Interview with DeepSeek Founder: We're Done Following. It's Time to Lead | Hacker News Reader

Interview with DeepSeek Founder: We're Done Following. It's Time to Lead

thechinaacademy.org

126 pointsoli567959 comments

Its success stems from a refreshingly unconventional approach to innovation. Liang Wenfeng's philosophy of maintaining a flat organizational structure where researchers have unrestricted access to computing resources and can collaborate freely.

What's particularly striking is their deliberate choice to stay lean and problem-focused, avoiding the bureaucratic bloat that often plagues AI departments at larger companies. By hiring people driven primarily by curiosity and technical challenges rather than career advancement, they've created an environment where genuine innovation can flourish.

AI development doesn't necessarily require massive resources - it's more about fostering the right culture of open collaboration and maintaining focus on the core technical challenges.

The model you described probably works great (not just in AI) as long as it's not your primary and direct source of revenue with which you must pay back investors. Once it becomes your primary and direct source of revenue and you must generate some returns for investors or meet some revenue targets, then whatever you're doing somehow has to align with that revenue stream (often ruining the fun).

You're describing a lot of tech companies like Google that had all these different orgs that were money sinks not related to a direct source of revenue and funded by dominance in search and high margins. And these programs didn't necessarily yield great creative products. Quite the opposite.

Whereas if you have some objective measure that's driving your decisions, like revenue or customer engagement (proxy for usefulness), you can drive great results.

I think either method can work if you have the right culture.

Having the right culture is easier said than done.

The enshittification of nearly everything can largely be attributed to the difficulty of maintaining that culture of open-ended creation without direct accountability to revenue.

have to disagree there - the enshittification of nearly everything is 100% attributable to the seductive nature of rent seeking (more specifically, trying to gouge ever more in recurring revenue from people who you try to ensure have no other options). even companies who have innovative products and positive revenue streams have gone down that road.

I mean, why do you think they are rent seeking in the first place?

(Hint: to make a profit and deliver profits to shareholders and investors)

> to make a profit and deliver profits to shareholders and investors

This is only part of the reason. It's really to signal that they can provide ever-increasing value. This year's 5% stock price increase cannot preclude next year's 5% stock price increase. It should be obvious this can't go on forever but it seems investors today either don't notice or don't care. It's precisely why enshittification happens: costs (things the business does to make their products/services more appealing to the customer) go down and revenues (the prices of the products/services) go up.

Making a car and selling the car is profitable (presuming the manufacturer is able to attract buyers). Making a car and selling the car while holding back software features unless the purchaser of the vehicle pays a monthly subscription fee is profitable and rent-seeking. The former used to be what happened and increasingly the latter is what's happening. It's not because the car companies weren't profitable in the past, it's because they want to show investors that they can continue to grow their profits.

the goal of nearly every single company out there is to make a profit, but not all of them engage in rent seeking.

Have you seen the "returns" for OpenAI, etc.? All cutting edge research is subsidized by government or megacorps in USA.

They are not profitable. The problem is that they have to find their way to profitability because investors and shareholders need to be paid back. And because they have to do that, you could say that it "compromises" on objectives that would more rapidly advance the field like openly sharing their reasoning architecture.

I think most of what's said here has value. But be wary of the survivorship bias. There are also a ton of flat, lean, problem-focused and curiosity driven startups that _don't_ succeed. Their success definitely has a lot to do their talent and how they work, but also a lot of luck, too.

I don't buy that. Allegedly Google lost to OpenAI because the compute resources were allocated evenly and then each team shared to other teams. So it became a popularity contest instead of meritocratic allocation. And then Pichai tried to merge all the different AI teams making it even worse. From rumors by connected people on podcasts.

There has to be some structure to put the best ones first. The key problem is how to judge that.

https://en.wikipedia.org/wiki/Adverse_selection

https://en.wikipedia.org/wiki/Tragedy_of_the_commons

"teams" are a clue to there being an underlying hierarchy and division which may not exist at deepseek. If its a smaller self-organising team of people, there would be no such effect.

It is also common knowledge that google's internal team and advancement politics are already pathological -- against a background of a winner-takes-all, cooperation does not work.

While I agree that Google's advancement politics are concerning, it's far fetched to say that there's a winner-takes-all aspect - there's still a lot of remuneration/power/recognition to go around for everyone, just not unevenly distributed.

DeepSeek's results speak for themselves - they've built a competitive AI model with millions that matches capabilities of systems costing billions.

While debates about resource allocation and organisational structure are interesting, what matters is their demonstrated ability to innovate efficiently.

The proof is in their technical achievement.

This sounded a lot like Valve’s company structure/lack of structure.

https://steamcdn-a.akamaihd.net/apps/valve/Valve_Handbook_Lo...

Sounds like an exciting place to work.

