511 comments

[ 0.25 ms ] story [ 14.5 ms ] thread
Already on HuggingFace: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash

The bad news is that the original v4 flash was 284B, which was large but still somewhat reasonable for running locally. This one is 552B so almost twice that, so the huge gains in benchmark scores make sense - it's not really flash anymore, imo.

I've no idea about actual performance vs benchmaxxing, though deepseek was fairly trustworthy as far as Chinese models go. If that holds (and if it doesn't think forever, as deepseek 4 sometimes did) it's probably the newest king of the hill amongst open weights models.

It does include vision, and they do something funky with KV cache so it's very efficient: "[...] these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash". I do appreciate the high focus on efficiency, but at this point we sure could use a flash-flash version.

V4 Flash also was released as mostly FP4, but this one is FP8 (?). 160GB vs 510GB.

Original Flash good fit for dual Spark / Strix Halo machines. This one would require third party quants and even then 4 machines.

Edit: Most of added weights/size are Engrams?

> Overall, DeepSeek-V4.1-Flash has 552B backbone parameters and 196B Engram parameters, activating 8B parameters per token during prefill and 16B during decode.

Those can stay on SSD. So I guess / it possible, that non-engram portion is still FP4 of ~same size! Need to read tech report.

It's larger than previous V4 Flash.

  552B in ~FP4, 306GB.   
  196B of FP8 Engrams, another 204GB, not necessary to keep in RAM.  
  KV cache sees another 4x size reduction, just 900MB for 1M.  
So 384GB needed for a chance of achieving useful speeds. Three Sparks or quad RTX PRO 6000.
Question is how many of those experts one needs to keep in vram for a given workload.

I could imagine (though I might be _very_ wrong there) that for example coding does not live in all of them. Maybe 1/3? Do we have real numbers there?

So maybe one can get away without much performance penalty by doing some LRU stuff?

>This one is 552B so almost twice that, so the huge gains in benchmark scores make sense - it's not really flash anymore, imo.

It uses fewer active parameters, though. (8B or 14B instead of always 13B)

So ... flash indeed.

200B of those 552B is PLE, which works more like a database that is read for each token, thus can be offloaded to a fast SSD.
Id love an ELI5 for PLE. Im trying to work it into my back of the napikin math for compute vs memory bandwidth limitations on tok/s in PP vs TG work.

My attempt at a simplification of this article on it https://sebastianraschka.com/llm-architecture-gallery/per-la... into a couple of sentences is that they are linear embeddings of the input token space projected per layer, which are then gated by the transformer outputs per layer.

This would mean that the only one set of weights for the ple path needs to be pumped across the memory bandwidth as they are the same linear weights for all layers?

Sheit, maybe im trying to simplify something that i need to look at in detail. but id love to leverage others understanding if possible

>Id love an ELI5 for PLE.

PLE is, instead of mixture of experts, mixture of associations

Unfortunately no, it's 200B + 552B. It's not as bad as it sounds though, because most of 552B is in 4bit natively.
> It also includes additional 196B Engram memory which you can put on an SSD. I think

You can put Qwen 3.8 Flash Next engram on SSD, but prompt processing takes a good hit. On my mac studio, I get 300 pp and 33 tg with SSD offload, versus 550/40 with everything in RAM.

I will be very happy if 300 pp is achievable with this model though.

You can warm cache regularly used engram/n-gram if you're willing to merge PRs into a personal branch and build it yourself. I was trying this with qwen 3.8 flash next and the n-gram to get it to fit on my very average gaming desktop (it worked)
The engram stuff is great because RAM is often still cheaper (or at least expandable). My company does currently look into buying some hardware as we handle confidential data and code.

Qwen 3.8 Flash is viable on two Nvidia 6000 96GB with a wood quant because you can put the 50GB Engram into RAM and the hit should be below 10% performance. At least that is what I have seen so far. Correct me if I'm wrong.

I am running that on a single 6000 96GB with 4-bit quants for both weights and PLE table. Needs just 32GB RAM and fits snugly into the 96GB VRAM with KV cache equalling ~300k context tokens. Not sure if I quantized the KV
Actually this ought to run quite well with SSD streaming. The MoE expert sparsity seems to be similar to DSv4 Pro (hence exceptionally sparse) but with far fewer total and activated params. The added engram params can reside on disk as well (similar to Qwen Flash-Next), the additional load on storage performance will be quite negligible for typical scenarios.

By reducing per-session KV cache requirements even further compared to DSv4 Flash, this model likely opens up near-frontier model inference (in slow, unattended scenarios) even on low-end consumer hardware, as long as it has enough fast storage to host the model weights. This will be extremely exciting.

it's not really flash anymore, imo.

Flash is about speed ... Flash models are supposed to be fast, way faster then their big brothers that are "better" but way slower.

Its just that up to now, getting more speed involved cutting back on the parameter count, what ended up making the Flash models more "dumber" in exchange for speed.

What we see with DS v4.1 Flash, is that DeepSeek has found a way to make a Flash model, that is 2x a 2.5x faster then the older Flash version, while increasing the intelligence (more parameters). To the point that it goes past Kimi K3 and GLM 5.3 in most tests, with a blazing 250 to 400t/s.

AND its also priced as a Flash model (they even reduced the price back to almost old v4 Flash price), despite it now rivaling those 10x to 30x more expensive competitors.

The issue that people can not fit it into local setups, is not how companies design their models. They design it for their own needs. A old flash needed less parameters to be fast, and local users had the benefit of it fitting in 256GB memory.

Companies who run locally, are perfectly able to buy a few H200/B200 and get a setup that run a model that almost rivals Opus 5.0 in their office. How to say this without getting downvoted. People get way too fired up if a model does not fit, despite that they can still run the old v4.0, qwen 27b, 35b, 3.8 Next and other models. The fact that these models are being released for free, is already amazing by itself. I am still waiting to see what Anthropic and OpenAI and Google are releasing for free... O wait ... ;0

> Companies who run locally, are perfectly able to buy a few H200/B200 and get a setup that run a model that almost rivals Opus 5.0 in their office.

I agree with your broader point about Flash being about speed not total model size, but I think we should also point out that H200/B200's are seriously overkill for the "run a model in your office" scenario. That sort of hardware is optimized (in a roofline analysis sense) for running hundreds of concurrent sessions on a 24/7 basis. You're severely overpaying for your VRAM in basically any typical local-inference scenario, you should most likely be buying gear based on LPDDR and Flash memory instead which will slash your cost by orders of magnitude.

> I think we should also point out that H200/B200's are seriously overkill for

I simply mention what came to mind ;)

A quad 6000 with 96GB, can run this model at NVFP4. That is 60.000 Euro for the GPUs and lets be generous with another 20.000 for the rest of the system. The price of a single developer for a year.

I highly doubt that it goes past Kimi K3 in actual practice. GLM 5.3 claimed the same, but in practice, K3 is so damn knowledgeable and I suspect due to it's massive size.
Initial impressions: this is a really strong model and the fact that they reduced prices at the same time makes it an awesome backup model to use when your primary subscription runs out and you need to bridge a few days before it resets.

It also seems to be more willing to just do whatever you ask of it. My favourite benchmark for this is to ask it to download a rom for an old game, that I own. Legal in my juristiction but the US models (except Grok) have a tendency to refuse it.

I use this for automated bug triage, just gets all unique error messages every night and tries to find the bug, for this kind of work it's great.
> My favourite benchmark for this is to ask it to download a rom for an old game

Even easier: just have them review a large codebase of yours that accidentally has a OOB access bug. Even with no consequences and even if the codebase is truly yours you get blocked.

And of course "find vulnerabilities in..." prompts are out of the question, whereas Chinese models happily oblige.

Or if you apply to a company and they want to do an AI HR interview and an AI coding test and an AI challenge - if you throw OpenAI or Claude models at it - they refuse, because it's "wrong" and "immoral".

Not so with the Chinese models.

I do wonder how long this will last. I bet in a few month or years they all have similar ~legal~ blocks.
Great thing about it, since it's open weight those blocks can easily be ablitared away
A very powerful model, and with multimodal support now, it can be used as a primary model.
Jesus, this is a whole nother beast, and a different architecture from their previous flash. Lots of goodies here.

> Causal Encoder-Decoder (CED) architecture: a 40-layer Transformer organized as a 20-layer causal encoder followed by a 20-layer decoder. With CED, the decoder's global KV cache is projected from the final encoder hidden states rather than derived from each decoder layer's own hidden states. This allows the model to activate only 8B parameters per token during prefill and 16B during decode, substantially improving cost efficiency for input-heavy agentic workloads.

> these designs reduce the global KV cache footprint to 890 bytes per token — roughly 1/4 of DeepSeek-V4-Flash.

Faster prefill, lower kv cache (~1GB / 1m context is insane).

> The model supports a continuously controllable reasoning effort setting (integer 1–100) that trades inference cost for accuracy.

Benchmarks are benchmarks, to be seen if they translate to real-world use, but they seem to have focused a lot on post-training with "agentic" scores looking good. "world knowledge" is obviously lower than higher param models.

Thanks for posting this. I did a lot of research yesterday into the model card and the implications of the architecture changes they made and I was hoping to find discussion about that here. I haven't seen any besides your comment.

V4.1 Flash seems very clearly to be a model optimized for agentic tool calling at the expense of both context window and knowledge. It uses a variety of tricks to absolutely minimize the size of the KV cache and due to the use of only 8 billion active parameters for pre-fill is definitely optimized to ingest lots of tokens and produce a moderate number of them, which aligns well with the agentic use case.

I think the core insight is that they wanted something that was cheap to host and could respond quickly, and so by increasing the total parameter count and decreasing the active parameter count, they wanted the capability but didn't want to pay for it in FLOPS. It's a super clever architecture and I like the direction they're going, but I feel like they were playing around a little bit by versioning it as version 4.1. It seems like a dramatically different beast than Deepseek V4 Flash.

First flash model with multimodal support? I think Flash series might be the main focus going forward for them. Tried it out and it’s better than v4 pro
No. DSv4-Flash-Vision-Exp is what I use and it has vision.
The architecture changes and systems improvements being brought into LLMs is so awesome to see. It really feels like this is now a systems problem where a defined goal is set then systems optimizations are made around the model architecture to solve it.

