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

301–310 of 420 posts

Re: DeepSeek V4 – almost on the frontier

#301

Post by 0xbadcafebee »

eckelhesten wrote: post #288
soerxpso wrote: post #141

Earlier quoted context omitted.

Some providers are based in the US or EU and would face legal repercussions for lying about what they do with your data. It's a bit more than "trust me bro". Off the top of my head, you can use Fireworks, for example, which is based in California and would face the same consequences for lying about their data policy as OpenAI or Anthropic would.

Meta is based in the US, yet they torrented TERABYTES worth of books to feed their AI. I’m not trying to be negative here, but your point is invalidated by that particular event in itself.

What, because they broke the law in one way, they'd break the law in every way? That's not how business works. The way business works is, I steal from other people to make a product, but then I don't steal from my customers, because if they find out, then I no longer have any customers. (Plus all their customers would sue them, which would both legally and financially tank them)

Re: DeepSeek V4 – almost on the frontier

#302

Post by forrestthewoods »

gpugreg wrote: post #289
forrestthewoods wrote: post #251

Naive Question: is DeepSeek V4 actually cheaper to run? Or is it cheaper because of other reasons? For example Anthropic running at a higher margin or DeepSeek at a larger loss?

I believe that DeepSeek-V4-Pro API at promotional pricing ( https://api-docs.deepseek.com/quick_start/pricing ) could run at almost exactly 200 % profit. If you take DeepSeek's numbers for DeepSeek-V3 ( https://github.com/deepseek-ai/open-infra-index/blob/main/20... ) and plug in ~3333 tps/GPU for DeepSeek-V4-Pro ( https://developer.nvidia.com/blog/build-with-deepseek-v4-usi... ) and a price of $7/hr per B300 GPU, th…

Good info, thanks! (Not sure why my original question got downvoted. It’s very fair to ask imho!)

Re: DeepSeek V4 – almost on the frontier

#303

Post by loehnsberg »

kamikazechaser wrote: post #230
cedws wrote: post #113

The biggest differentiator for me: DeepSeek just does what I ask. I've tried using both GPT and Claude for reverse engineering recently, both refused. I even got a warning on my OpenAI account.

In my experience GLM 5.1 has been excellent when paired with IDA Pro (DeepSeek v4 pro comes in close second, Kimi straight up refuses). Claude can only do reverse engineering if you throw it into some sort of hero/saviour mode then gradually pivot into red team (though it gets easily tripped).

Among the inexpensive models (and I include Grok 4.3 in this list), GLM 5.1 really sticks out!

On my personal test bench, when compared to other inexpensive models, GLM 5.1 provides the answers that I would consider most complete or satisfying (these are subjects that I consider myself an expert in). The answers tend to be more comprehensive, nuanced, and include references that I would consider the correct ones (if given access to web search).

I also find it a joy to code with, somewhere between Sonnet 4.6 and Opus 4.6 (have not tested Opus 4.7 yet).

Finally, just gauging by pelicans, it kind of stick out: https://simonwillison.net/tags/pelican-riding-a-bicycle/

Re: DeepSeek V4 – almost on the frontier

#304

Post by ethbr1 »

Wowfunhappy wrote: post #219
ethbr1 wrote: post #215

Earlier quoted context omitted.

This. It's going to be tricky for the frontier model labs to argue they didn't intentionally design their models to do so, when the models take illegal actions. I'm not even sure how one would construct a viable legal argument around that for SOTA models + harnesses, given the amount of creative choices that go into building them. It'd be something like "Yes, we spent billions of dollars and thousands of person-hours…

> I'm not even sure how one would construct a viable legal argument around that for SOTA models + harnesses, given the amount of creative choices that go into building them. I'm not a lawyer, but to me the legal case seems pretty obvious. "We spent billions of dollars creating this thing to be a good programmer, but we did not intend for it to reverse engineer Oracle's database. No creative effort was spent making it…

> No creative effort was spent making it good at reverse engineering Oracle's database.

That's the bit that's going to be nasty in evidence. 'So you didn't have any reverse engineering in your training or testing sets?'

Re: DeepSeek V4 – almost on the frontier

#305

Post by gpugreg »

forrestthewoods wrote: post #302
gpugreg wrote: post #289

Earlier quoted context omitted.

I believe that DeepSeek-V4-Pro API at promotional pricing ( https://api-docs.deepseek.com/quick_start/pricing ) could run at almost exactly 200 % profit. If you take DeepSeek's numbers for DeepSeek-V3 ( https://github.com/deepseek-ai/open-infra-index/blob/main/20... ) and plug in ~3333 tps/GPU for DeepSeek-V4-Pro ( https://developer.nvidia.com/blog/build-with-deepseek-v4-usi... ) and a price of $7/hr per B300 GPU, th…

Good info, thanks! (Not sure why my original question got downvoted. It’s very fair to ask imho!)

Probably nothing personal. It feels like the climate of HN is shifting towards more negativity (and less quality) during the last few months.

Re: DeepSeek V4 – almost on the frontier

#306

Post by itissid »

So RPI/QRSPI like skills (e.g. https://github.com/mattpocock/skills and https://github.com/humanlayer/humanlayer/tree/main/.claude/c... and https://github.com/dfrysinger/qrspi-plus ) for working with claude code work well enough for me that they can reliably* produce code that matches the plan/spec in a way they did not till December 2025.

I have a gut feeling that these models can do just as well, has someone run a reasonable size task — >=1-2 days of designing and planning — and see it work well with these models?

* For me what worked well was the grill me skill(or its variation) at the design stage, the hygiene I followed here was have it ask one question at a time, resolving dependencies at the design stage and reading the hashed out plan closely. The use of a couple of other MCP tools like a documentation server like deepwiki and arxiv for grounding. Other tricks I use are having high signal tests and having claude either be able to read logs and code at the same time or embedding it in the execution(e.g. as a debugger, repl or devtools)

Re: DeepSeek V4 – almost on the frontier

#310

Post by varispeed »

cedws wrote: post #113

The biggest differentiator for me: DeepSeek just does what I ask. I've tried using both GPT and Claude for reverse engineering recently, both refused. I even got a warning on my OpenAI account.

I myself got refusals often for legitimate data analysis work. I am starting to lean on buying powerful hardware little by little until I get suitable rig to run local models that make sense.
Post reply on HN