For the curious, Valve uses a version of the "lattice organization". This style or management structure is attributed to Bill Gore, creator of Gore-Tex.

https://participedia.net/method/lattice-organization

Either you give people a clear idea what you will build and they will organize accordingly or you organize and they will guess what they are supposed to build according to the org structure.

https://en.wikipedia.org/wiki/Conway%27s_law

I mean, I guess you could call it unconventional since it was the status quo of basically every super-massive tech company way back in the early days of the tech sector, but has since been utterly eclipsed by like 4 companies the size of nations that can't seem to ship a single app without the input of 6,000 people.

Ah yes, classic staying lean and exposing your database[0]

https://www.wiz.io/blog/wiz-research-uncovers-exposed-deepse...

A faux pas for sure but we can't pretend Google, Meta, and OpenAI all haven't had similar issues even with their much larger team sizes.

2023 and 2024 interviews, https://www.lesswrong.com/posts/kANyEjDDFWkhSKbcK/two-interv...

> Liang Wenfeng is a very rare person in China's AI industry who has abilities in “strong infrastructure engineering, model research, and also resource mobilization”, and “can make accurate high-level judgments, and can also be stronger than a frontline researcher in the technical details”. He has a “terrifying ability to learn” and at the same time is “less like a boss and more like a geek”.

Maybe worth adding that the interview is from July of last year. This is not a recent interview. Still interesting but was not what I was expecting.

On the other hand, if you release something innovative in January, you probably had to already be on the right track in July.

I think this was super interesting, it sounds like he’s leaning more into “open” than openai is.

“In disruptive tech, closed-source moats are fleeting. Even OpenAI’s closed-source model can’t prevent others from catching up.

“Therefore, our real moat lies in our team’s growth—accumulating know-how, fostering an innovative culture. Open-sourcing and publishing papers don’t result in significant losses. For technologists, being followed is rewarding. Open-source is cultural, not just commercial. Giving back is an honor, and it attracts talent.”

Inspiring and what the Valley (at least in part) use to represent. People doing cool shit as an end, not a means.

Says someone who personally has 50-100 billion USD. And no, it's not net worth through corp shares. The guy is essentially his own LP.

Correction: I meant 5-10 billion USD, too drunk yesterday

It's a great interview throughout, but I was thrown off by this strange question (which I found to be much more interesting than the answer):

> An Yong: What do you envision as the endgame for large AI models?

I don't know if it has a different meaning/connotation in Chinese, but reading this metaphor with a Chess connotation scared me. If there is a game, who are the players? what is the victory condition? will there be a static stalemate, or a definitive win? and most importantly, will there be an opportunity for future games after it, or is this the final game we get to play?

It’s a pretty common phrase for “what’s the ultimate goal?”

I don’t think it’s meant to be taken as a chess metaphor.

While less figurative, I don't see how "what's the ultimate goal for large AI models" makes it less scary.

Some of it might have to do with my having recently watched Dune Prophecy (set in the aftermath of the Butlerian Jihad) but this recent rapid progress in AI is putting me in somewhat of an apocalyptic mindset.

In a world where work doesn't exist, what happens to a society built on work ethic? What is an economy without labor?

It didn't sound awkward or weird to me at all. I think you took a very common word, and then extrapolated it out in a chess context when it's nothing to do with a chess context.

Isn't "endgame" a common expression to mean "the end", "the place where there's no progress anymore" etc ?

Sorry, am I the only one who finds this sort of formulation in regard to large AI models existentially intimidating?

i guess? but what are you hinting at exactly?

I am legitimately confused at other people around me thinking that this exponentially evolving technological explosion will end at a steady state that will be at all familiar to us.

It's a worn-out metaphor, but I can't help but think of horses marvelling at this mechanical carriage thing, wondering what's the endgame of that.

  > will end at a steady state that will be at all familiar to us.

admittedly, it probably wont, but i think a lot of replies like mine above are because your hinting at something vague... what's the concrete worry?

  > horses marvelling at this mechanical carriage

so you mean to say, the end-result of all this is humans will be out-of-the-job so to speak?

Indeed.

But the biggest thing is that I just don't think that there is anything resembling a steady-state outcome. There's a really interesting book called "Singularities: Landmarks on the Pathways of Life" from (oh, wow, time flies) two decades ago, which describes in each of its chapters a major biological transition whereby a new "technology" was adapted by a particular organism, and then rapidly took over the whole of Earth's biosphere, by virtue of enabling it to either utilize its environment better or evolve more quickly or both. Some example topics covered are: ATP, RNA, Proteins, Membranes, Oxygen utilization, Multicellular organisms, etc. Each change being so dramatic that whatever came before almost loses relevance.

I'm concerned about the rise of AI being the next such singularity, after which we humans would become pretty much irrelevant to the state of the world. Not just out-of-the-job, but similar to horses (or better yet coral reefs), living in some specific ecological niches, away from where the "interesting" stuff happens, with our existence in peril due to forces out of our control.

