Hacker News

Thanks for the clarification - Google does publish more than others - and I actually really appreciate the work they are doing with the Gemma models, which are truly competitive open models. I do wish they’d publish more in depth papers on their Gemma models but appreciate that they are open weights.

They weren't the first to do MTP like this, and arguably did it wrong: the MTP heads are kept in a separate file and have to be welded in by the inference engine.

Qwen 3.6 shipped with working MTP first, and had working MTP in llama.cpp first.

Chinese labs are also still behind, so they’re incentivized to collaborate and have no reason to do it in private.

I suspect their tune will change if they ever take the lead..

I don't really see the moat for frontier AI labs being "more efficient models" although that could help their margins - I think moats will be built by expanding the horizontal and vertical market expansion - like Anthropic is doing the most at the moment

Who is financing DeepSeek and what are they expecting in return?

I seriously am far from fear mongering and doomsday mentality, but I just can't see how OpenAI and Anthropic can have a successful IPO if the quality gap between the free and paid continues to narrow like that...

> Probably because American AI companies are on the hook for quite a lot of investment money

That's a lot of words to say it's just capitalist greed.

Do you think that DeepSeek are building their models for free, or something? They aren't "on the hook" for anything?

What's with all the China glazing about this stuff? They release some open-source work and people act like they are suddenly the beacon of freedom and transparency.

Yes, challenger Labs publish out of necessity. It is a marketing strategy. People assuming open source means giving something up, but the reality is that Z.ai has a revenue of some $100M and it would be about $0M if they never open sourced their models.

Wouldn’t that just help the American labs anyway though? Or do they assume they’ve actually already figured this stuff out and kept it secret?

I'm afraid I'm even balking at the word "pioneering" in context with US frontier labs. They are probably doing a few new things, right, but they are not blazing any trails for others to follow along, the Chinese are.

> Publishing by necessity

It's more a cultural thing. Sharing progress is just in their blood.

Chinese papers and techniques have been very influential and copied by US labs.

Multi-head Latent Attention (MLA), Multi-Token prediction, MoE architecture are some of the most famous examples.

I don't think it's very common to believe the Chinese people are bad guys. It's the government and its control of the people that's the problem. And no, I don't think the US is immune to that sort of problem either.

Honestly it's just a hierarchy difference between the two countries. In the US, tech/fin/military companies have the upper hand compared to the government (fragmented between 2 parties). Despite the sharades with Anthropic, Tech-fluencers are in control. Compared to china, the government (dictatorship) has more control over Tech companies (take any example from the past 10 years). For them, undermining the US AI supremacy is an objective, and releasing open weight models is the way, and I'm all for it.

Let's not get crazy here. You can acknowledge that the Chinese AI industry has some structural advantages right now without trying to claim anything else. China is still a brutal autocracy.

> All of the optimizations Deepseek have done are in software and it goes down to the PTX assembly level

DeepSeek are still using NVIDIA (PTX) to train on, but for inference have already transitioned to Huawei Ascend chips, and inference speed is what this paper is addressing.

Anthropic almost certainly also has optimized software down to the assembly level, considering this take-home interview challenge they published: https://github.com/anthropics/original_performance_takehome/... which is all about instruction-level performance optimizations. That they don't prioritize UI fixes just means they consider other things more important.

> Compared to Anthropic who are celebrating in fixing a flickering issue in a terminal app which took months to fix.

It's funny, because if you ran Claude Code on a slow terminal, the cause of the flicker was obvious: They kept dumping the entire history of the chat back into the terminal in a number of situations, and relied on the terminal to them end up in the correct state.

All frontier labs are working down to the PTX level (and lower)

Google hasn’t published much in depth ML work since T5 (which was hugely influential at the time) - most Gemma releases are 1-3 page model card pdfs these days with no in depth analysis. Even TurboQuant is shaking out to have basically been a rehash of previous work without proper attribution. I do think Microsoft is doing some interesting things with smaller models but haven’t read much research, interested in any refs you might have to share!

I don’t see how Anthropic is in a better position. They have a slight edge in model quality right at a time when we’re getting a taste of what cheap, “good enough” AI looks like. They don’t own their own compute. And their own arrogance and lies have alienated a huge chunk of their customer base and alerted everyone to the dangers of being dependent on them.

> Another data point on this is the black market for Claude tokens in China [1]. The chat logs themselves are a commodity to train models.

anyone with IQ higher than 130 (thus qualified for actual AI R&D) would be questioning something obvious here -

if they are already doing such dodgy stuff with the aim to maximize profits, why would those resellers have large amount of logs with actual American model responses to sell to those AI labs in the first place. shouldn't they just post train & customize some leading Chinese open source models to pretend to be Opus or GPT for the vast majority of their users (as classified by some models) who don't know much about expected Opus behaviours & not skilled enough to tell the differences?

that is actually the interesting bit not covered in your censored version of the story line, it is also what happens on the ground. your censored version of the story implies that those dodgy resellers using stolen credit cards, pooling accounts with stolen IDs and illegally selling very personal logs would somehow be honest enough to spend extra $ to ensure their victims (aka paying users) can actually use real Opus and GPT. LOL

dude, you failed this IQ test miserably.

Nope, people seek it out because government tells them to pay taxes _or else_.