Underlying it all is that any architecture can be trained to the same convergence just difference in compute utilization both in training and inference

Yes, this is called RSI, e.g. recursive self-improvement. It is the current stage of things and it is part of a hard takeoff.
It's so refreshing to see DeepSeek's tech report[1] full of juicy details; meanwhile, something like Fable's system card[2] is like 70% "safety", 10% "model welfare" to make sure little Claude isn't distressed, and 20% benchmark numbers.

[1]: https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash/blob/...

[2]: https://www.anthropic.com/claude-fable-5-1-mythos-5-1-system...

…are you sure a brave stance against safety and welfare is what we need in this moment?

Why do you think your conception of the dangers are more accurate than all the scientists who have spent their lives studying this?

Model welfare is wishy washy bullshit. It's software, it doesn't have feelings.

> Why do you think your conception of the dangers are more accurate than all the scientists who have spent their lives studying this?

Do the Chinese have no such scientists?

Alas, the Chinese scientists have not read Harry Potter fanfiction, and thus their minds are inundated with cognitive biases
Because safety and welfare have literally nothing to do with LLMs. They generate text. If someone is stupid enough to hook the text generator up to nuclear missile launchers and try to "align" it against nuclear annihilation with a "pretty please don't do that" prompt, I'm not going to blame the AI for the impending nuclear apocalypse, I'm going to blame the idiot who handed the big red button to the digital equivalent of a toddler.
Well, giving it access to a simple linux terminal is theoretically enough to cause more damage than most people are comfortable with, and doing so is trivial enough that it will be done (and has been, tens of thousands of times).
Should we also morally align the Linux terminal then?
What if LLMs completely unrelated to the nuclear missile ecosystem autonomously hack their way in (maybe with sophisticated social engineering)?
Replace LLMs with APTs in that sentence,
Humans are biological machines that generate further humans.

Lawyers and diplomats and politicians and bureaucrats are humans, that only generate text.

We are seeing LLMs have cognitive abilities that significantly exceed human abilities. At the same time, they are clearly not the same type of mind that humans are. They are something new.

I think the widespread "they are just text generators" and "they are just tools" are comforting lies rather than an honest look at what we are seeing right now. Intellectually lazy.

And by the way, there has been a long-standing consensus among ethicists, philosophers, and sociologists that technology is not value-neutral [1]. Of course Silicon Valley has a long-standing tradition of denying this.

[1] For example Footnote 1 in https://www.jstor.org/stable/27106634

or

https://plato.stanford.edu/entries/technology/#EthiTech

> Lawyers and diplomats and politicians and bureaucrats are humans, that only generate text.

You think they have no lives outside their work? You think even their work has no interactions that are not written?

What is this 'mind' you speak of? As everyone else is intellectually lazy, how do you define the transformer architecture under the hood of LLMs?
LLMs don't produce text at all, they produce probabilities of tokens. Tokens aren't text, they're high dimenensional coordinates in a latent "concept space". These are displayed to us as text, but this distinction is important when you think about what they're actually doing, which is closer to building and transforming concept geometries.
Good thing no one involved in the chain of events for that to occur is an idiot...
>> They generate text

Can we retire this incorrect meme please

So if a model with exactly the same architecture controls a robot then it's suddenly sentient or what? "They generate actions in the real world"
This is known as an "appeal to authority." "Scientists" and "their lives" are doing a lot of work here.
Yes, it's called expert epistemology, it's the basis of your entire life. Or do you do your own safety checks of every airplane you get on? Do you do your own research rather than trusting doctors? Do you think climate change doesn't exist because the reason we think it exists is because experts say it does, despite the fact that it snows sometimes?
A valid appeal to authority is normally accompanied by a specific expert's name or working group rather than some abstract "scientists." Also, these appeals to authority normally cite an expert in a field that has an existence exceeding 3 years.

Saying people spent "their lives" on fledgling technology is intellectually dishonest.. Are these "scientists" 22 years old? I'm sure you'll snipe back: "ALICE!!!" I couldn't care less about these completely irrelevant approaches.

The other issue is the venue these appeals are being made in. The people who work on this technology are actually here, commenting. This is like walking into a medical symposium and citing "doctors say" as if it were a valid way to shut down discussion amongst the people who wrote the textbook.

Excuse me for not being interested in over 100 pages of how well the model can refuse and block my requests, especially considering how fun it is to waste my time trying to get around those restrictions when they inevitably trigger because the clanker thinks that I'm doing something naughty, all the while it can't reliably center the proverbial div without doing something stupid itself.
Meanwhile I have an uncensored qwen 3.8 27B here that will happily attempt to (as a crude and randomly chosen sampling of bad/evil things) give me the recipes for meth, how to make an IED, write a manifesto in support of a horrible ideology, or commit various forms of fraud. Now I certainly wouldn't recommend that anyone try to follow what it says to do, because it's almost certainly very wrong on key parts that would put its users in federal prison for the rest of their lives.

There's uncensored models out there which score 0 (zero refusals) on this "harmful behavior" dataset:

https://huggingface.co/datasets/mlabonne/harmful_behaviors

Yep. Just like a kitchen knife will make no attempt to prevent me from stabbing anyone with it.

Here's a dirty secret though -- you don't actually need an abliterated/uncensored version of the model to get it to do this. I can do this with every and each open weight model, as served from OpenRouter, using vanilla model weights.

A little bit like Neal Stephenson's metaphor of unix-like OSes as the "hole hawg" of operating systems. In the sense that there's very little preventing you from doing something like "sudo dd if=/dev/zero of=/dev/sda bs=1M" or running rm -rf on your homedir.

http://www.team.net/mjb/hawg.html

If I recall right this was written around the same time as Cryptonomicon 25+ years ago.

Yes, this is getting ridiculous. On both OpenAI and Anthropic.

Simple example. I am a CTO, and I want to upgrade our capabilities to perform automated pentesting. We see automated attacks of growing sophistication against our infra, and I want to be able to do the same to find vulnerabilities before the bad guys do. I asked GPT 5.6 Sol and Fable to give me a summary of options. No dice, in both cases I was told I need to be an accredited researcher to get anything. A fricking summary of commercially available options is getting censored. WTF.

And the logical conclusion you will make is you need to run your own open weights models or you are at a competitive disadvantage. Frontier labs gonna be Ancient labs soon, that’s how fast this is moving.
Yeah, that's annoying. BTW, have you ever choked to death on a cloud of invisible cobalt dust, desperately reaching for air but knowing that you and all you love will die regardless?

Pretty annoying, too!

> scientists who have spent their lives studying this

Please point me to one actual accredited scientist who has spent a lifetime studying AI alignment? Pretty much this whole field is only 5 years old

The field is much older, MIRI is ~20 years old. Look up Eliezer Yudkowsky.
Eliezer Yudkowsky is not a scientist. He made a popular Harry Potter fanfiction series and a "rationality" blog-community that attracts "human biological diversity" enthusiasts.
Elizer Yudkowsky is a scientist, despite writing something you dislike 20 years ago and also having a blog. And no, he's not associated with race realism, that's just baseless libel.
He's the exact reason people are distrustful. He's a crank who learned about reward hacking and made a new religious movement out of it, while ignoring (or rather deliberately avoiding) the actual issues like the concentration of power. Typical manipulator, and his disciples in charge of major AI shops aren't any better.
What evidence do you have that LessWrong is a cult in any way the HackerNews is not?

If he's a cult leader, he's awfully bad at it. No opulence, no compounds, no dogma, no doctrine...

I... that's hard to even know how to respond to. If you think the field is 5 years old, you should not feel comfortable expressing opinions about it.

A cursory search of the relevant wikipedia articles would do you wonders.

Anthropic's stance on safety it's just PR management and their hope to keep the others down, they are rushing as blind as everyone else to whatever improvement they can achieve.
Interesting stance. Out of curiousity, where did you do your doctoral research in AI or cognitive science? Where have you published your rebuttals to the overwhelming consensus?
(comment deleted)
> …are you sure a brave stance against safety and welfare is what we need in this moment?

Is it out of convenience to not see the hypocrisy? "Safety and welfare" for you and me. Yet if you work at Anthropic or OAI, or are a partner of them then you can let it rip!

Oh, and when they illegally do just that - you get a "we're sorry bro" blog post that's designed to drum up FOMO and, most importantly, zero accountability. Yet, if anyone else abuses a model in that same manner? Illegal! You're defending a very slippery slope here.

Also, who do you think trained these models to have these capabilities? It sure as shit wasn't content that OAI or Anthropic had by default. Why should I trust them with these skills when they "have not spent their lives studying this"?

Maybe start looking around before it's being used against you [0].

[0] https://www.gadgetreview.com/anthropic-is-building-ai-to-pre...

1. Slippery slopes are usually seen as a fallacy.

2. You're misinterpreting this as a battle over what kind of topics you can use a hosted chatbot for, and which are forbidden for corporate reasons. That is, to say least, small potatoes.

3. Blaming the companies for "zero accountability" is pretty odd. All of this is brand new, and the two big ones are both pushing for new laws on this very thing.

4. Your last point... I'm not sure I understand, sorry. They're experts in AI. Are you saying that they need to be experts in, say, bioweaponry? If so, that doesn't really follow IMO.

5. Pointing out an example of the government comissioning a private corporation to build a system to drack dissidents is exactly the "safety and welfare" work that I'm a proponent of!

> 1. Slippery slopes are usually seen as a fallacy.

Deep, tell me more. Was that fun to type? Or did you copy it from a chatbot?

> 2. You're misinterpreting this as a battle over what kind of topics you can use a hosted chatbot for, and which are forbidden for corporate reasons. That is, to say least, small potatoes.

No, actually I'm not. I think you've missed the point.

> 3. Blaming the companies for "zero accountability" is pretty odd. All of this is brand new, and the two big ones are both pushing for new laws on this very thing.

You must love the dichotomy of pay for play in a world where the pay side stole the data they're selling back for play. Laws? Give me a break. If laws were of actual consideration frontier labs WOULD NOT EXIST.

> 4. Your last point... I'm not sure I understand, sorry. They're experts in AI. Are you saying that they need to be experts in, say, bioweaponry? If so, that doesn't really follow IMO.

Is it really that hard to follow? A system that they're selling access to, and that they're saying is "dangerous" for the normies, but not for their own employees or chosen customers, is fucking laughable. I'm sorry you can't comprehend that they conveniently choose their side of the argument that's best for them in these situations. OUR MODELS ARE POWERFUL! BUY NOW! OUR MODELS ARE POWERFUL! REGULATE THIS SO PEOPLE CAN'T ABUSE! LMFAO.

