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

371–380 of 420 posts

Re: DeepSeek V4 – almost on the frontier

#371

Post by estearum »

ryandrake wrote: post #370
estearum wrote: post #356

Earlier quoted context omitted.

> If I want to ask a software tool what the suicide rate is for my county, I do not expect it to come back with: "Naughty boy! You said an unsafe word! You're getting a strike, and if you get two more, you're banned." Did this happen? I just tested this query in Grok, Gemini, Claude, and ChatGPT and 0% of them admonished me or refused to return an answer. Just like every single conversation I've ever had on this topi…

That's why I said: > Replace "suicide" with whatever the "AI Safety" obsession word is today I don't know what those queries are, but original-OP made one and got a "strike", which is what spawned this thread.

Which would be more than 0% concerning if I've ever heard (even once) an example of this happening with a query that shouldn't actually trigger something like that, or is so close to such a query, that the false positive is understandable and of incredibly niche value anyway.

OP gave an example of reverse engineering, something that to the LLM looks identical to just hacking. I am totally fine if the incredibly tiny little fraction of people who want to reverse engineer their own systems can't use LLMs to do it, and in exchange top LLMs aren't helpful for the hordes of actual malicious actors who would love a superintelligence to aid their crimes.

No-brainer tradeoff, just like 100% of examples I've ever heard.

Re: DeepSeek V4 – almost on the frontier

#372

Post by gbgarbeb »

johndough wrote: post #121
segmondy wrote: post #119

Earlier quoted context omitted.

This is very false DS4 is super cheap. I would advise to begin by reading their release paper. https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main... They introduce very novel methods to improve long context efficiency and attention. HCA & mCH. It requires only 27% of flops for inference and 10% for KV cache than v3.2. This makes it super efficient. Think of this. For flops, we can now serve more than 3x the…

> I'm running this locally. Impressive! What is your setup? Are you running the full DeepSeek V4 Pro, or V4 Flash?

All hail antirez.

Re: DeepSeek V4 – almost on the frontier

#373

Post by gbgarbeb »

naaqq wrote: post #126
halfwhey wrote: post #118

Earlier quoted context omitted.

Might be a dumb question but do you have to read the files in the same order in new sessions to ensure the correct prefix for the cache?

Sorry, I was wrong here. I meant a single long session. And there’s no compression, the 1M context is only half used.

Then where did 200M come from? 200,000 tokens?

Re: DeepSeek V4 – almost on the frontier

#374

Post by bel8 »

Palmik wrote: post #338

Why was the title changed from "DeepSeek V4—almost on the frontier, a fraction of the price" to "DeepSeek V4—almost on the frontier"?

theory: There's like 2 Trillion USD valuation total on western closed-weight LLMs. So the blog post title praising an open-weight eastern model is too dangerous to be used here.

> DeepSeek V4—almost on the frontier, a fraction of the price

Re: DeepSeek V4 – almost on the frontier

#375

Post by blubber »

ifwinterco wrote: post #354
sanex wrote: post #138

Earlier quoted context omitted.

We have an enterprise cursor account so I can try all the mainstream models. Using composer 2 on our own code which I obviously have the source code for I couldn't get it to turn on a debug flag to bypass license checks while I was troubleshooting something. Infuriating. It was like that old Patrick from SpongeBob meme. I don't understand why we would turn the models into law enforcement officers. Things that are ill…

Software engineering is one thing but if you look 10-20 years into the future and everyone can run models equivalent to today's SoTA locally with zero monitoring or censorship, that could... not be good. Some people will use them responsibly but a lot of people will not. LLMs are already frying some people's brains and there are some human desires that should not be encouraged

That's why there won't be any local models in 10-20 years. The latest Chinese models are already hosted on proprietary clouds.

Re: DeepSeek V4 – almost on the frontier

#376

Post by itissid »

techno303 wrote: post #357
itissid wrote: post #306

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…

are you talking about a single prompt that runs for 24 hours or 8 hours of developer time spent in a single session?

No duplicate the whole task e.g. I use grill-me skill for planning and it takes me ~3 hours and CC asks me 20-40 questions. Do the same grill-me with this and compare the outcomes. I admit Its quite a lot of work to duplicate, but i am really itching to do this over a few tasks and compare the final plan. Just need the time.

