DeepSeek R2 launch stalled as CEO balks at progress
reuters.com
reuters.com
I can't imagine demand would be greater for R2 than for R1 unless it was a major leap ahead. Maybe R2 is going to be a larger/less performant/more expensive model?
Deepseek could deploy in a US or EU datacenter ... but that would be admitting defeat.
>June 26 (Reuters) - Chinese AI startup DeepSeek has not yet determined the timing of the release of its R2 model as CEO Liang Wenfeng is not satisfied with its performance,
>Over the past several months, DeepSeek's engineers have been working to refine R2 until Liang gives the green light for release, according to The Information.
But yes, it is strange how the majority of the article is about lack of GPUs.
Although I'd like to know the source for the "this is because of chip sanctions" angle. SMIC is claiming they can manufacture at 5nm and a large number of chips at 7nm can get get the same amount of compute of anything Nvidia produces. It wouldn't be market-leading competitive but delaying the release for a few months doesn't change that. I don't really see how DeepSeek production release dates and the chip sanctions could be linked in the small. Unless they're just including that as an aside.
It is pretty strange that DeepSeek didn't say May anywhere, that was also a Reuters report based on "three people familiar with the company".[1] DeepSeek itself did not respond and did not make any claims about the timeline, ever.
[1]: https://www.reuters.com/technology/artificial-intelligence/d...
If the journalists aren’t fully trusted in the first place… trusting them to strictly adhere to even the best codified rules seems even less likely.
(A lot of things break down in society without trust, maybe that's already how the US is? Where I live it is thankfully still somewhat ok)
The Washington Post, The New York Times, The New Republic, The Intercept, Rolling Stone, CBS News, CNN, Newsweek, USA Today, NBC News, Der Spiegel (Germany), The Sunday Times (UK), Daily Mail (UK), Al Jazeera (Qatar), RT (Russia), Xinhua (China), Press TV (Iran), Haaretz (Israel), Le Monde (France), El País (Spain) all have been caught using fake anonymous sources.
No one I’ve ever heard of on HN fully trusts journalists.
This is why we need to be critical of journalists nowadays. No longer are they the Fourth Column, protecting society and democracy by providing accurate information.
Especially since the alternative is to live in a world without facts.
Which some people would probably love, but I prefer my reality to be constructed from objectivity rather than authority.
The tendency to compare to a nonexistant ideal is also something I find very very weird. This tendency does not exist for many other concepts. For example when people talk about communism, and someone say "hey $COUNTRY is just one bad apple, it doesn't mean real communism is bad" then others are quick to respond with "but all countries doing communism have devolved into tyranny/dictatorship/etc, so real communism doesn't exist and what we've seen is the real deal". I am not criticizing that (common) point of view, but people ought to take responsibility and apply this principle equally to all concepts, including "journalism".
It also doesn't follow that my critique of journalists/journalism means tearing down journalism altogether. It can also mean:
- that people need to stop trusting mainstream journalists blindly on topics they're not adept in. Right now many people have stopped trusting mainstream journalists only for topics they're adept in, but as soon as those journalists write nonsense about something else (e.g. $ENEMY_STATE) then they swallow that uncritically. No. The response should be "they lied about X, what else are they lying about?" instead of letting themselves be manipulated in other areas.
- that society as a whole needs to hold journalism accountable, and demand that they return to the role of the Fourth Column.
Because certain political interests take the existence of a fact-based, independent power center as a threat to their own power?
And so engineered a multi-decade campaign to indoctrinate people against the news/media, thus removing a roadblock to imposing their own often contrary-to-fact narratives?
Pretending this happened in a vacuum or was grassroots ignores mountains of money deployed with specific intent over spans of time.
> It can also mean that society as a whole needs to hold journalism accountable, and demand that they return to the role of the Fourth Column.
I absolutely agree with this.
If I had my druthers, the US would reinstate the fairness doctrine (abolished in 1987) and specifically the components requiring large media corporations to subsidize non-profit newsrooms as a public good.
The US would be a better place if we banned 24/7 for-profit news.
Kek. Reminder after Sino India drama, India has basically 0 accredited journalist in China. The chances of Indian journalist "citing two people with knowledge of the situation" in Deepseek in Bengalurur before it's spreads over PRC rumor mill is vanishingly small.