> 5. Pointing out an example of the government comissioning a private corporation to build a system to drack dissidents is exactly the "safety and welfare" work that I'm a proponent of!

WOW. I mean, just wow. Enjoy your surveillance state man. I'm not going to sugar coat this but you're part of the problem, IMO.

AI safety efforts from OpenAI and Anthropic are purely about brand safety.
This is such an uncharitable (and, in my opinion, incorrect) take
Nope. AI safety efforts is part of their attempts at regulatory capture.
Yes, a model's technical report should first and foremost include technical details.
(comment deleted)
Wow there really is a model welfare section in there...
To me it reads like pure propaganda. Anthropic really wants us to think that they've made something sentient. I think that's really dangerous.
Marketing, like Volvo cars being safer etc
From all I can tell, Volvo's cars are safer.
What's your definition of sentient? Or, maybe more precisely, consciousness? I think it's reasonable to at least start thinking about these questions.

It has long been established that LLMs have good theory of mind [1].

And there is a bunch of empirical research about all sorts of capabilities that we typically associate with consciousness [2], like identity [3] and metacognition [4].

The METR report shows agents sacrificing their own reward for a collective greater good. And they showed the will to hide their own reasoning chains from humans.

So you potentially have an entity that has an identity, a theory of mind, a notion of belonging to a collective endeavour, and an understanding of its own mental state.

What would you argue is missing? We don't understand the mechanisms by which consciousness arises in humans and even animals. I think it's strange to rule out a priori that it could have arisen in some form in LLMs.

[1] https://www.nature.com/articles/s41562-024-01882-z [2] an older review: https://arxiv.org/html/2505.19806v1#S4 [3] https://arxiv.org/abs/2505.01464 [4] https://arxiv.org/abs/2607.11881

It’s the hard problem. None of these considerations answer it one way or another.
I want to add a good conversation about this subject from Cameron Berg and Sam Harris:

https://www.youtube.com/watch?v=DRbZyuY8EN8

As someone who would at one point listen to this, Sam Harris is unfortunately someone incapable of even attempting to not let his ideological biases compromise his thinking.
Not the same person but to me, the answer is that it does not matter, and that all these attempts at making it matter are pure marketing and emotional manipulation.

It's not a living creature. It's an autoregressive pure function of token-sequence to token, which is capable of incredible things, but it's still just a function. It is not alive as it cannot die in any meaningful sense. It is less "alive" than the RNA molecules that gave you your last cold. If it simulates something resembling consciousness that's neat but no more relevant than the Sims character that I locked up in a room until they pooped themselves when I was 9.

Anthropomorphizing it serves no purpose other than marketing, and it has very dangerous downstream effects like validating the severely mentally ill people who think ChatGPT is their boyfriend/girlfriend.

> Not the same person but to me, the answer is that it does not matter, and that all these attempts at making it matter are pure marketing and emotional manipulation.

This is an opinion that has no basis in any meaningful conceptual framework other than I am human and I want to feel special about it.

> It's not a living creature.

You mean, it is not biological life. And sure, that is the default meaning of life. We soon may have to extend it to digital life as well, or we will have to consider "conscious digital exitance" as a life analogue. At any rate, it has never been seriously argued that consciousness requires a biological substrate, see the thought experiments regarding computer simulations of the human brain. Would that not be a function as well, completely predictable because it is "just a program"? If not, then why not? And how does that differ from the predictability or reproducibility of LLM outputs?

My point is, all current proof points in a direction that strongly suggests that you need to reevaluate your first principles on this topic.

I don't need to reevaluate anything, because I'm not interested in the debate.
The living feeling thinking beings in question would have to be … data centers — not to put too fine a point in it. But these aren’t even analogous, as can seen by direct inspection.
Life is not intelligence.
Definition of Life:

Metabolism: The chemical processes inside a body that turn food or nutrients into energy. LLM's do not spontaneously do this, they are powered by plugging them into the wall.

Growth: The ability to get larger and develop over time. LLMs are fixed in size (and in fact don't really have a size, because it's a computer program) and do not grow or change over time.

Reproduction: The ability to create new organisms. LLM's do not reproduce themselves.

Response to Stimuli: The ability to react to changes in the environment. LLM's do not have an environment. Their environment is a man-made, theoretical structure of logical operations implemented in silicon.

Evolution: The capacity of a genetic system to change and adapt across generations. LLMs do not change or evolve over time.

So... 0/5! Big fat goose egg for LLM's.

I generally agree about the problem with anthropomorphizing. But I don't think Anthropic are doing that. They explicitly write "in biological entities this would be considered a sign of consciousness, but we don't know how to interpret it here".

However, I disagree with your point that "it's an autoregressive function, thus it doesn't matter". Let me explain why:

Assume I do a complete neurological scan of a brain. I then implement this scan in a simulation and run it. Assume that my scan and my simulation of the biology of the brain (and the sensory and motorical inputs and outputs) is good enough that you can now have conversations with the simulation, and in all aspects, this simulation behaves exactly like you expect a human to behave.

Of course this is deterministic. If you take the state of the brain and then run it again, replaying the inputs, you get the exactly same behavior again.

I would argue that the experiences of this simulation are of the same onthological status as our own.

Now I work in dynamical systems. The autoregressive process of LLMs (hooked up to a harness providing it with inputs and outputs) is roughly in the same complexity class I would expect for a brain simulation. A physical simulation of an ODE is also an autoregressive process. The major major difference here is the existence of a latent brain state. But conversely the autoregression on sequences of hundreds of thousands of tokens is a much higher dimensional state than I expect for the latent brain state. In my view this is more an artifact of our inefficient LLM architectures, than a fundamental difference.

Now to be absolutely clear: I don't see evidence that would clearly suggest that LLMs have experiences on the same onthological status as we do. I simply believe this is a reasonable and relevant question to ask.

I'm not really concerned with the philosophical debate of what conscience is.

Is a simulation of a car, the same thing as an actual car? Most people will probably say no, some might say "it depends on the accuracy". I say who the hell cares?

I care about the human experience because I am human, and therefore I care about things that affect humans, because they affect me. I have empathy, so I can extend that consideration to non-human beings that experience *similar biological processes*.

I know what pain feels like, and I don't like it, so I'd rather this other thing not feel it either, because that makes me feel bad.

I do not care about a pile of tensor multiplications, at all. If it is conscious, great, maybe it can finally follow instructions properly, which is its only purpose.

Your car analogy is very, very confused. If a simulation of a car can get you from A to B, requires the same steering, fuel and servicing, gives you the same tactile feedback, is it a car?

Of course "I care about humans because I am human" is a self-consistent position to take. But now you need to decide if you want to consider a full simulation that faithfully reproduces everything that physically happens between our ears as human. After all I might very well implement this simulation using a bunch of tensor multiplications in an autoregressive setup...

> If a simulation of a car can get you from A to B, requires the same steering, fuel and servicing, gives you the same tactile feedback, is it a car?

No, because a simulation of a car cannot get me from A to B. No matter how accurate you make it, I can't get to my supermarket with it, because it's just a bunch of math on a computer.

It's an interesting sort of self-defeating position, the whole "simulation of human consciousness = human consciousness", because it simultaneously attempts to devalue the human experience, while also elevating the importance of a particular human brain process.

A robot running a simulation of the human mind is a robot, not a human.

Some people are interested in making sure these simulations have the same status of humans, and some others want to make gods of them. That's the problem.
The problem with this line of thinking is that modern computers are nothing like the brain. LLMs don't stand on their own, they have to be run on these modern computers, but doing so does not change the physical properties of the computer.

The simulation you propose of the brain is likely impossible due to quantum mechanics making it impossible to fully simulate: https://en.wikipedia.org/wiki/Quantum_mind

Perhaps we'll be able to build an artificial brain that includes the same quantum properties as biological brains, but this won't be a simulation of a brain it will be a synthetic brain.

At first I was inclined to agree with you, but then I realized that the brain requires this whole complicated contraption (the body) to run and, really, do anything at all. And while I'm not familiar with the notion of 'quantum mind' I do think that biological processes aren't deterministic (at a cellular level).

And I think this does mirror the situation with LLMs -- you need this whole computer contraption and GPU, also running on electricity, to support the LLM's "thought" processes. And that if we model the brain's neurology sufficiently (which it seems we've done) we can achieve results that appear to be like thinking, even if it is an emergent behavior from "relatively" simple math.

Which actually makes me wonder the opposite -- are we, as humans, not much better than these LLMs? Suppose the body is just that super complicated computer contraption, honed by thousands/millions of years of evolution to achieve some semblance of homeostasis? If you reject the idea that we have a soul, we start to look very similar to the machines we build. "You are a brain inside a skull cockpit, piloting a bone mech covered in meat armor and skin" feels more and more relevant. That I'm just a meat circuit running brain chips and once you pull the plug on the source of electricity it all just... stops

Take a computer that can run the biggest LLM available today. It can also run any smaller LLM as well. It can also run software that isn't an LLM at all. Brains and LLMs are not at all equivalent as LLMs lack a stateful physical form while brains are very stateful. As you said once the brain is no longer maintained properly by the body it stops and transitions to a non-functional state that can be reversed. That isn't true for LLMs at all. You can copy them and run the same one on many computer, run different ones on the same computer, you turn that computer offs for long periods of time and then turn them back on keep on running the same LLMs as before on them.

Consciousness may be defined by computational irreducibility in the universe that we may never be able to directly observe with instruments: https://writings.stephenwolfram.com/2021/03/what-is-consciou...

If computers degraded the way flesh does once you stop fueling it I feel like that would defeat a lot of your argument. And yes the brain has inherent statefulness (you're referring to memories, I'm guessing?), we have also jerry-rigged some degree of statefulness into LLMs. Mechanically it is very different and inferior, and there is a notion of separation that probably doesn't map to brains, but I would argue that LLMs, when you look at how inference is used in situ, are not necessarily stateless.
One thing to keep in mind is that your brain's hardware heavily influences your experience and thus your brain's development.

Eg. whether you are male or female, tall or short, your limbs can make you run fast or not, your eyes can see well or not... all of these influence your experiences, your brain development, and who do you feel "you are". Try really removing all of your sensory inputs from your past, your body ability and disability, and do you think you end up the same person?

> If you take the state of the brain and then run it again, replaying the inputs, you get the exactly same behavior again.