  > Not just out-of-the-job, but similar to horses (or better yet coral reefs), living in some specific ecological niches, away from where the "interesting" stuff happens.

interesting, i could see that happening... one does wonder to what end ai is being developed so aggresively.

  > our existence in peril due to forces out of our control

i think i understand what you are saying

while its a different thing i think there are already a lot of systems in the world that collectively slipping out of our control, even if nominally we have agency (see climate change, or social media, market bubbles etc)

just my personal bias, but i feel that if ai "takes over" so-to-speak its in a long line of succession of our technologies controlling us instead of the other way around... idk does that make sense? probably im not contributing much to this conversation; just my thoughts...

Yes.

Definitely not.

I think it means more "the place where there's no existential risk any more"

Please say more. What is this state of AI evolution where there's no existential risk any more?

for AI, probably when it's so entrenched in everyone's lives that it's treated as an essential aspect of computers and computerised devices

Oh, I see how at that point AI won't face any existential risk any more. Too bad about us humans though

oh, right, I wasn't using "existential risk" in the sense of AI taking over the world or not, just explaining how the term "end game" is usually used in this context. could be something as simple as a startup getting acquired and therefore no longer having the existential risk of running out of money and dissolving.

I think this project is awesome and am quite disappointed with some cynical commentary from large American labs.

Researcher at Meta or OpenAI spending hundreds of millions on compute, and being paid millions themselves, whilst not publishing any of their learnings openly, here a bunch of very smart, young Chinese researchers have had some great ideas, proved they work, and published details that allow everyone else to replicate.

    "No “inscrutable wizards” here—just fresh graduates from top universities,    PhD candidates (even fourth- or fifth-year interns), and young talents with a few years of experience."

    "If someone has an idea, they can tap into our training clusters anytime without approval. Additionally, since we don’t have rigid hierarchical structures or departmental barriers, people can collaborate freely as long as there’s mutual interest."

Why did you group Meta with OpenAI here?

Is that why if you ask it... it says it's based on ChatGPT4 ?

Most LLMs do this due to the proliferation of ChatGPT-generated content in the training data.

At the heart of all progress is the mantra that "best idea wins".

Maybe DeepSeeks creative use of RL within LLMs will open up founder and VC interest in using RL to solve real problems - I expect to see a cambrian explosion of high growth applied RL startups in engineering,logistics,finance,medicine

Doesn't matter if and how much they used OpenAI's models. The only important thing that matters is that they managed to disrupt the status quo, Silicon Valley will need to be more aware going forward.

They are nice words, ironically though their product is an exact clone of a US product (apart from the data stealing discussion). You could argue the cheaper aspect is innovating, but that's what China has been doing for many products.

To the best of my knowledge, there's nothing quite like R1-Zero released by OpenAI or others, they seem to really be pushing innovation.

Relevant ARC-Prize post and discussion from yesterday: https://arcprize.org/blog/r1-zero-r1-results-analysis https://news.ycombinator.com/item?id=42868390

An iteration at best, some Chinese electric cars have a better battery then Tesla and are cheaper. Yet would hardly call them innovative as he claims the chinese should become. It's actually what the Chinese have been doing, copying and making it cheaper with slightly different features.

Didn't expect to be cheering for Chinese AI companies and Facebook over mega funded US tech corps, but here we are.

Were fortunate that not all SOTA AI models are controlled by US Tech corps. Right now they're in the "maximum marketshare at all costs" stage, but they'll be looking for their ROI after achieving a dominant share. I trust OpenAI the least, it's still early on in the AI age and they look like the company that they were formed to prevent.

Can only hope that DeepSeek, Facebook, Qwen and Mistral continue to release open models. Unfortunately if a companies motivation is ROI from cloud hosting then they're going to be incentivised to stop releasing their models as OSS to prevent competition which we've seen with Mistral's best models although in their latest model released today under Apache 2.0 the CEO is saying they’re renewing their commitment to Open Source [1], so we’ll have to see how long that holds. We're also starting to see that from Alibaba whose latest Qwen2.5-Max model is only available through their Alibaba Cloud. Luckily Facebook business model isn't reliant on cloud hosting so we should continue to expect Open models from them. So far efficiency seems to be DeepSeek's competitive advantage as despite being OSS they're still the cheapest hosting provider [2] despite other hosting providers not having to recoup any R&D and training costs.

[1] https://x.com/arthurmensch/status/1884972984202338450

[2] https://openrouter.ai/deepseek/deepseek-r1

Its open source vs closed source, not China vs US

but this new dimension of geopolitical competition is now sidelining the cautionary anti-AGI populace, which was honestly probably saving us a few years

Reply on news.ycombinator.com