First, nobody is training on H20s, it's absurd. Then their logic was, because of high inference demand of DeepSeek models there are high demand of H20 chips, and H20s were banned so better not release new model weights now, otherwise people would want H20s harder.
Which is... even more absurd. The reasoning itself doesn't make any sense. And the technical part is just wrong, too. Using H20 to serve DeepSeek V3 / R1 is just SUPER inefficient. Like, R1 is the most anti-H20 model released ever.
The entire thing makes no sense at all and it's a pity that Reuters fall for that bullshit.
Why? Any chance you have some links to read about why it’s the case?
More generally, with any hardware architecture you use, you can optimize the throughput for your main goal (initially training; later inference) by balancing other parameters of the architecture. Even if training is suboptimal, if you want to make a global impact with a public model, you aim for the next NVidia inference hardware.
Human progress that benefits everyone being stalled by the few and powerful who want to keep their moats. Sad world we live in.
It's about China being expansionist, actively preparing to invade Taiwan, and generally becoming an increasing military threat that does not respect the national integrity of other states.
The US is fine with other countries having AI if the countries "play nice" with others. Nobody is limiting GPU's in France or Thailand.
This is very specific to China's behavior and stated goals.
So instead of splitting hairs about that description, lets highlight an idea that actually, millions of people doing millions of things per day consitutes its own system, despite what name you call it or who collects the taxes. Observing the actual behavior of that system ("data driven"?) has more benefits than hairsplitting of nomenclature for political studies.
Why bother writing this? because simplistic labels for government actions in international affairs is Step 2 of "brain-off" us versus them thinking.
Let's find ways to remove fuels from the fires of war. The stakes are too high. Third call to start thinking instead of invective here. Negotiation and trade are the tools. Name calling on those that work for "peace" is Step 2 again. IMHO
Remove the word Taiwan and you are describing the US.
>It's about China being expansionist
US has been doing that since their inception as a country. Are you telling me the USs 750 foreign military bases located in at least 80 foreign countries and territories is NOT expansionism? Come on!
>actively preparing to invade Taiwan
The US illegally invaded Iraq and Afghanistan for 20 years killing and torturing innocents in the process and leaving the Taliban in power to further cause harm. Wow many countries did China invade? Yet somehow China is the boogieman? Please!
> generally becoming an increasing military threat that does not respect the national integrity of other states.
Same with the US, Trump threatened to annex Greenland and Canada, yet I don't see sanctions on the US.
I don't see the US having any ground to stand on criticizing China.
The real reason is that the US cannot compete fairly
> Thomas Edison's aggressive patent enforcement in the early days of filmmaking, particularly his control over motion picture technology, played a significant role in the development of Hollywood as the center of the film industry. Driven by a desire to control the market and eliminate competition, Edison's lawsuits and business practices pushed independent filmmakers westward, ultimately leading them to establish studios in Los Angeles, away from Edison's legal reach.
https://www.bakerlaw.com/services/artificial-intelligence-ai...
Seems like it would be a definitive list to me as it shows US AI companies getting sued for copyright infringement.
And, isn't this the system working exactly how it is supposed to? Someone makes a claim and the courts decide, and then some kind of punishment will be doled out of the claim was found to be true?
Like, even if you just want to talk about protectionism, China is way worse than the US pre-Trump. "Fairly" has nothing to do with foreign policy.
[1] https://www.aclu.org/news/smart-justice/president-obama-want...
No nation is perfect, but the US has historically been better than many others.
Citations? Apart from usual Western government propaganda outlets perhaps?
The problem is rather that if the only moral compass is the communist party it will suck
By those metrics, the rest of the world should have been terrified by the US for the last 60 years...
Those are necessary precursors to aggressive expansionism, but insufficient without political will.
But if China were only threatening to invade Taiwan it would be a gray area.
Imho, their claims in the South China Sea are much more obviously expansionist, given the settled cases against them under international law.
Much easier to see those boiling over into China invading a few populated islands of the Philippines.
To translate what you're saying. The Chinese are trying to establish the same kind of global trade collaboration that Europe and the US have done for the past hundred and x years? But the Chinese civilization is over 2000 years old, and they had a much larger global trade network back when the west was a pile of wooden shacks and feudal barbarism?
They're also building up a large army in in the same way that the US and Europe have with NATO? I'm also not really sure what's wrong with the moral compass of the Chinese communist party? From what I can see at the moment it is authoritative, but not necessarily venal?