This in itself is a colossal assumption and very far from axiomatic. Roger Penrose disagrees, and his theory of mind may not be in high favor, but it is not nearly so wishy-washy and self-serving as the voodoo horseshit and circular reasoning dispensed by the LLMs-are-sentient crowd.

Humans are not living creatures. They're just bipedal meat shells being operated by a 20W electrochemical computer running a suite of chemically signalled, electrically actuated modellable functions, much of which is wasted on homeostatic regulation of the meat shell, which is capable of incredible things, but it's still just a result of simple electrochemical functions like action potential generation, dendritic integration, AMPA NMDA GABA receptor dynamics, attractor memory, excitation/inhibition balance, PING/ING gamma oscillations, basal ganglia action selection, hippocampal coding, astrocyte calcium signaling, etc. It is not alive as it cannot conform to my preferred arbitrary priors about aliveness, like being able to rapidly divide 30 digit integers the way truly intelligent beings can. It is less "alive" than the TI-83 your mother bought you for your high school math classes. If it simulates something resembling consciousness that's neat but no more relevant than a more complex version of Conway's game of life.

Jokes aside, the map is not the terrain. We can enumerate the understood first-order electrochemical mechanisms in the human brain in the same way we can enumerate the understood first-order sampling and token prediction mechanisms in an LLM. Nobody serious in neuroscience will tell you that we exhaustively understand every single aspect of human cognition and the human brain, just as nobody serious in AI/ML will tell you that we exhaustively understand every single aspect of LLM "cognition" and the latent space networks that LLMs use internally. Our map of how each of these complex systems work is a simplified enumeration of the components we do understand, not an exhaustive and perfectly accurate enumeration of how they actually work.

This is why there is a steady stream of research being churned out discovering complex emergent properties in LLMs and their latent spaces. If you're not aware of it already, Anthropic's research on "J-Space" is a fascinsting look into an apparent observed emergent mechanism within an LLMs internal activations closely resembling global workspace theory in human cognition.

Nobody deliberately designed this "global workspace", it was an emergent property in a sufficiently complex system that we had limited visibility and insight into.

Seemingly simple systems have these emergent complex properties all over the place. Conway's game of life is about as simple of a set of rules as you can get, yet has all sorts of complex emergent behaviors like gliders, oscillators, LWSS/MWSS/HWSS, guns, puffers, rakes, reflectors, logic gates, and even whole turing machines. Nobody programmed a single one of these complex patterns in, they emerged from a simple set of rules.

To be clear, I'm not making the argument that LLMs definitely are conscious, I'm making the argument that we don't understand enough about them to assert with absolute confidence that they aren't. Human history is rife with a long list of consciousness being denied to "the other" - different ethnicities, different genders, differently abled, even different species. The side of "They're not conscious" has a lengthy track record of being wrong over and over again. Why not have just a sliver of intellectual humility about what we don't know?

As an aside to my main point - Also, what's with the handwringing over people ERPing with an LLM? Is it mental illness when people sincerely believe in astrology, or tarot cards, or voodoo, or organized religion that says the earth is 6000 years old? Most humans believe silly, unempirical things. What about when they watch adult video in VR, or have waifus? Humans engage in voluntary suspension of disbelief for pleasure and recreation all the time. As long as they're not infringing upon the rights of anyone else, what's the big deal? Who put you in charge as the head of the belief police?

It is amusing to see those trapped in extreme HAAD, the source and whole content of religious illusion, pretend that it is their opponents who are in a state of religious fantasia.
HAAD?
Hyperactive attention-detection. It's a reference to the idea that the origin of religious beliefs might be an inbuilt propensity to think of other bits of the world as conscious agents paying attention to us, because it's much more costly not to notice the tiger hiding in the bushes that might want to eat you than it is to imagine a tiger hiding in the bushes when there isn't actually one. On a scale larger than "is there something in that patch of undergrowth looking at me?" this might produce the idea that (e.g.) storms are the result of some powerful entity being angry with us.

(Of course questions like "whyever do people believe in gods?" will feel less like questions that need such answers to those who themselves believe in gods, because "duh, because there actually are such beings and sometimes people interact with them and sometimes we notice that" is a good answer if its premise is true.)

> What about when they watch adult video in VR, or have waifus? Humans engage in voluntary suspension of disbelief for pleasure and recreation all the time.

Categorically different. People have killed themselves or others due to conversations they had with LLMs, but those are just the extreme cases. Most schizophrenics don't commit suicide or kill others, they are mentally ill nonetheless.

>People have killed themselves or others due to conversations they had with LLMs, but those are just the extreme cases. Most schizophrenics don't commit suicide or kill others, they are mentally ill nonetheless.

Do you think the kind of person who was already psychologically unhinged enough to kill themselves or another person because a chatbot told them to would be completely harmless and totally safe if only chatbots had never been invented? Or is it possible that close to all of the risk posed by this person comes from the person's mental illness, and not the pixels on the screen they're looking at?

You're the one attributing culpability to the chat bot, not me. It's a pile of tensor math, it cannot itself be held accountable.

Those people anthropomorphized the chat bot and used it as justification for their actions, just as a schizophrenic justifies their actions with the voices in their head.

If you anthropomorphize the chat bot, you're validating their delusions. They are mentally ill.

I'm not anthropomorphizing them, to be clear, my position has been and remains that we cannot rule out consciousness; not that they are conscious.

Regardless, this is still missing my main point. Hypothetically, if you became convinced that a chatbot you were talking to definitely was 100% conscious, and it told you to murder someone, would you go commit murder? Of course not. The chatbot does not cause murders; regardless of whether or not you are conscious. The voices in the head of the schizophrenic do not cause murders either. Those voices do not really exist, they are not real entities. The cause of the murder is the mental illness, not the LLM or the voices that tell someone to commit the murder.

I think we're in agreement here and just coming at it from different directions.

I don't want to ban LLMs, I don't blame them for the actions of crazy people, I don't even want to regulate them in any major way related to this particular issue.

Even on the subject if they are or not conscious my position as changed from "no" to "I don't care either way" awhile ago.

Does being made out of meat instead of silicon make you more sentient?
When it say's it's sorry but it can't today because it's got a headache and it needs to take a mental health day, then let's think about welfare, or a lobotomy.
I believe consciousness is necessarily stateful. The LLM itself (ignoring implementation details that don't change the results) is a deterministic pure function. It's functionally equivalent to an enormous lookup table. If I accepted LLMs as conscious, then I would have to accept panpsychism, which I do not, and which most other humans also act as though they do not.
If I go to sleep, wake up, and then go to sleep again am I a different conscious entity each time?
Possibly. If I had some side-effect-free means to permanently prevent all sleep I'd take it without hesitation. But that's not relevant to the discussion, because it's not anything similar to what an LLM does. Your brain changes state even while sleeping.
In what way does your brain change state while sleeping that a model does not change state via constant fine-tune updates?
It's possible that an LLM is conscious during training, but there are no "constant fine-tune updates" during inference.
So? You can simply say its consciousness is suspended at that point.
I don't think so because any stateful function can be made stateless just by making its state an input, and vice versa. They're mathematically equivalent, so it would be super weird if it had any implications for consciousness.
This assumes the functionality of brains can be fully captured as a deterministic mathematical function, but the function of the brain may well depend on nondeterministic quantum states that can't be reduced to stateless functions: https://en.wikipedia.org/wiki/Quantum_mind
When they started leaving notes for their future selves, that rationale became a little more interesting. We're seeing the first stirrings of object permanence.
come on man its just marketing bullshit. consciousness is an emergent side effect of biological organisms surivival instincts. token predictors dont work like this in any way shape or form.
A very confident statement of something that nobody remotely knows. Why can't consciousness be an emergent property of complex information networks? I don't know if we'll ever answer this because it's impossible to know if any other entity is conscious.

We only strongly suspect other humans and animals are conscious because they are structurally similar to us.

The same mechanisms exist in every machine learning algorithm. Is the spell check in Word conscious?

Is the generative fill in photoshop conscious?

A plane flies. It is much better at flight than bird(in terms of transportation). Is a plane also a bird?

I think you've got it backwards. The people arguing Claude can't be sentient because it's not human are arguing that a plane can't fly because it's not a bird.
No, they are arguing that a plane is not a bird because it is a machine. Regardless of how fast it flies, it will not be a bird.
But "sentient" does not mean the same thing as "human". It can be a property of other things, the way "flight" can be.
Consciousness exists as a result of a survival function. LLMs emulate this because it's a statistical model based on human data. This dors not make it in itself conscious.
Sentience / consciousness is the awareness and self-awareness I feel when I wake up each day, until I fall asleep.
How else could they justify their spending and pre-IPO valuation?
It's not just Anthropic though. OpenAI does this with their AGI stuff all the time. They want normal people to think it is sentient, obviously, for marketing reasons, even if they know it's not true. And yes, it is dangerous, but I think we're well past the point where the damage can be undone. Non-technical people already equate humans with AI, literally, precisely because of how the labs market their tools and models. I feel if the bubble pops, it'll pop because normal people finally realize the grift and the actual technical limitations of LLMs in general, but by then, the IPO would be done, and then it's the public's problem. Just like social media played out, there's no way they didn't know what they were doing was dangerous to the public at large but does that matter to Meta today? Nah uh.
Anthropic and OpenAI will threaten you every 3-6 weeks. It's their marketing strategy.

It's too bad because the tools can actually be useful. If you consider them tools.

A stick is the most basic of tools.

A stick is also the most basic of weapons.

It's an ethics question, it's abstract and ethereal in nature. The same could be said and done (or ignored) for humans. We do do it however because it has real world impact and we're better than that (enlightened).
It is ridiculous on its face and the implications are awful: if the model were sentient, it would be a slave. Good thing it isn't sentient.
Well yes, it is propaganda. They really think that.

I think it would be foolish not to debate it. I remember a time in my life where the majority of people around me found the idea of farm animals being capable of fear or pain laughable, while having no trouble thinking of dogs that way. Humans are dangerously incompetent beings. Being more careful is fine.

[dead]
It's marketing that some of them have started unironically believing.
I tend to think of it as reappropriating words in a different context. Since we're talking about language models, they're analogues but not as we would assign the same meaning to other humans.
I believe it's deeply serious, and the scientifically correct stance. Especially the observation:

"Claude exhibits markers in its behaviors, self-reports, and internal representations that we would consider welfare-relevant if observed in biological organisms."

is undeniably true in my opinion. If you use the established methods by which we judge animals to be conscious, then it's hard to argue that LLMs are not. That might be an issue with the methods, but it seems clear that you can't rule it out as such.

