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?
DeepSeek V4 – almost on the frontier
251–260 of 420 posts
Re: DeepSeek V4 – almost on the frontier
#252Earlier quoted context omitted.
> I even got a warning on my OpenAI account. This idea of software threatening the user with consequences is totally wild and dystopian. Fellow developers, what kind of world have be built? This is insanity. Imagine if my hammer told me, "Hey, you shouldn't use me on screws--only nails. Do it again and I'll self-destruct!" WTF people, stop making this kind of software!
I think it's closer to asking a remote (human) assistant to do something that someone doesn't want done (e.g., view the source of a closed-source product, whether through reverse engineering, going into their office, or social engineering) and that remote assistant company saying, "Please stop asking our assistants to do that." You can still use an IDE (hammer) to reverse engineer anything you want.
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).
Re: DeepSeek V4 – almost on the frontier
#253The 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 arm64-elf, because this was easiest to test, and then also after more hours of back and forth the arm64-darwin port. (with crossbuild and github actions). It did cost me with all the subsequent fixes around $8 API costs.
So the experience: at the beginning deepseek was amazing. When it started to get expensive (china day time), I switched from Pro to Flash. No problem, same results. Some bitfield implementation was too complicated so I had to wait for Sonnet 4.6 tokens, kimi-2.6 did the rest. For the very hard problems I asked gpt-5.5, but this was only for one problem. minmax was horrible. didnt follow rules, and made lot of silly stuff.
But when the deepseek context window got filled, deepseek also started to become stupid. So either /clear, or /export and strip the file. And start a new session with the cleared sessions. kimi was overall better, but running into limits with my cheap moderate subscription. Paying private for it, as my companies' token budget is usually out after a week of work.
All in all it is worth it. My next compilers (perl 5+6=11) will be done with deepseek and kimi also.
regarding decompilation: recently we had to decompile a firmware for a USV we bought, but doesnt work on a new system. It only worked on a raspi. So I decompiled it with ghidra, and told my colleague, easy, that's how you do it. But my colleage didnt know about token budgets yet, and already threw opus at it. CoPilot Business account. He had working C files immediately, compilable for our new system. It ended up the USV was not beefy enough. But Opus was fantastic. The code was very short and simple C though.
Re: DeepSeek V4 – almost on the frontier
#254Re: DeepSeek V4 – almost on the frontier
#255Earlier quoted context omitted.
All sorts of tools try to prevent dangerous/destructive uses In fact probably every single piece of commercial software you use had you sign a contract saying you wouldn’t do it
> All sorts of tools try to prevent dangerous/destructive uses But they don't threaten their users or have an "N strikes and you're out" policy. I take those safety caps off of all the chemicals in my garage because I'm a grown-ass adult and those caps are a pain in the butt. I would not expect the manufacturer of a solvent to show up at my house lecturing me about safety and threatening to ban me from buying his pro…
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.
Re: DeepSeek V4 – almost on the frontier
#256… waiting patiently for llama.cpp support to land
Re: DeepSeek V4 – almost on the frontier
#257Earlier quoted context omitted.
We need that lawsuit to happen already so we can establish precedent. The person in the driver's seat of the Tesla should be at fault. The engineer using the llm should be at fault. The person behind the gun not the manufacturer should be at fault.
In the America, whoever has the most money is liable. It's not worth it for the legal industry otherwise. The lawyer earns his pay by convincing the court that whatever established precedent doesn't apply to his case.
Unfortunately.
Re: DeepSeek V4 – almost on the frontier
#258The 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…
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 member published years ago). I think the review might have included a Vonnegut quote as well, and that I ultimately figured it out it was the quote that was making Claude refuse. I may be misremembering the author though.
Mistral had no such refusals, but their OCR is lesser quality.
Re: DeepSeek V4 – almost on the frontier
#259I realize this post is about the pelican test, but in regards to coding, has anyone tried out the advisor strategy with V4?[0] e.g. Have V4 call out to Opus when it's uncertain, but otherwise handle execution. The results with Sonnet/Haiku in the blog post seemed promising, so I'm curious how it would go with these latest open models. [0] https://claude.com/blog/the-advisor-strategy
That first graph (SWE-bench Multilingual) is a crime
Re: DeepSeek V4 – almost on the frontier
#260Earlier quoted context omitted.
Have you given GLM 5.1 or Kimi K2.6 a shot for coding? They outperform Deepseek v4 pro.
> Kimi K2.6 a shot for coding? They outperform Deepseek v4 pro I think this probably depends quite a bit on the specific problem. I'm finding that Deepseek v4 Flash often outdoes Kimi 2.6 on a variety of coding problems that involve complex spatial reasoning
Really? I've found kimi k2.6 to be really good for vision and spatial stuff. Gemini has been the only subjectively better one but gemini isn't reliable in a loop