Money is just a physical representation of the ability to get what you want. The problem is not money. It’s the fact that we live in a “me” society.

If your moat is “please don’t copy my outputs”, you don’t have a moat. There is no such thing as a distillation “attack”.

US AI companies trained their own models on vast amounts of copyrighted and publicly available content without obtaining permission. There's no moral high ground here.

Besides "attack" being a ludicrous name for distillation, note how your article says "accuses", also it's mostly about Alibaba, not DeepSeek (although it's mentioned there). Both Dario Amodei and Sam Altman publicly claimed that DS used their outputs to train their models, and knowing the differences between all these models by heart, I believe they're simply lying through their teeth to sway the public opinion and/or the policy. These models are absolutely nothing alike, and distillation necessarily makes student's outputs similar to teacher's. This is very visible in Z.ai models (which were trained on Gemini outputs to the point that they repeated Google's conditional prompt injections in the CoT, and later on Claude where it started repeating their CoT as well) and certain Google models which were trained on Claude's outputs in a roundabout way. Distillation always shows up in the result.

And certainly they have no idea whether these outputs (assuming they ever existed and it wasn't made up) were used for training. The article mentions that DS made 150k requests. This isn't much and might have been just an eval or a benchmark to compare their own model against. It's really hard to believe DeepSeek had any Claude outputs anywhere in their training schedule, since it's just too different. Besides training on random vibecode of course, which is mostly written by Claude.

While I don't agree with your comment being downvoted, I don't think distillation is either an "attack" nor is it "stealing". The idea that someone else gets to decide how I use tokens that I pay for is ludicrous.

Imagine if your casio calculator would come with a ToS that says you can't use it to develop a competitor calculator or any other tools. Or that your hammer can't be used to make other tools. Or, closer to the HN crowd, imagine MS in the 90s saying that you can't use their OS to build competing services to MS. They'd be laughed at and be split immediately if they tried that.

The only thing they can do is to refuse serving tokens (and even that's debatable, if we get to tokens being commoditised). But that's gonna be a game of whack-a-mole, and they know it.

The paper actually references testing their DSpark speculative decoding strategy with Qwen 3 4b, 8b and 14b models so while I doubt they will release builds themselves, they’ve open sourced (DeepSpec) their training pipeline for this so we will likely see folks adopting for other models.

There was a recent exodus from Qwen of researchers who supported their open source efforts, I’m not sure we will see many new open models from them past the 3.6 series.

> To compete in that direction, USG needs to learn from CCP to "seize the means of production"

No they don't. The U.S. Government is free to launch their own AI labs if they wish -- and even compete with the private sector -- but that doesn't mean they have to confiscate existing investments and capital. But Congress is unlikely to do that, because we've learned in the course of history that in well-functioning competitive markets, publicly-operated services tend to be worse than private ones across multiple dimensions.

Chinese companies are largely where they are not because they're state funded, but because they operate in ways that would be considered criminal in the U.S. If they didn't constantly trespass on OpenAI and Anthropic to try to reach parity, they would be too far behind to produce innovative research.

China is just taking a lot of ideas from the USG when it was doing things correctly and is using those for innovation.

In this case, it feels like they are just funding multiple independent pure research projects and letting the chips fall where they may.

Doesn't even really seem like Europe can coordinate that.

OpenRouter sometimes chooses a very expensive provider. Try the floor slug or choose directly the provider. I moved to just putting 5 dollars directly on deepseek instead of going through OR.

the casino has no clocks, as one HN user put it some time ago.

I second ccusage, it's nice

https://ccusage.com/

It's in the JSONs in ~/.claude, but last 30 days only I think. You can have the model analyze history. So for correct history you'd need to run history analysis on a cron job or something. Kinda hacky.

Yeah can definitely see a world where china pivots and we're stuck with closed/closed

Mistral...don't fumble this

I believe the comment you replied to was talking about the cost on providers like OpenCode vs Deepseek API. Deepseek API is even cheaper than the other providers for the same deepseek models.

This applies to every provider. OpenAI seems to be the worst hoarder.

US labs do it too.

US labs are the biggest data broker in the current history. They collect everything, dumb fuck.

> in a rational market.

Unfortunately the market is often not rational in this way.

Hype within retail market means there are suckers willing to buy. Institutional market knows there are suckers when the hype is high. Both would drive the price up, and retail investors the ones left when it falls.

IPOs with massive bags can be wework or spacex, it all depends on vibes. If they buy a couple more articles doomposting and glazing AI on the financial times right before exit they will def find a bunch of boomers to buy their bags. If the narrative changes before they IPO its over.

No one's buying that shitbox.

With MoE models like Deepseek’s and with multiple Crescent Island accelerators, the aggregate memory throughput actually doesn’t look that bad. Two Crescent Island gets roughly 1400GB/s and Deepseek-v4-flash with 13B parameters active nets roughly 100t/s which is decent for a small team or great for a single user.

More Crescent Island scale up, although not likely entirely linearly.

But all GPU inference work like this, it’s not specific to Intel. Just Intel promises more affordable cards with big memory so they’re attractive.

Every day, the rate of “could post a picture of 40k tech priests and have it taken unironically” goes up, and it’s starting to get concerning.