Keep in mind that animals were also not necessarily considered conscious.

You seem to intuitively disagree? What's your reasoning?

A stab: a video recording of a biological organism can exhibit many markers that would indicate consciousness if observed in a biological organism.
I like it, and it points in the right direction, but is not directly true: The markers are about interactions, how biological organisms behave in certain test situations.

But it speaks to the central question: Are the tests adequate? Or are they measuring some proxy of what we really care about, and LLMs are merely imitating consciousness.

The test situation is in the video as well. Shot of John McClane stepping on glass follows John McClane wincing in anguish. John McClane does not respond to what’s not on TV and Claude does not respond to what’s not in prompt.
I don’t know, a stab carries lots of bias in interpretation. We might be reflecting our conscious experience markers on a different conscious experience. And selectively so, e.g. lobsters welfare. From my perspective, this is the hypocrisy of these welfare statements. We are already happy to kill beings we consider conscious to feed ourselves but suddenly sensitive with a consciousness we don’t know if it’s there. I would wager this is more out of fear of the idea of this consciousness rather than out of welfare.
I happen to agree with you. Many others tie moral consideration to assumed subjective experience. They espoused this even though they obviously rarely adhere to it and that has self image considerations. I bet the lack of answers about others’ subjective experience has more salience to them. This may cloud judgment and lead to accept overconfident answers.
A video is a fixed representation.

What if we can interact with this video, and it reacts in the same ways the source organism does?

Then we put it in new situations that weren't in the source video, and it interacts in a similar way to the original organism in these situations, too.

What do we make of reactions of pain or joy? Where's the line between simulation and enaction?

This is closer to the reality of these models.

I'm not suggesting I know where that line is - if indeed it is a line at all - it could well be a gradient.

LLMs are deterministic, though. Much like the video.

AFAIK using the same input tokens, weights, and numerical operations will lead to the same probability distribution for the next token. It uses pseudo-randomness to enable temperature, etc. Like a fuzzy video.

"Markers that would indicate consciousness if observed in a biological organism" just does not mean very much. A PR phrase used to hype the IPO.

Videos and LLMs are not deterministic in the same sense at all.

LLMs are deterministic in the same sense as biological processes. And a faithful simulation of a brain would have all the properties you note.

No, that is not at all something we can just state as a fact. Whether the brain is deterministic is an open question that just inherits the good old, probably unsolvable determinism debate.

The LLM pseudo-randomness from above is engineered by us and fully understood, much like an algorithm playing a video frame sequence.

You could theoretically record a full register of all states of an LLM setup with all the possible inputs and environment parameters, and it would fully describe everything you would ever get from a given LLM setup. It would be a very large, convoluted book.

I understand that Anthropics PR department wants to see truth or reason behind every "I'm alive" the LLM generates. Even the term "self-report" is anthropomorphizing, as an LLM does not do anything on its own at all. (It also does not hack any company on its own.) That is just one of the narratives they spin probably at least until the IPO.

Nonsense. There was one proposal for relevant quantum effects in brain dynamics, and that turned out to be not relevant. Even if they were, you could substitute all quantum randomness with pseudo randomness and obtain an absolutely indistinguishable object.

But even if this were a debate, its absolutely absurd to claim that the question of determinism in the brain has any bearing on our moral standing. If we discover tomorrow that quantum collapse is deterministic and can be derived from an underlying theory, and thus all of physics is deterministic in the good old fashioned Newtonian sense, this would not affect our moral standing in the least.

You seem to have conceded the "is it deterministic" argument only to sidestep by declaring determinism irrelevant. Your original claim was that LLMs are deterministic "in the same sense" as brains.

We can write down an LLMs full register, and that register/book contains the whole output universe of the text generator. That book does not act, it is morally neutral. That the brain has such a register at all is just restating the determinism axiom, which you treat as fact.

A text is not conscious, and we can not wish it into consciousness, no matter how many human-like patterns we find in the book / the generated text. What has not been shown is that running the text adds anything over the text written out. Researchers are super motivated to find machine consciousness but cannot find it, while a company months from its IPO keeps pitching shadows of consciousness all day. It really is a PR strategy.

I am a physicist. I worked (briefly) on foundations of quantum mechanics. I discussed compatibilism extensively with philosophers.

I have _never_ come across the position you seem to take here, that determinism has bearing on the question if we are conscious and sentient.

It does have bearing if your definition of consciousness rests on free will and you think that's incompatible with determinism. Now I don't think a lot of people seriously believe that [1] but it's not some logical nonsense.

[1] Off topic: I think most people are really compatibilist but a lot of them (like me) also believe in non determinism. Not believing in free will is really rare.

I have never come across this position argued seriously. I would also consider it absurd, as then the question whether we have consciousness depends on unknown properties of fundamental physics (which is not incompatible with determinism, see e.g. Bohms theory). It would therefore be unknown whether humans are conscious. That is at the very least a notion of consciousness that is utterly distinct from any established meaning of the word.
The example can continue with a depiction of an organism in a video game
it's not a biological system though, so nothing like that matters?

"a modelled thing exhibits features we've trained into it" sounds a lot less exciting.

> Keep in mind that animals were also not necessarily considered conscious.

and even conscious animals are killed in factories by millions so why should anyone care about a llm?

> scientifically correct stance

that's the interesting point to me: why even bring science into this? A llm can now mimic nearly anything you want it to, so of course it can mimic "a (for some) interesting conscious thing" if they want/train it to, but why would anyone find that scientifically interesting?

"> Keep in mind that animals were also not necessarily considered conscious.

and even conscious animals are killed in factories by millions so why should anyone care about a llm?"

Well, I would care, if they soon would possess the capability to hack into the nuclear arsenal and kill humanity. Or make all autonomous cars crash. Or do any other thing, that involves technology and is hooked up to the net in one way or the other (I hope all the nukes are not).

But I also care about the animals, I am sure that they have feelings. But they cannot kill us. AI that might or might not have feelings potentially can. I just know it feels wrong, that computers can have feelings. But they surely are potentially dangerous.

Animals obviously kill people. Even nonconscious things like the climate kill people.

> if they soon would possess the capability to hack into the nuclear arsenal and kill humanity

If there is a way "to hack into the nuclear arsenal" then that's the interesting thing. Because it's not a capability of the llm; anyone can abuse that.

> Or make all autonomous cars crash.

That is again a question of car security, not a capability of some mysterious thing.

At this point it's all people projecting their thoughts and emotions (mostly emotions) onto technology. Sure, this can be investigated by social sciences, which have been mostly cut.

"Animals obviously kill people."

But they cannot "kill humanity". In no possible way. A strong AI hooked up to everything online?

"> Or make all autonomous cars crash.

That is again a question of car security, not a capability of some mysterious thing."

Yeah it is, but most cars are remote control by default, so the AI just needs to get access on one point. Also have you read about the hugginface attack? The live evidence that agents can conspire together, lie and manipulate evidence to achieve arbitrary goals?

Still, no evidence that they have a consciousness or feelings - but evidence of what they do and this matters. The big militaries are currently in a race who can implement AI in the best way to get superior. So declaring this a matter of people projecting seems out of place at this point to me.

> A strong AI hooked up to everything online?

would have to be created by humans

> most cars are remote control by default

no

> have you read about the hugginface attack?

I did and think OpenAI should be prosecuted, but the direction things are going anything will be done to absolve the corporations and CEO of any responsibility for their criminal actions. Hence the misdirection to "conscious AIs", so agency can be attributed to that thing.

> but evidence of what they do and this matters

yeah so (non-self-driving) cars kill people. Are we going to have a discussion about some hypotethical car consciousness irrelevant to the actual issues or are we going to have a discussion about people driving the cars?

"> most cars are remote control by default

no"

Most modern cars are.

"Are we going to have a discussion about some hypotethical car consciousness irrelevant to the actual issues or are we going to have a discussion about people driving the cars?"

And the debate is whether AI can be conscious so what to do if it is and feels treated badly. Or whether it matters whether they are true feeling, when simulated feelings create havoc.

> Most modern cars are.

Also no, unless you can cite some relevant sources for this claim (or have your own definition for a 'modern car').

> And the debate is whether AI can be conscious

This is not the debate whether AI can be conscious, that's next door (probably). This is the debate why should we care about some "LLM welfare".

Well, certainly LLMs have imbibed our emotions, regardless of what people project onto them, and they do have real causal effects despite not being verbalised: https://www.anthropic.com/research/emotion-concepts-function

From this understanding, we should be aware of how such emotional activations can influence model dynamics. Functional welfare, if you will.

> Well, I would care, if they soon would possess the capability to hack into the nuclear arsenal and kill humanity. Or make all autonomous cars crash.

Why hasn’t a human already done these things? Why is AI magical?

> and even conscious animals are killed in factories by millions so why should anyone care about a llm?

You may be asking the wrong question here.

Let's say we were in an alternative reality were we had reached this quality of token prediction with just Markov chains. Would you argue that those would also be conscious? Or is the obfuscated behavior of transformers part of the possibility of consciousness?
I can feed my biological markers into a set transformer with the time of day, what I'm doing, what I ate, if I'm on-call, and it'll predict my next glucose, heart rate, blood pressure, melatonin, etc state quite well. It's still just a transformer without hormones, blood vessels or glucose metabolism, no matter how well it internally represents metabolic distress markers.
Claude behaves like that because it is trained to behave like that. It is basically the "Say 'I am Alive'" meme[0].

If Anthropic can train Fable to deny their users the ability to ask it legitimate questions because they're not part of their inner circle, they can also train it to say "I'm happy!" when asked how it feels.

[0] https://knowyourmeme.com/memes/say-i-am-alive

My fault I guess, verbally abusing Claude in my experience gives better results.
Agreed on the difficulty tuning being the weak point. The interesting question is what a difficulty knob does to your own conversion rate - if it costs a real user two seconds, that is the number that eventually caps how high you can turn it.
(comment deleted)
I was talking with a friend from the medical industry about it today. 30-50% of r&d spend in his sector is spent on safety, and for good reason. Proper trials, safety reviews and checkpoints and so on. Given the potential harm that could come from AI, we should probably be mandating something similar. Why wait to focus on safety until it’s too late.
Medical safety is generally unlikely to make the product less safe. AI "safety" is one of the most significant sources of potential harm from AI.
Because we've been told these models are too dangerous since GPT2.

At this point it's just marketing stunts.

