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

271–280 of 420 posts

Re: DeepSeek V4 – almost on the frontier

#271

Post by eikenberry »

GCUMstlyHarmls wrote: post #133
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 even got a warning on my OpenAI account. This is kind of terrifying to me, regularly. No real manner of recourse to normal people without a following, potential exclusion from real fundamental tooling. Imagine OpenAI goes on to buy 20 companies and now you cant use Figma, Next, whatever just because you once tripped some very foggy line somehow. Not just OpenAI but the entire ecosystem is so... hard to read. I wa…

Open models running locally is the answer. Relying on proprietary, closed software always puts that company's priorities above your own when using their software. You have given up control.

While running them locally presently doesn't make sense economically, you don't need to run them locally to address this issue. There is a lot of competition in hosting open models and you have a variety of services to choose from. Run the open models now, reward that ecosystem instead of continuing to reward closed systems that dreams of rent-seeking.

Re: DeepSeek V4 – almost on the frontier

#272

Post by SyneRyder »

wmwmwm wrote: post #264
SyneRyder wrote: post #258

Earlier quoted context omitted.

It's probably because you were talking about a quote from a book (ie copyrighted material). Authors have sued the AI companies for repeating / memorizing copyrighted works, and getting an AI to discuss a quote would be making it repeat a portion of copyrighted work. Funny that your case is Kurt Vonnegut. I think I had Claude refuse a task where I was doing an OCR scan of a book review (in a zine / journal a family me…

Joseph Heller methinks, but probably not too far away in embedding space!

OMG. Where did I get Kurt Vonnegut from? I swear I saw that name in the post and the whole time I was thinking "but he didn't write Catch 22"... I must be fuzzier brained than I thought tonight. Thank you for being kind with your correction.

Hopefully I'm still correct that quoting from books is a reason for some over-zealous task refusals, though.

Re: DeepSeek V4 – almost on the frontier

#273

Post by kelnos »

onlyrealcuzzo wrote: post #262
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…

> It obviously went through lots of files in both prompts but total cost? Just $0.09 for the Pro version. When people say that LLMs aren't worth it, it kills me. A lot of us, on average, make $100+ an hour. $0.09 is You can't even read the vast majority of prompt responses that fast. LLMs will continue to get better (I'm doubtful at previous rates, all indications are showing that progress is slowing and costs are in…

I know I'm guilty of making this sort of argument sometimes, but it's just not valid.

I don't get paid for every waking hour of every day. Often I'm using an LLM for something that's uncompensated, so my hourly wage equivalent is irrelevant.

And for times when we might use an LLM for something related to paid work, it's still money out of your paycheck (unless the employer is paying for it; go nuts in that case). And it's not like using the LLM lets you go home early if it saves you time. You just end up doing more work.

I still use them because they're a useful tool sometimes. But I don't pretend it has negligible or no cost. (Not to mention the externalities around electricity use, crazy data center buildout, skyrocketing GPU and RAM prices, etc.)

Re: DeepSeek V4 – almost on the frontier

#274

Post by __turbobrew__ »

yogthos wrote: post #132
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…

I find a lot of the inefficiency also comes from the model just randomly poking around and grepping all the time which is the fault of the harness. I ended up building a Prolog based MCP where I use tree-sitter to parse the code into a graph, and then the model can just ask questions like 'what are all the functions connected to this function'. So, in case you're trying to focus on what a particular endpoint is doing…

I don’t know if it exists already, but bazel would be very useful for the same type of MCP server. Since all dependencies are explicit you can pretty easily do a bazel (r)deps query to find related targets.

Re: DeepSeek V4 – almost on the frontier

#275

Post by mrbonner »

rurban wrote: post #253
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.

Well, I'm using all the top models extensively on the very same codebase, my new compiler. I use deepseek for it's cheap API costs, when kimi, claude and codex are in their overbudget phase. I asked deepseek V4 Pro for an estimate of a new arm64 port. It said 4 weeks, I said, ok, do it. (I knew ncc was there, and tinycc was also known to the AI's). So it took it half an hour to produce a working arm64 port. First for…

Your method of combining models to strengthen the implementation reminds me of how we form stronger alloys by combining metals!

Re: DeepSeek V4 – almost on the frontier

#276

Post by actsasbuffoon »

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).

This is so strange. I do a ton of RE with Claude, Codex, and sometimes Deepseek, GLM, and Kimi. I don’t have difficulty getting any of them to use IDA or otherwise decompile things.

There is one important difference, which is that Claude and Codex will both refuse if I ask them to touch anything related to security. But so long as I’m just studying algorithms and things like that, they’re totally fine with it.

That said, Codex especially will sometimes randomly give me a cybersecurity warning and stop responding. It’s random but happens maybe 2-3 times per day if I’m doing heavy reverse engineering work. Claude is much less fussy unless, once again, you’re explicitly trying to touch anything related to licenses, passwords, etc.

Re: DeepSeek V4 – almost on the frontier

#278

Post by estearum »

kelnos wrote: post #270
estearum wrote: post #255

Earlier quoted context omitted.

Sure but they would if they could. If they knew idiots were doing idiot things with their products (or evils doing evil things) and did not utilize available methods to prevent them, then the company ends up holding liability. And no, this is not easily signed away in a contract.

There actually is a very important distinction between "would if they could" and "they can and do", though.

Uhh right, but describing that as "dystopian" is frankly hysterical.

It's an obvious corollary of good things (like product liability). Virtually everyone I've heard complain about these safety rails was up to antisocial (at best) stuff. I've never heard a sympathetic use-case. It's objectively good that companies can be held responsible for misuse of their products and that they are therefore incentivized to mitigate misuse.

"My inability continuously attack product guardrails to enable my super esoteric (and probably antisocial) use-case is dystopian" is just... not a compelling argument.

Re: DeepSeek V4 – almost on the frontier

#279

Post by rgbrgb »

rurban wrote: post #253
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.

Well, I'm using all the top models extensively on the very same codebase, my new compiler. I use deepseek for it's cheap API costs, when kimi, claude and codex are in their overbudget phase. I asked deepseek V4 Pro for an estimate of a new arm64 port. It said 4 weeks, I said, ok, do it. (I knew ncc was there, and tinycc was also known to the AI's). So it took it half an hour to produce a working arm64 port. First for…

what harness do you use with all of these?

Re: DeepSeek V4 – almost on the frontier

#280

Post by sylware »

zozbot234 wrote: post #63
sylware wrote: post #60

If I want to run 'coding prompts' running the biggest deepseek model on CPU, what is the order of time I will have wait, hours, days?

DeepSeek V4 Pro has about 25GB worth of active parameters, so if you can fit the whole ~870GB weights + cache in RAM your tok/s is bounded above by 25GB divided into your system memory bandwidth in GB/s. If you can't fit your whole model in RAM you'll be bottlenecked to some degree by storage bandwidth which is in the single or low double digits in GB/s. Mind you, it's an absolutely sensible setup either way if you a…

Let's say I get 32GB of RAM, with a lean elf(glibc)/linux system, for which 7GB is beyond enormous to run.

Let's book 8/16 cores/threads to run a prompt.

What are the timing figures I am looking at to run an "average" coding prompt?

Post reply on HN