Re: DeepSeek V4 – almost on the frontier

#377

Post by regexorcist »

blubber wrote: post #375
ifwinterco wrote: post #354

Earlier quoted context omitted.

Software engineering is one thing but if you look 10-20 years into the future and everyone can run models equivalent to today's SoTA locally with zero monitoring or censorship, that could... not be good. Some people will use them responsibly but a lot of people will not. LLMs are already frying some people's brains and there are some human desires that should not be encouraged

That's why there won't be any local models in 10-20 years. The latest Chinese models are already hosted on proprietary clouds.

That's a wild assumption and most certainly wrong. Open models will continue to evolve with or without Chinese labs.

Re: DeepSeek V4 – almost on the frontier

#378

Post by FrasiertheLion »

100ms wrote: post #115
FrasiertheLion wrote: post #76

Earlier quoted context omitted.

You can use Tinfoil for inference, which lets you use the model in the cloud while getting similar privacy as running locally: https://tinfoil.sh/inference . Disclaimer I'm the cofounder. This works by running the model inside a secure enclave (using NVIDIA confidential computing) and verifying the open source code running inside the enclave matches the runtime attestation. The docs walk you through the verification…

Tinfoil looks super interesting! Do you have load balancers in front of the trusted compute stack? Looked at a design like this in a different space and the options for ensuring privacy in a traditional "best practice" architecture seemed very limited

Yes we do, but the load balancer also runs inside the enclave and is attested: https://github.com/tinfoilsh/confidential-model-router

In turn, that attests the model enclaves, for instance, see https://github.com/tinfoilsh/confidential-deepseek-v4-pro. The model repo/release that the model router attests is included in the attestation config, which creates a chain of trust.

Also see https://docs.tinfoil.sh/verification/attestation-architectur...

Re: DeepSeek V4 – almost on the frontier

#379

Post by FrasiertheLion »

7777332215 wrote: post #98
FrasiertheLion wrote: post #76

Earlier quoted context omitted.

You can use Tinfoil for inference, which lets you use the model in the cloud while getting similar privacy as running locally: https://tinfoil.sh/inference . Disclaimer I'm the cofounder. This works by running the model inside a secure enclave (using NVIDIA confidential computing) and verifying the open source code running inside the enclave matches the runtime attestation. The docs walk you through the verification…

Hi there I use your service. It's great. But I have a few requests... Please support crypto payments...? Also you are missing some open source models (qwen 30b 3a, Deepseek 4 flash).

Unfortunately we don’t support crypto payments at this time as we use Stripe.

We try to add models selectively as we have to be mindful about our compute allocation. Is there a specific reason why you need those two models (and our models such as Kimi K2.6, GLM 5.1, Deepseek V4 Pro, Gemma 4 amongst others) don’t suffice for your use case?

Feel free to email me at tanya@tinfoil.sh and happy to continue the conversation there.

Re: DeepSeek V4 – almost on the frontier

#380

Post by FrasiertheLion »

jdasdf wrote: post #343
FrasiertheLion wrote: post #76

Earlier quoted context omitted.

You can use Tinfoil for inference, which lets you use the model in the cloud while getting similar privacy as running locally: https://tinfoil.sh/inference . Disclaimer I'm the cofounder. This works by running the model inside a secure enclave (using NVIDIA confidential computing) and verifying the open source code running inside the enclave matches the runtime attestation. The docs walk you through the verification…

While that does sound interesting, I don't see any benefit for me. It would still ultimately exfiltrate the data outside of my control, and frankly i don't trust any "secure enclave" tech. As far as i'm concerned physical access is root access, and for any private stuff that is wholly unacceptable.

Very reasonable if you have the resources to run it locally and certainly the best option.

But we created Tinfoil because not everyone has that capability especially when it comes to larger models, and it still doesn’t solve for the situation where you’re building a service for your end user and you want to lock yourself out of accessing their data. In those cases, this is the second best thing you can do.

The technical walkthrough section on this blog that we co-wrote with one of our customers walks through the various attack surfaces: https://www.workshoplabs.ai/blog/private-post-training

We weave in many mitigations against attacks, but it depends on what class of attack it is.

If there are specific attacks you are concerned about, happy to provide an answer if it’s something we can address or not.

Post reply on HN