monolithic models.

I was sceptical of the US sanctions but this seems like a real win if this can be taken all the way to its logical conclusions.

View on HN

Yeah the sanctions will (not sarcastically) actually improve the world on a number of fronts. Increasing diversity of compute, forcing decentralization of manufacturing, etc. etc.

The sanctions will (not sarcastically) massively harm the world because Nvidia may no longer be a free money cheat code. I like having an easy economic strategy for investing...

The world doesn’t have to optimize policy to increase the profits of a single American company.

Chinese stocks are pretty reasonable right now, if their market has dealt with the insider trader mess then it might be a good time to onboard. It isn’t for the feint of heart however.

Markets used to be places to make money more smart (efficient allocation of capital) but have somehow degraded to index fund buys that track average economic growth of a few hot stocks that are expected to at least not get cold anytime soon.

also increase smuggling, theft, espionage, crime, sabotage.

There are much better ways to increase diversity

PRESIDENT TRUMP: "You don’t think we can. You don’t think we do that to them? We do. So we do a lot of things." https://singjupost.com/transcript-maria-bartiromo-interviews...

This is a mistaken belief. Sure you get all those negative aspects to being degrees, just like you get them under all other conditions … Chinese, Russian, Israeli espionage over the last ~80 years, anyone?… but you cannot actually get diversity without isolation that permits actual diversity to emerge.

Diversity is not pouring oil into water and using the polluted oil-water in lieu of oil and also in lieu of oil. If you want actual diversity you need differences that are separated from each other. It is precisely what has been collapsing for the last 80+ years, actual real diversity, precisely because unique separate groups and clusters have been shattered, scattered, mixed, and polluted.

Even AI is now accelerating this collapse of what is really a form of human biodiversity, or should it be called cultural diversity, as AI is causing a conformity of thought. There are several reports and papers on that phenomenon already.

It’s absolutely ridiculous to claim that somehow those factors will increase over the prior situation simply because we increase actual, real diversity of unique things; not this fake, fraudulent, delusional diversity that has forced on us like a toxic sludge dump that has destroyed human diversity as everyone increasingly consumes the same “content” slop and eats the same food slop, and has the same cultural and musical slop.

Nit-pick: smuggling is when you import goods into a country without informing the relevant government bodies. When it comes to GPUs, it's one country that has declared an export ban. Chinese port authorities wont care if you declare you're importing a container with 16000 Nvidia GPUs as that's still legal.

>The idea implies that eventually crowd-sourcing an open AI is probably technically feasible

It's already technically feasible: https://www.primeintellect.ai/blog/intellect-2

People are doing it: https://nousresearch.com/nous-psyche/

Deepseek-R1 is at the level of GPT 4.1 already, it's open-weight, open-source and they even open-sourced their inference code.

I don't know why everyone keeps echoing this, my experience with Deepseek-R1, from a coding perspective at least, has been underwhelming at best. Much better experience with GPT 4.1 (and even better with Claude, but that's a different price category).

I'm not arguing which model is better for your use-case. I'm saying in general as it's "powerful" as GPT 4.1 in a lot of benchmarks, and you can peak underneath the hood, even make it better for your said use-case

Do you mean V3? V3 is 4.1 level or above.

I got the impression that 03-mini or 03-mini-high were meant for coding? GPT 4.1 was meant for creative writing, not coding?

It’s good at a lot of things:

  GPT‑4.1 scores 54.6% on SWE-bench Verified, improving by 21.4%abs over GPT‑4o and 26.6%abs over GPT‑4.5—making it a leading model for coding.

https://openai.com/index/gpt-4-1/

They are trained on these "benchmarks", that's why they score better.

If they were trained on those benchmarks they would score 100%

They show how bad they are, they cannot score 100% on benchmarks they were trained on.

This is everyone with every model.

People sang praise from the roof for Google's Gemini 2.5 models, but in many things for me they can't even beat Deepseek V3.

What would be an example of 2.5 Pro failing against R1 (which is what you'd actually want to compare it to)?

R1 sometimes fails against V3 for me too, so its not a specific dig against Gemini.

In terms of code and science, Gemini is way, way too verbose in its output, and because of that it ends up getting confused by itself and hurting the quality of longer windows.

R1 does this too, but it poisons itself in the reasoning loop. You can see it during the streaming, literally criss-crossing its thoughts and thinking itself into loops before it finally arrives at an answer.

On top of that, both R1 and Gemini Pro / Flash are mediocre at anything creative. I can accept that from R1, since it's mainly meant as more of a "hard sciences" model, but Gemini is meant to be an all-purpose model.

If you pit Gemini, Deepseek R1 and Deepseek V3 against each other in a writing contest, V3 will blow both of them out of the water.

Agreed on the last point, V3 is terrifyingly good at narrative writing. And yes, R1 talks itself out of correct answers almost as often as it talks itself into them.

But in general 2.5 Pro is an extremely strong model. It may lose out in some respects to o3-pro, but o3-pro is so much slower that its utility tends to be limited by my own attention span. I don't think either would have much to fear from V3, though, except possibly in the area of short fiction composition.

In my experience, all reasoning models feel (vibely) worse at structured output like code versus comparable non-reasoning models, but far better at knowledge-based answering.

A lot of software (eg. ollama) has confusingly named Deepseek's distill/finetunes of other base models "DeepSeek-R1" as well. See eg. https://www.threads.com/@si.fong/post/DKSdUOHzaBB

I wonder whether you're actually running the proper DeepSeek-R1 model, or one of those lesser finetunes?

SETI@Home style peer2peer open GPU training network is something I’m looking into as well.

Possible and has been done, but super-slow and inefficient resulting in long training times for small models. To keep compute occupied you need to pass gradients very fast.

Do you mean this one?

https://blog.lambdaclass.com/introducing-demo-decoupled-mome...

This is what piqued my interest in the first place

Yes but could you break it up into chunks of sets of gradients to compute? I know that compute needs the full chunk to compute a set. Again, things I’m exploring but ultimately no different than just having the full dataset on disk and just scaling out compute nodes in ro mode.

I suppose its exciting, but whether that is a good thing depends entirely on how much you think AI technologies pose existential threats to human survival. This may sound hyperbolic, but serious people are seriously thinking about this and are seriously afraid.

https://i.imgur.com/we5ggtS.jpeg

Reply on news.ycombinator.com