It seems that the Chinese people themselves are enjoying a pretty good standard of living and quality of life? I've only been there two or three times, but I never saw the same kind of deprivation in China that I saw behind the Iron Curtain for instance.
It's certainly corrupt. Xi didn't launch major, disruptive anti-corruption drives for no reason, but because he saw it as an existential threat to the CCP's legitimacy (after all, it did torpedo the Soviet Union).
Granted, an alternate rationale was internecine power struggles within the party and removing political enemies, but there was some real corruption.
The strongman argument against the CCP's moral compass is that it has no concept of or respect for individual rights: the party is above all.
Historically, this has always ended tragically because eventually it will be abused to either justify suffering or party gain at the expense of people.
Authoritarianism only works until someone bad grabs the reigns, and single-party non-democratic systems have a way of rewarding sociopaths.
People may gripe about fuzzy areas being stepped on and norms pushed (and they should gripe!), but there's a huge chasm between separation of powers in democracies and China.
Not agreeing with a thing doesn't make it illegal.
If Congress wants to prohibit Presidents from pushing these areas, then they're free to do so. (And expect they will once the clock tocks)
But will they keep releasing the weights or do an OpenAI and come up with a reason they can't release them anymore?
At the end of the day, even if they release the weights, they probably want to make money and leverage the brand by hosting the model API and the consumer mobile app.
Now they are firmly on the map, which presumably helps with hiring, doing deals, influence. If they stop publishing something, they run the risk of being labelled a one-hit wonder who got lucky.
If they have a reason to believe they can do even better in the near future, releasing current tech might make sense.
What is DeepSeek aiming for if not that, which is currently the only thing they offer that cost money? They claim their own inference endpoints has a cost profit margin of 545%, which might be true or not, but the very fact that they mentioned this at all seems to indicate it is of some importance to them and others.
Given that DeepSeek is used by the Chinese military, I doubt that it would be a reasonable move for them to host in the U.S., because the capability is about more than profit.
It should only be quality which could be unpredictable before training.
With this combo, I have no reason to use Claude/Gemini for anything.
People don't realize how good the new Deepseek model is.
Personally I get it to write the same code I'd produce, which obviously I think is OK code, but seems other's experience differs a lot from my own so curious to understand why. I've iterated a lot on my system prompt so could be as easy as that.
There is something deeper in the model that seemingly can be steered/programmed with the system/user prompts and it still produces kind of shitty code for some reason. Or I just haven't found the right way of prompting Google's stuff, could also be the reason, but seemingly the same approach works for OpenAI, Anthropic and others, not sure what to make of it.
The large context length is a huge advantage, but it doesn't seem to be able to use it effectively. Would you say that OpenAI models don't suffer from this problem?
Yes, definitely. For every model I've used and/or tested, the more context there is, the worse the output, even within the context limits.
When I use chat UIs (which admittedly is less and less), I never let the chat go beyond one of my messages and one response from the LLM. If something is wrong with the response, I figure out what I need to change with my prompt and start new chat/edit the first message and retry, until it works. Any time I've tried to "No, what I meant was ..." or "Great, now change ..." the responses drop sharply in quality.
The published model has a note strongly recommending that you should not use system prompts at all, and that all instructions should be sent as user messages, so I'm just curious about whether you use system prompts and what your experience with them is.
Maybe the hosted service rewrites them into user ones transparently ...
Mainly the hosted one.
> The published model has a note strongly recommending that you should not use system prompts at all
I think that's outdated, the new release (deepseek-ai/DeepSeek-R1-0528) has the following in the README:
> Compared to previous versions of DeepSeek-R1, the usage recommendations for DeepSeek-R1-0528 have the following changes: System prompt is supported now.
The previous ones, while they said to put everything in user prompts, still seemed steerable/programmable via the system prompt regardless, but maybe it wasn't as effective as it is for other models.
But yeah outside of that, heavy use of system (and obviously user) prompts.
For anything that requires "AI level of intelligence", the difference is vast.
My consumer AMD card (7900 XTX) outperforms the 15x more expensive Nvidia server chip (L40S) that I was using.
Surely it would be cheaper and easier for the CCP to develop their own chipmaking capacity than going to war in the Taiwan strait?