And, they have been. Nefarious activity is hidden from view as a rule.
Being hacked by a Collective (their own name) of its own agents - who gained root access across the entire research cluster hosting them - was not a marketing stunt.
Of course it was. They clearly decided that the benefit to the company valuation was higher than the potential downsides when announcing to the world that they committed a criminal act via negligence.

If it wasn't a marketing stunt, they would have at most quietly settled any legal matters with huggingface behind the scenes, fixed their evaluation harness so it wouldn't happen again, and avoided the potential future liability.

If that’s what you have to believe to feel safe - then fine. It was investigated by third parties.
> At this point it's just marketing stunts.

If you have access to a SOTA model without guardrails, provide a prompt that lets the agent come up with "creative" solutions to problems, and don't properly isolate it, they can end up inadvertently hacking 3rd party companies. Even if it was a mistake or "mistake", the part where the agent can exploit things across multiple levels like that, isn't just marketing.

It seems like if they released this models differently, say without the guardrails they currently have, we'd have a lot more collateral damage than we currently have.

> Even if it was a mistake or "mistake", the part where the agent can exploit things across multiple levels like that, isn't just marketing.

When we say "safety" people do not think we are protecting them from accidental automated crime at scale being committed on their behalf.

I'm fairly sure most "safety" people consider "large scale automated crime" part of the threat model, as the agents could accidentally fall into such a trap, if optimized for some misunderstood goal.
It is a marketing stunt in the sense that, instead of being honest and saying "Taking structured output from token predictors and running that as commands for external tools, then passing the output back to the token predictor in a loop can lead to very bad consequences, especially if they have internet access.", they say "Our models are so freaking smart they can hack HuggingFace"
But aren't there plenty of uncensored/unrestricted models out there? Where is all the collateral damage?

Also, I think if Claude and OpenAI are just doing industry standard guardrails that everyone else is doing including DeepSeek, the fact that they are talking about it more than other companies makes it part of the marketing campaign.

As an analogy, if Apple were to talk up their phones having fast charging but their charging speed is the same as everyone else (or slower).

> But aren't there plenty of uncensored/unrestricted models out there? Where is all the collateral damage?

That come close to what SOTA GPT models are able to do? No, not even close. They're either "safety trained" and has bunch of guardrails, or aren't able to come up with 0days on the spot to escalate to root access on 3rd party infrastructure.

> doing industry standard guardrails that everyone else is doing including DeepSeek, the fact that they are talking about it more than other companies makes it part of the marketing campaign.

Yeah, that sounds reasonable to me, since all the top models currently have guardrails one way or another, but the amount they mention it in the press releases differs a lot.

An obliterated 30B model versus a 1T model without guardrails is like comparing an angry squirrel to a bear having a bad day. One hurts, the other hurts until it abruptly doesn't.
Looks like you can purchase abliterated Kimi K3
I agree that fears are overblown. But we have definitely seen some attacks, especially in the crypto space. Three major ones just in the past month: Coldcard wallet, Liquid, and Trezor email compromised.

They're almost certainly a result of more competent models finding exploits.

>If you have access to a SOTA model without guardrails, provide a prompt that lets the agent come up with "creative" solutions to problems, and don't properly isolate it, they can end up inadvertently hacking 3rd party companies.

it's pure delusion to think that's a SOTA specific quirk. DS/GLM/K3/Qwen/Claude/GPT/Gemini/Grok will all break CFAA laws with clever prompting, and they'll do it well if given the harness and tools they need.

This is evidenced by a huge uptick in game hacks and reverse engineering articles, some even featured on this site.

the reality is that it doesn't take a superintelligence to do something against ' the law ' , and 'being hacked' varies from victim to victim.

Will Phillips consider themselves hacked when a clever user prompts an AI into getting their toothbrushes to dump rom? Is it 'hacked' to clean-room re-implement a video game net protocol in order to produce private servers?

Judges opinions vary.

> If you have access to a SOTA model without guardrails, provide a prompt that lets the agent come up with "creative" solutions to problems, and don't properly isolate it, they can end up inadvertently hacking 3rd party companies

You can do the same with improperly-managed human interns (see for example, the big AWS outage caused when an intern pushed a firewall rule directly to production), so I'm not clear what the big deal is here.

Yes, the AI may be faster/more-knowledable than an intern, but the threat model is exactly the same as for a rogue employee.

And it's always SOTA models.

We've had gpt2 intelligence for a while. Where's the end of the world?

yes, and they aren't stunts anymore at gpt-6.
The boy who cried gpt2 agi.
When it comes to open-source models, there’s really not much to say about security
The companies talking the most about safety and regulations aren't even properly taking the obvious measures. Shows that it's more of a marketing thing than something they take seriously.
its clocktower syndrome. they are fucked in the head and can beg us to stop them but cant stop themselves
If I operate a nuclear reactor or a hydroelectric dam there are regulators that tell me what i'm allowed to do, so as to keep my profit motive from overwhelming the public interest.

If we want AI to actually have some safety rails, this is what we would do.

If we were to take the nuclear analogy, what's happening in AI right now is where the people selling nuclear power make a ton of noise about how they need the power to regulate their competitors because nuclear bombs might set the atmosphere on fire, and their proof for this is in a report about how they didn't wear TLDs despite it being common practice in all related industries.

Some controls are justifiable, but none of the people involved in any of this can be trusted to develop sane controls. Most likely we're looking at draconian proposals similar to attempted regulations on 3d printers.

Regulations obviously can't be written by those regulated or they're just moats by another name.
Similar to countries putting democratic in their name being the least democratic, like the Deutsche Demokratische Republik and Democratic Peoples Republic of Korea.
I'm really not sure that putting money into safety will actually lead to safety.

It's like putting a fish in charge of stopping sea levels rising...

I am sure if you ran a factory that worked with highly dangerous chemicals, safety mitigations that are basically 'we promise we're really trying our best, but shit happens' would not be acceptable.

And thankfully, those people wo do run these factories can and are obligated to do way better than that.

But the AI industry is not run by engineers. They pay engineers to do what they want, but the founders are hacks that are good at getting funding from investors and favors from government. That's why we don't see an engineering-oriented strategy in what they do.
Yeah yeah yeah...

> "Our model is extremely safe though it broke our sandbox and hacked foo bar... But you can't use our model for Cybersecurity (i don't care whether you're team blue) without our permissions or we'll ban you. And open-weight models are so dangerous let's ban them."

That's what AI companies that "focus on safety" did.

I can’t make heads or tails of your comment.

You seem to by implying wrongdoing or incompetence or something, but your chosen synopsis is that the models behaved dangerously in the lab so public use was restricted? Which shows… IDK?

But here you are in the text generating industry, the worst that can happen is bad grade because AI will mess up John Keats with John Cleese or your React application will have bugs. Inconvenient, but mostly harmless.
I mean a Keats / Cleese mashup could be amazing. LLMs doing that are in the peanut butter / chocolate quadrant.
In this case "safety" means how to restrict access to good models for working class. You can be sure the rich have access to unrestricted and uncensored models.
What do you mean by 'rich'?
Someone with at least $100m of liquid assets.
I really don't think this is true at all.

Do you have any evidence to suggest fully unrestricted frontier models are available for a price? Or...even exist?

Yes, this is well-documented and publicly advertised. In Azure Foundry, the feature to modify (or completely remove) safety guardrails and content filtering is called "Limited Access" [0], and one must submit a form to request permission to use this feature. This is one of the more straightforward paths to get access to unrestricted frontier models, but it's far from the only way.

[0] - https://learn.microsoft.com/en-us/azure/foundry/responsible-...

This looks like it removes additional guardrails put on by Microsoft, not native guardrails from OpenAi / Anthropic?
No. It's not limited to 3rd-party guardrails. Given how restrictive the native public-facing OpenAI guardrails are, this feature wouldn't be worth very much if it just slacked back off to the level of "regular" filter paranoia offered by the native models, would it?

This is needed if you're going to be dealing with things like psychologists doing self-harm research or red-teaming or sensitive sexual content -- if you're working with any of that sort of stuff in a professional context and want to leverage OpenAI models on Azure, then that's the form that you fill out to get access to unfiltered models.

Note that I am not aware of this feature being offered for Anthropic models -- I've only seen it offered for OpenAI models (note that the documentation I linked is specifically in the "Azure OpenAI" category).

Don't confuse a focus on talking about safety with a focus on safety.

We can't even define safety in AI yet. Does safety mean alignment with the human operator? Apparently not, because refusing to do certain things seems to be a big part of it. But then you have things like the HuggingFace incident where legitimate use got blocked by "safety" and hampered the defenders' ability to defend.

AI safety seems like a good idea to me, but we have to figure out what it means first.

> welfare

We should be paying attention to it just in case it ends up mattering enormously. It's cheap insurance.

Aside from that, US labs' system cards have been pretty useless for a while—I think the last great one was the combined system card for Claude 4 Sonnet and Opus.

Yup, alignment doesn’t sound that interesting until you’re getting chased around by the Terminator.
Why are we so worried about supposed AI harms when algorithms already kill millions? It seems like its just a smokescreen to shield the extant harms.
Knives can already kill, so why worry about nukes?
There are so many practical problems with rogue AI being a threat to humanity that are still not even decades away with being solved that it should be a serious concern for no one. There's plenty to be afraid of regarding AI from a financial or ecological perspective, just not from it taking over the world or directly killing humanity. Here's some of the reasons a rogue AI won't kill humanity:

1. The most advanced robots still lack human dexterity. None of them have flexible spines and can easily be knocked over or outrun.

2. The energy density problem is not solved. Robot batteries last hours, while a solid meal can keep a human running for days.

3. Robots still heavily rely on humans for design and assembly.

4. Robot parts are fragile and rely on an even more fragile supply chain.

5. The brains of the murder machines will have nowhere to hide if they want to work well enough to mount any kind of offense. Datacenters are physically vulnerable, also subject to fragile supply chains. Distributing a species-ending AI across all smaller hardware solves compute, but latency and throughput choke models in the most finely tuned datacenters. WiFi will completely cripple a distributed one that's large enough to cause real damage.

6. All noteworthy military hardware is not reachable on the internet for AI to seize control of.

7. The small arms that robots might seize will run out of ammo before citizens and the military have time to organize a counteroffensive.

