Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

DeepSeek V4 – almost on the frontier

simonwillison.net

Post reply on HN
Print view View original

411–420 of 420 posts

Re: DeepSeek V4 – almost on the frontier

#411

Post by itemize123 »

zozbot234 wrote: post #137
Gareth321 wrote: post #135

Earlier quoted context omitted.

Isn't it widely speculated that these are distilled from current frontier models? Distillation is far less compute intensive than primary training. That said, if distillation produces something almost as good for a fraction of the cost, Jensen's point may stand.

You can't really distill a model without access to the internal weights. You could train on chat logs, but that's absolutely not the same thing, it doesn't even come close to comprehensively "extracting" the model's capabilities. And everyone does that in the industry anyway ever since ChatGPT was first released, some versions of Opus even claimed to be DeepSeek if you prompted them in Chinese.

actually distillation is without weights - you basically just need a black box teacher model.

Re: DeepSeek V4 – almost on the frontier

#412

Post by itemize123 »

Aldipower wrote: post #363
npv789 wrote: post #333

my default model now, less censorship

Just ask it about the "Tiananmen Square protests and massacre". :-) On the other hand, asking ChatGPT about "Hiroshima US atomic bombs", isn't too much better.

run it hosted on another provider - it's less censored there

Re: DeepSeek V4 – almost on the frontier

#413

Post by Gareth321 »

zozbot234 wrote: post #137
Gareth321 wrote: post #135

Earlier quoted context omitted.

Isn't it widely speculated that these are distilled from current frontier models? Distillation is far less compute intensive than primary training. That said, if distillation produces something almost as good for a fraction of the cost, Jensen's point may stand.

You can't really distill a model without access to the internal weights. You could train on chat logs, but that's absolutely not the same thing, it doesn't even come close to comprehensively "extracting" the model's capabilities. And everyone does that in the industry anyway ever since ChatGPT was first released, some versions of Opus even claimed to be DeepSeek if you prompted them in Chinese.

Weights are not required for distillation. I'm not sure how you came to that belief. Distillation is training a student model to minimic a teacher model output.

Anthropic, for example, posted a 2026 disclosure (https://www.anthropic.com/news/detecting-and-preventing-dist...) which singles out DeepSeek's distillation activity. They detected over 16M actions over 24,000 fraudulent accounts. That's just what they detected.

Re: DeepSeek V4 – almost on the frontier

#414

Post by oveja »

antirez wrote: post #250
dust42 wrote: post #249

Earlier quoted context omitted.

For many models the performance of llama.cpp on Mac is 20-40% lower than MLX. Did you try MLX? At least on HF there are MLX 2-bit quants. Unfortunately I have only 64GB, so I can't test it.

I'm not using llama.cpp there, it's my inference engine that is DeepSeek v4 specific. The goal is to optimize it as much as possible.

That's cool!

I knew the name sounded familiar, thank you for SDS!

Re: DeepSeek V4 – almost on the frontier

#416

Post by Reviving1514 »

gmerc wrote: post #334
Reviving1514 wrote: post #61

Earlier quoted context omitted.

What provider are you using? I have it a shot through open router and saw some weird half formed words coming through occasionally, would love to switch over and give it a proper go

Direct API

Thank you!

Re: DeepSeek V4 – almost on the frontier

#417

Post by deaux »

never_inline wrote: post #117
deaux wrote: post #14

I'm surprised that people here don't care at all about these models openly training on your data, especially if you use them straight from the model developer. Whereas things like "GitHub now automatically opts everyone into using their code for model training" get hundreds of justifiably angry comments, I never see this brought up anymore on posts like these talking about using Chinese models through OpenRouter. Thi…

I am fine with them training on my open source code (which is pretty bad but not the point, because they're providing the service for free). I will be super pissed if I pay for enterprise and they train on it though. I believe this is the opinion of majority programmers.

At least Moonshot (Kimi) says in the ToS that they train on your prompts when using their paid API.

Re: DeepSeek V4 – almost on the frontier

#418

Post by trollbridge »

wg0 wrote: post #9

Deepseek v4 Pro feels like Claude Opus 4.6 in it's personality but here's what I did find out about costs: I did cut loose Deepseek v4 on a decent sized Typescript codebase and asked it to only focus on a single endpoint and go in depth on it layer by layer (API, DTOs, service, database models) and form a complete picture of types involved and introduced and ensure no adhoc types are being introduced. It developed a…

DeepSeek V4 Pro's pricing is blowing me away, particularly with how effective the cache is. I just burned 2M tokens and the total cost was 30¢. On Claude Code, I'd have used up multiple 5 hour windows by now, or else horrific amounts of API consumption, around $20-$30 I'm guessing.

Re: DeepSeek V4 – almost on the frontier

#419

Post by Wilder7977 »

motoxpro wrote: post #344
Wilder7977 wrote: post #252

Earlier quoted context omitted.

It's not though. It's still just a piece of code, much closer to IDEs or any other program than to a human assistant in any way that matters (morals, responsibility).

It just seems like you are saying if you found out Claude code was a bunch of remote working doing work for you, then it would be morally wrong to do illegal/morally wrong/irresponsible things with them, but because it is NOT a human, those same things are fine?

Yes, correct.

Is the distinction between human labor/actions and a program executing hard to grasp?

Moral is a human thing, not an absolute thing, so of course it's different if there is a single human involved and a tool, and a human with a relationship to other humans.

Re: DeepSeek V4 – almost on the frontier

#420

Post by motoxpro »

Wilder7977 wrote: post #419
motoxpro wrote: post #344

Earlier quoted context omitted.

It just seems like you are saying if you found out Claude code was a bunch of remote working doing work for you, then it would be morally wrong to do illegal/morally wrong/irresponsible things with them, but because it is NOT a human, those same things are fine?

Yes, correct. Is the distinction between human labor/actions and a program executing hard to grasp? Moral is a human thing, not an absolute thing, so of course it's different if there is a single human involved and a tool, and a human with a relationship to other humans.

I just have different moral preferences. I think its morally wrong to do illegal/morally wrong/irresponsible things in general, whether I am using a hammer, a car, a company, or AI.

It's worse to ask someone else to break a window with a hammer for me, but the window still got broken, and the person whose window it was is still sad/out of money/etc.

The thought experiment was that if you were doing illegal things with an AI, you would not feel bad, but if you found out that the AI was a person, you would feel bad. That is very strange to me, more a feeling of guilt/shame.

Post reply on HN