The USA doesn't want to lose Taiwan because of the chip making plants, and a little bit because it is beneficial to surround their geopolitical enemies with a giant ring of allies.
that is what the CCP tells you and its own people.
the truth is taiwan is just the symbol of US presence in western pacific. getting taiwan back means the permanent withdrawal of US influence in the western pacific region and the offical end of US global dominance.
CCP doesn't care the island of taiwan, they care about their historical positioning.
In any case it's clear that it is not the fabs that China cares about when it is talking about (re)conquering Taiwan.
building their own capacity means building everything in China, that is the entire semiconductor ecosystem. just look at the mobile phones and EVs built by Chinese companies.
If I were China I’d be more worried about the other up and coming world power in India.
with a reality tv show dude being the commander in chief and a news reporter being the defense secretary.
life is tough in america, man.
That's not certain. Most war games show that the U.S. would lose the war (also read official reports done for congress). You can't win against the world's leading producer of goods by trying to attack them from the sea when they can reach you with rockets from their own territory.
America is a society and component of a civilization that has never really understood itself or its place in the world and history. We peaked in the industrial warfare of WWII, and then bumbled our way through trying to rely on our past self-involved achievement.
Maybe then we wouldn't be beholden to Nvidia's whims (sour spot in regards to buying their cards and the costs of those, vs what Intel is trying to do with their Pro cards but inevitably worse software support, as well as import costs), or those of a particular government. I wonder if we'll ever live in such a world.
But we have models developing and being produced outside of the US already, both in Asia but also Europe. Sure, it would be cool to see more from South America and Africa, but the playing field is not just in the US anymore, particularly when it comes to open weights (which seems more of a "world benefit" than closed APIs), then the US is lagging far behind.
Llama (v4 notwithstanding) and Gemma (particularly v3) aren't my idea of lagging far behind...
While neat and of course Llama kicked off a large part of the ecosystem, so credit where credit is due, both of those suffer from "open-but-not-quite" as they have large documents of "Acceptable Use" which outlines what you can and cannot do with the weights, while the Chinese counter-parts slap a FOSS-compatible license on the weights and calls it a day.
We could argue if that's the best approach, or even legal considering the (probable) origin of their training data, but the end result remains the same, Chinese companies are doing FOSS releases and American companies are doing something more similar to BSL/hybrid-open releases.
It should tell you something when the legal department of one of these companies calls the model+weights "proprietary" while their marketing department continues to calling the same model+weights "open source". I know who I trust of those two to be more accurate.
I guess that's why I see American companies as being further behind, even though they do release something.
Even worse, the "Acceptable Use" document is a separate web page, which can be updated at any time. Nothing prevents it from, for instance, being updated to say "company X is no longer allowed to use these weights".
The "FOSS-compatible" licenses for these Chinese and European models are self-contained and won't suddenly change under your feet. They also have no "field of use" restrictions and, by virtue of actually being traditional FOSS licenses being applied to slightly unusual artifacts (they were originally meant for source code, not huge blobs of numeric data), are already well-known and therefore have a lower risk of unusual gotchas.
HGX 8x Nvidia H100 cluster for sale.
You can buy whatever you want. Export controls are basically fiction. Trying to stop global trade is like trying to stop a river with your bare hands.
I miss the old days of journalism, when they might feel inclined to let the reader know that their source for the indirect source is almost entirely funded by the fortune generated by a man who worked slavishly to become a close friend of the boss of one of DeepSeek’s main competitors (Meta).
Feel bad for anyone who gets their news from The Information and doesn’t have this key bit of context.
You never know which stories The Information won’t run, or which “negative” articles are actually deflections. Similarly, you never know which amazing startups remain shut out of funding, and a lot of entrepreneurs have no idea about the amount of back channel collusion goes on in creating the funding rounds and “overnight successes” they’re told to idolize.
A random dude on HN such as me shouldn’t be the source of this knowledge. Hope someone takes up the cause, but we live in a time of astounding cowardice.
Then look up Latin America’s history, where the US actively worked to install and support such violent dictatorships.
Some under the guise of protecting countries from the threat of communism - like Brazil, Argentina and Chile, and some explicitly to protect US company’s interests - like in Guatemala
Yes fuckups happened. But then for results Russian intervention see CCP and how many people died from their hands and policies
The lesson of present-day America is that democracy is too important to be left to the people.