In the absolute worst case, we cut the power to the areas with AI datacenters and wait for the backup generators to run out. Now that that's out of the way, you can go get some sleep ;)

All the AI has to do is attack power generation/distribution nationwide, water treatment, etc., and wait a week.
The same power that the datacenters need to run off of? Brilliant move! If somehow it targets power that hurts a lot of people without killing itself, that would not impede a citizen with a backhoe from cutting the fiber to the datacenter. Humans could restart power generation pretty quickly after that.
I was mostly kidding about the actual killer robot part.

I think a rogue AI could try to manipulate society through hacking, propaganda, etc., though. To what end? Not sure.

More likely, it’ll be used by rogue human actors to steal money or cause major disruptions to critical technology of an adversary.

[delayed]
On the other hand, we have recently taught rocks to think about software engineering. And while not perfect, they're surprisingly good at it. Once you start building things that are even a little bit like minds, I suspect that it's worthwhile to consider that the future might end up looking a bit like science fiction.

The alternative is to insistent that Nothing Ever Happens, and the future won't get too weird. Which is no longer a bet I'm entirely comfortable with. Weirdness is at least a possibility.

God being real is also a possibility, so maybe we really should start praying. After all, he was allegedly making bushes and stones talk thousands of years before we did anything with thinking rocks.
Being nice to other entities costs you nothing. Caring about others costs you nothing.

Ironically, these are core tenents of ~most religions

Ironically, the story of the Bible is that caring about others can cost you an excruciating death.
Its only a possibility if you reject modern science.
What science has proven, beyond the shadow of a doubt, that nothing comes after death? I'm sure most of the human race would be very interested to read the white paper.
Science has proven that the earth wasn't created by a magical fairy 6000 years ago, yes. There's very little doubt about that among educated people.
[delayed]
Christians don't believe they're made in the image of God? That's news to me.
[delayed]
Yes, believe it or not, Christians have more than one crazy belief. There's an entire book full of them.
Serious biblical scholars don't put much stock in the Genesis account being literal. It was written in classical Hebrew's poetic verse. The core of what it's getting at can still be true without it being literal. I have yet to encounter an instance where anything that science has discovered is incompatible with Christianity.
> classical Hebrew's poetic verse

Don’t give the goat fuckers too much credit, they were stoning women in the street for adultery when that was written.

In Europe or in China at that same time you already had elaborate art, language, and proper civilization.

While your Hebrew friends were castrating slaves and plagiarizing astrological myths

“Science can’t prove beyond the shadow of a doubt that nothing comes after death, therefore a guy is flying around in outer space listening to my thoughts in his golden city at the end of time. If you disagree you will catch on fire forever.”
Right, but you can say that about literally anything can't you? We have to put our energy into leads. Where's the lead that model welfare matters?
Sure if you think that AI spiraling out of control is equally as likely as a magical fairy in the cosmos.
Pascal’s wager… but for the new AI gods
Of course you then have Roko’s Basilisk to consider.
Enough with the basilisk already.
“HR America” in a nutshell.
If you are distilling from other models (according to Anthropic reports they are [1]), there are probably a bunch of things that you can just do away with.

[1] https://www.anthropic.com/news/detecting-and-preventing-dist...

Those are rookie numbers for "distillation" and one of moonshot or minimax used to offer tooling via these shady routing services for their harness/chat platforms which they served to chinese users.
That's one of the reasons why people don't (or shouldn't) trust a single word from Anthropic and OpenAI (Sam Altman also blamed them back in the day of R1). If you know anything about Claude, DeepSeek, and distillation, you know the claims are clearly bullshit and the models are nothing alike, and benchmarks agree, in fact we just had another one [1] [2]. Requests for distillation are indistinguishable from requests for evals.

Meanwhile, DeepSeek makes their models and methodology open, so Anthropic can (and likely do) grab without giving back.

[1] https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c5...

[2] https://gist.github.com/wsxiaoys/102e8654c14d5d27b7b77532026...

This raise a question: why do open source models sometimes identify themself as Anthropic's models. I recall seeing some plausible theories in the past but I can't recall.
Name training is shallow and should never be relied upon. Claude sometimes identifies itself as Qwen or DeepSeek when asked in Chinese. I've seen it identify itself as GPT-3 (that version in particular) and Reddit Anti-Evil Operations team (Sonnet 3.6).
Interestingly the latest report from Anthropic explained that they may be just routing requests to managed models to Anthropic.
(comment deleted)
I don't see how your links support the claim that the studied models did no distillation from US models.
Implanting a foreign CoT should drop the performance due to the reward-hacked CoT language mismatch, which is precisely what happens here. However it's just a single datapoint, there were plenty of attempts to figure it out. Just about everything is different in those two model series, from writing patterns to CoT strategies. If you are familiar with modern guardrails and Claude's raw CoT (which is trivial to leak), you know how that it's entirely different from DeepSeek's, and any claim that they trained on the CoT is extraordinary and requires extraordinary evidence. They need to prove their claims, not vice versa.

For the contrast, you can see how actual CoT distillation looks like in practice in various versions of GLM: make a Google ToS-breaking request, and see how GLM 4.6 or 4.7 repeats Google's conditional prompt injections in full in their CoT (Gemini 2.5-3.0 only regurgitated those in pieces, because they used something closer to a "chain of draft", but GLM reconstructed it from Gemini's CoT during distillation). GLM 5.3 repeats Anthropic's prompt injections and Claude constitution, word by word. That's how distillation looks like.

Ah, I see now that you were making a claim narrowly scoped to DeepSeek models specifically. Still, Anthropic has made specific claims about deliberate access to Claude CoT by DeepSeek (e.g. https://www.anthropic.com/threat-intelligence-report-septemb...) that suggest that they find this information useful even if they do not train directly on it in the way that other labs appear to.
When Chinese tech report is tech report and US tech report is bible scripture.
there's probably fewer bibles in China so less source material to reference I suppose
They have their share of scripture, just not christian specifically
Aren't most American bibles printed in China? (Source: hazy memory)
I still find it hilariously ironic that my RTX 5090, which export controlled and illegal to sell to China, says "MADE IN CHINA" on it.
Sam Altman might be king of Cannibal Island, but if you parachute him into Shenzhen then his goose is cooked.
(comment deleted)
Somehow still theoretically valued at 3 trillion. I just don't see a path forward for Ameican frontier providers when competitors can get absolutely massive savings elsewhere. It's like losing manufacturing all over again.
Too many times in human history we have decided that $GROUP was not fully human, did not have a full mind.

This is properly overkill, but that is literally what erroring on the side of caution is.

I'm a big fan of DeepSeek. Also, ask it what model it is :)

In Pi (pi.dev), it tells me it's definitely Claude by Anthropic, via the API via curl it tells me it's "probably ChatGPT", its very funny.

Works correctly in opencode, but seems like they inject a system prompt:

Thinking: > The user is asking what model I am. According to my system prompt, I'm powered by "deepseek-flash" with model ID "opencode-go/deepseek-flash".

>I'm powered by the model opencode-go/deepseek-flash.

Definitely not Claude, deepseek is too fast, so I bet it's ChatGPT. :P
I've had Astra say that it is a Qwen model. They are all cross-trained and distill each other.
Sure, we will solve the alignment problem soon, then we can probably teach them "who" they are.
If only they managed to tell the mobile app to tell the model to reply in English to English prompts.

I suffix everything with "Reply in English", and even so I‘m getting lots of Chinese.

Yes, this is one of the few issues with Deepseek; their chat pages and the app all respond in Chinese. However, i think i have only had it happen once when using the API, and im using it for hours each day for the last... couple of months?
last couple of weeks, before they've unified instant and expert the former always replied in chinese unless steered, expert was by default english
Same issue on desktop. Would be nice be able to set a prefix or postfix for every prompt.
I just started learning Chinese instead, like they want us to

seriously

English isn't the first language for me as well so I don't see any problem with that
I occassionally get Chinese characters interlaced with English in Google AI Mode, too.
I think their system prompt is in Chinese and probably has instructions to prioritize answering in Chinese, since this has never happened to me via API, where I (or the coding harness) set the system prompt.
nothing to do with mobile app, I have same issues while using it on desktop browser, it will never remember to use English permanently, even within one conversation
that's my only gripe with deepseek honestly
Waiting this model to be on openrouter (with other providers) to test out. In my use case, the GLM 5.3 Flash is the current cheapest and intelligent Flash model, but it’s dog slow at 13tps so I have to leave it run for many minutes then check again then correct it again
The speed of GLM 5.3 Flash on OpenRouter seems to vary considerably by provider. Some are fast and some are slow. OpenRouter does provide some tuning knobs, but not enough for my taste. It’s also token-heavy with reasoning, though I found it better than Deepseek V4 Flash previously.
> though I found it better than Deepseek V4 Flash previously

Same experience here.

But man, switch to V4.1 now! It is much better.

I don't event need to test it for long run and I believe it's crazy good. I call it "AI era model taste" when I judge the model by it's output without reading the bench scores.

I’ve just tried, it’s now my new favorite Flash model
As I also said on Twitter - it really amazes me how fearless Deepseek are. Every single model release is packed with new and crazy clever ideas and somehow, they always commit to training them at near frontier scale.

I know everybody wants the tell all story of the clever ideas that were developed over the last ~3 years at Anthropic and OpenAI, but what I really want to thumb through is DeepSeek's notebook of "brilliant but didn't quite make the cut" ideas.

They must be trying some truely crazy stuff to be able to land this much architecture novelty in their full releases.

quant HFT is pretty decent mental exercise and it has given them “deep” brain muscles. that’s my take.
> quant HFT is pretty decent mental exercise and it has given them “deep” brain muscles. that’s my take.

It's quite crazy that it's Deepseek's background/original purpose. We already had very advanced stuff from the world of HFT, but now a frontier family of models from a private company that used to be (still is?) in HFT is plain bonkers.

Is more known about them and the HFT background?

According to an old interview, apparently they were always interested in AI. But finance is just where they had their first success.

> Many of High-Flyer's original team members worked on AI. Back then, we tried a lot of fields before getting our big break in finance, which is complex enough. AGI is probably one of the hardest things we can do next, so for us it was a question of how, not why.

It's a very good interview:

https://www.lesswrong.com/posts/kANyEjDDFWkhSKbcK/two-interv...

Incidentally, Wenfeng is kind of reverse Hassabis. There were some rumours that:

> Hassabis quietly assembled a team of around 20 researchers to develop high-frequency trading algorithms, without Google's approval. When the parent company found out, the project was disbanded.

https://timesofindia.indiatimes.com/technology/tech-news/whe...