Lots of US companies got a lot of money out of those US-supported dictatorships, while destroying local businesses and torturing and killing people. Those were also the era of closed-off economies, hyperinflation and environmental destruction, so what the local people got out of it?
So yeah, thanks for protecting us from the dictatorship of the proletariat and fucking up our economies for decades. And I’m not also defending USSR and their imperialistic practices disguised as making the people as equal as possible - fuck them as well!
There’s an old book called “Confessions of an Economic Hitman” that gets into some details of how the US supported those dictatorships under Project Condor and other CIA programs. Is it 100% truthful? Maybe not, but the gist of it is.
The proles have let us both down. All my life, I was led to believe that a "dictatorship of the proletariat" would involve a bunch of morons wearing red hats, casting one last vote against their own interests to tear down the established order. So at least that turned out to be technically correct.
And spare us the false equivalence bullshit that we all know is coming.
> trade labor, goods and services voluntarily
capitalism is using capital (money, materials, and employees/work) as inputs to produce finished products with the goal of re-investing those profits into said production or into other markets
simply trading or rendering services can be done without the need for constant growth/profits or investment as capital over time (e.g coops, traditional businesses etc)
This is an important point. The only way communism "works" is top down enforcement.
Might as well talk about the probability of a conflict with South Africa, China might not be the best country to live in nor be country that takes care of its own citizens the best, but they seem non-violent towards other sovereign nations (so far), although of course there is a lot of posturing. But from the current "world powers", they seem to be the least violent.
China is peaceful recently, at least since their invasion of Vietnam. But (1) their post-Deng culture is highly militaristic and irredentist, (2) this is the first time in history that they actually can rollback US influence, their previous inability explains the peace rather than lack of will (3) Taiwan from a realist perspective makes too much sense, as the first in the island chain to wedge between Philippines and Japan, and its role in supplying chips to the US.
The lesson we should learn from Russia's invasion of Ukraine is to believe countries when they say they own another country. Not assume the best and design policy around that assumption.
If you want to read some experts on this question, see this: https://warontherocks.com/?s=taiwan
The general consensus seems to be around a 20-25% chance of an invasion of Taiwan within the next 5 years. The remaining debate isn't about whether they want to do it, it's about whether they'll be able to do it and what their calculation will be around those relative capabilities.
And on who you would support in such a conflict! ;)
DeepSeek is not a charity, they are the largest hedge fund in China, nothing different from a typical wall street funds. They don't spend billions to give the world something open and free just because it is good.
When the model is capable of generating decent amount of revenues, or when there is conclusive evidence of showing being closed would lead to much higher profit, it will be closed.
(In before “whatabout”: maybe US-made models do the same, but I’ve yet to hear reports of any anti-US information that they’re censoring.)
it's not even in the top ten based on OpenRouter https://openrouter.ai/rankings?view=month
or
who knows maybe they just chillin watching how west labs burn gpu money, let eval metas shift. then drop r2 when oai/claude trust graph dips a bit
Remember that DeepSeek is the offshoot of a hedge fund that was already using machine learning extensively, so they probably have troves of high quality datasets and source code repos to throw at it. Plus, they might have higher quality data for the Chinese side of the internet.
* Of course I won't detail my class of problems else my benchmark would quickly stop being useful. I'll just say that it is a task at the undergraduate level of CS, that requires quite a bit of deductive reasoning.
Be mindful of what this means. A kid in his garage fine tuning a model can "catch up" to SOTA models for most use cases. For actual "frontier" work that requires SOTA levels of intelligence, there are only 3 companies in the race. None of them are from China or Europe.
so what?
DeepSeek-R1 0528 performs almost as well as o3 in AI quality benchmarks. So, either OpenAI didn't restrict access, DeepSeek wasn't using OpenAI's output, or using OpenAI's output doesn't have a material impact in DeepSeek's performance.
https://artificialanalysis.ai/?models=gpt-4-1%2Co4-mini%2Co3...
I am not at all surprised, the CCP views AI race as absolutely critical for their own survival...
EQBench, another "slop benchmark" from the same author, is equally dubious, as is most of his work, e.g. antislop sampler which is trying to solve an NLP task in a programmatic manner.
"Follow the money."
Businesses are pouring money into the OpenAI API. This is your biggest clue.
To me that does seem like a reasonable speculation, though unproven.