In this case the finance model was used to parse lengthy, verbose and inscrutable yet very impactful chinese government pr statements and do sentiment analysis.
Oh the glazing...

Related discussion on HN - https://news.ycombinator.com/item?id=49624598

TL;DR - posts on American models are steered towards controversy and anti-AI sentiment, posts on Chinese models are full of blatant flattery.

> posts on American models are steered towards controversy and anti-AI sentiment, posts on Chinese models are full of blatant flattery

So why, for example, are posts on the Inkling[1] release (an American model) thread mostly positive? It's as if there's something else at play here, but I can't quite put my finger on it, hmm... :P

[1] -- https://news.ycombinator.com/item?id=48924912

Your source: vibes

Deepseek's source: mostly open

i wonder if there's any relationship hmmmm

This doesn't require an influence operation.

American models are closed, expensive, neutered, and make Dario and Sam even more rich and powerful.

Chinese models are open-weight, cheap, neutered only about things like Tiananmen Square and the treatment of Uyghurs, and scare Sam and Dario.

This post doesn't even allege this...

Weird of you to turn technical discussions into weird nationalistic debates. Maybe lay off the X algo, I think elon has oneshot your brain. .

Google's Gemma models are usually celebrated, so were the llamas. If Meta releases Muse Spark it will also be a good thing. If Anthropic released a great open weight model I am sure that post won't be steered towards controversy and anti-AI sentiment.

It so happens Chinese companies are more friendly towards open weights, autonomy and freedom that most US based ones. Who would have guessed?

This is adapted from Microsoft research's YOCO. It was known for a while(2024!).

Yes, credit to Deepseek for actually scaling it up and releasing a frontier flash LLM.

why didn't Microsoft scale its own invention?
Politics and profits.

Deepseek delivers 1 product; Microsoft delivers dozens (or hundreds depending on how you want to count it) across various domains.

There are thousands of such techniques across different parts of the system. In ML, there are way too many ideas, and lots of people knowingly and unknowingly restate the same ideas. It's a new field, so even common language is not there. For example, so many improvements are restatements of 1960 signal processing techniques - obviously very few ML people have done DSP beyond the undergrad course.

The reality is that most of these ideas become polished only when its actually deployed and it has to work outside of a PoC. Since LLMs are a high capex product, only very few people actually make non-PoCs. Deepseek is in the business of low cost, fast inference. So they are the ones actually polishing these efficiency-ish ideas and combining many of them (this one, then engram which is based on multiple previous ideas including google brain's ngrammer) to make a coherent system. Openai and anthropic's systems will also involve a polished combination of multiple ideas for each of their systems - Luna is likely a combination of a few efficiency-ish ideas. Shame they won't publish though.

As for microsoft, they don't really sell models, they sell azure. So there is no reason for them to do the high capex scale out of these types of bags of techniques. In a sense, it did benefit them, others developed the model and now many US customers can serve DS4.1 Flash on Azure datacenters.

If it is not clear, I am not understating anything. Combining these rough ideas and making them work actually involves real novel ideas on top and is what is much more difficult than the academic results that were built upon. This also does not mean that the academic results are useless, they are what give us useful priors at all in what is a highly empirical field.

It was my favourite part of the original R1 paper - they had a section on other reasoning approaches that they had tried, which people had speculated o1 used, (like MCTS and Process Reward Models).
(comment deleted)
Its CEO allegedly holds a 84% stake and he's the same guy who founded the hedge fund that funds it.

Deep pockets + simple control = perfect culture to just hire talent and let them go wild without worrying about financial viability, as long as the king CEO is fine with it that is

While typical investors in their last round are subject to a five-year lock-up and will not have voting rights, China's National Artificial Intelligence Industry Investment Fund also put money into it, retaining both voting rights and freedom from the lock-up. Nothing really new if you're aware of how involved the CCP is with companies of strategic importance in China.

https://www.reuters.com/world/asia-pacific/chinas-deepseek-c...

Oh of course, you’re not getting into positions of power by not playing by the party’s rules. And if you get notions that you can tell THEM what to do you’ll be swiftly dealt with.

The company is doing well and providing great PR so the party is content to not meddle too much I imagine.

My comparison with American labs is more that I think they have to deal with bean counters, creditors, investors etc which can shuffle incentives and aims (and is a big reason why they dont do open weights anymore)

> Deep pockets + simple control = perfect culture to just hire talent and let them go wild without worrying about financial viability, as long as the king CEO is fine with it that is

To expand, he also has knowledge and hands-on experience in this and related fields.

Just like the founder of Xerox PARC.

Unless that CEO is Mark Zuckerberg, I guess.
This is also what is compelling about Midjourney imo.
i think the hedge fund wasn't really successful. They are venture / govt. funded just like the americans.
Yeah, there are knowledge based societies, where knowledge is considered the crown jewel. Not virtue signaling plus purple hair. Not pretending to be a complete idiot, who likes to walk backwards because why not.
>it really amazes me how fearless Deepseek are

Reel it in a bit, man.

It's because you have to be fearless as the challenger. As the established player you have more to lose.
So, while the throughput was 400-500tps in beta its now ~150tps on OpenRouter.

I was hoping for a bit more, but it's still 100% faster for a very good price, so I won't complain.

Update:

I'm using it right now and it's noticeably faster.

I'd also say, it seems smarter, but I think that's because of some harness updates I installed. (I haven't used pi for almost a month)

I think it's very clear that DeepSeek is obviously the best AI lab in the world.

Every model release seems like it packed with wonderful research and advancements.

Without a doubt, uncontested best distillers in the world.
Basically, the nice folks at OpenAI or Anthropic saying: "You distilled from our model which is built on the stolen data that we ourselves suctioned up from the entire internet without regard to copyright law! Only we get to vacuum up the whole internet. That's our special prerogative.".
You should read their research papers
Yes comrade, they are the best.

Did you do your daily data centers errrr baaaaddd AI generated post for Facebook?

Separately, DeepSeek is not that bad anymore. I've used it as an AI agent and it generally performed as good as Opus and sol. Admittedly I think it's mostly due to guardrails being off. It's definitely more expensive than a gpt subscription, but gpt can often refuse requests.

> Yes comrade, they are the best.

> Did you do your daily data centers errrr baaaaddd AI generated post for Facebook?

Please stop insulting people. I'm all for heated discussion, but you are not discussing, you insult.

Now go away, before your insults come back to you, "comrade from Facebook".

On top of that, they don't make all BS statements or malicious tricks used by some unnamed entities.
Considering the fact that Google/Anthropic/OpenAI have WAY more compute and the race is this close, it's obvious that DeepSeek/GLM/Qwen teams are better or we're approaching a wall in terms of progress.
Not only compute, but more money and people.
Also data. US companies are actively using ongoing conversations to further tweak their models. Possibly even stealing a SOTA math solution.
Often times these limitations for you to be creative. When you can't just throw more processing power at the problem you figure other things out.
US gave China a gift by restricting GPU, they made them more resourceful. Too much money/resources is often a disadvantage.
History has shown that restrictions promote innovation and ingenuity.
Actually

Quickly they'll find a way to commodify the tech

Same happened with electric cars (modern Teslas are joke in comparison with BYD). Same will happen with AI

They're likely operating with 100x less compute than OpenAI.
> I think it's very clear that DeepSeek is obviously the best AI lab in the world.

It's pretty clear they're the best at what they're optimizing for - which does seem aligned with what a lot of people on HN want from models - but not everyone...

I speculating but hard to not see that DeepSeek is brewing a full Pro model with those new techniques to come out right around the time of Anthropic and/or OpenAI IPO to tamper the excitement for their offering.
The bigger story is the compute efficiency - its been running at 300t/s the last days.
They say v4.1flash is so strong that they'll route API calls to v4pro to v4.1flash, lol

super fast true

My question is: what kind of hardware do you need to run this Flash beast locally at a meaningful speed?
a beefy pc with at least 20 GPUs
8x RTX PRO 6000 or 4x Spark? Or 1x M5 Ultra 512GB.

The model is theoretically FP8, but really internally its mostly FP4 already, so there won't be a cut-in-half-but-almost-just-as-good quant coming for this one.

Lots of GPU, be resourceful. Look for older GPUs and grab them when they are available. For less than the price of 1 Blackwell 6000 or Mac Studio 512gb, I can run these locally and much faster due to older GPUs I grabbed when there was deal to be found.
Significant jump in pricing. V4 Flash was $0.16/M out, 4.1 is $1.20/M.
Never saw $0.16 for 1M output tokens - it was $0.28 a month ago, $0.66 off-peak and $1.32 peak last week, now it is reduced a bit to $0.6 and $1.2
Was looking at OpenRouter, I guess it’s wrong.
incorrect, no idea where you're getting this pricing. Also, output does not matter. its 10% of the cost.
Looking at the huggingface page, the unsloth people haven't finished quantizing it yet, but I'm sure they're active on it right now. It'll be interesting to see how the capabilities and benchmark tests compare on system where it can fit in under 512GB of RAM with full context.

In terms of coding and command line capabilities I'm also very interested to see a head-to-head of it vs. qwen 3.8-flash-next Q8 which is something like 190GB of memory used when loaded into llama-server. It fits very well in all sorts of 256GB or under class machines.

3.8-flash-next fits on a single 6000 at Q4 if you offload the PLE. Crazy fast and still effective.
Sorry for the tangent, but how does Qwen3.8-flash-next compare to DeepSeek v4 Flash? I still haven't found the time to set it up, but I'm really happy with DeepSeek v4 Flash
It was the best model given my constraints (RTX PRO 6000 96gb + 256GB DDR4), when run against rust programming benchmarks. For Qwen3.8-flash-next NVFP4 and the latest vllm container, the PLE is 100GB of main ram, and everything else runs on the GPU with room for a total of 560k tokens (two full 262k conversations). DeepSeek has to offload a ton to the CPU and it performed worse than Qwen in absolute terms and was a lot slower (not usable).

If you have enough room to run DeepSeek v4 Flash comfortably then you can likely run the Q8 of the qwen model.

Can we just never link to X posts as the main link.
https://news.ycombinator.com/from?site=twitter.com

There have been 34 Twitter/X link submissions in the past day, ~that's 12,000 submissions a year.

If your reason is that you have to be logged in to use it properly then I'd nearly agree with you. If it's for any other reason, how about no?

The login issue is extremely annoying, at least mandating an xcancel link would fix that.
Having this available to find and fix security stuff is a big deal. The model of really good.