Thread by @_KarenHao

As someone who has reported on AI for 7 years and covered China tech as well, I think the biggest lesson to be drawn from DeepSeek is the huge cracks it illustrates with the current dominant paradigm of AI development. A long thread. 1/

First, what is DeepSeek? A Chinese firm that was able to produce an open-source AI model with roughly 1/50th of the resources of state-of-the-art models yet still beat OpenAI’s o1 on several benchmarks. 2/

Much of the coverage has been focused on US-China tech competition. That misses a bigger story: DeepSeek has demonstrated that scaling up AI models relentlessly, a paradigm OpenAI introduced & champions, is not the only, and far from the best, way to develop AI. 3/

Thus far OpenAI & its peer scaling labs have sought to convince the public & policymakers that scaling is the best way to reach so-called AGI. This has always been more of an argument based in business than in science. 4/

There is empirical evidence that scaling AI models can lead to better performance. For businesses, such an approach lends itself to predictable quarterly planning cycles and offers a clear path for beating competition: Amass more chips. 5/

The problem is there are myriad huge negative externalities of taking this approach - not least of which is that you need to keep building massive data centers, which require the consumption of extraordinary amounts of resources. 6/

Many journalists have written extensively about those externalities.

When these massive data centers come to town, they distort power supplies (per @business’s @Leonardonclt, @naurtorious, @andretartar)… 7/
bloomberg.com/graphics/2024-…

…they take up drinking water - yes, drinking water, because of the water quality required to cool these facilities - which threatens to hike water prices for families… 8/
theatlantic.com/technology/arc…

…they extend the lives of gas & coal plants, worsening air quality (per @washingtonpost’s @evanhalper) - not just because fossil fuels are the fastest way to surge energy supply but because 24/7 facilities can't run solely on intermittent renewables. 9/ washingtonpost.com/business/2024/…

Then there are the global impacts. More coal & gas means more carbon emissions, accelerating our climate crisis. Hence why Google & MSFT saw a ~50% & 30% jump in emissions since 2019 & 2020 respectively (from @darakerr at @npr). 10/
npr.org/2024/07/12/g-s…

OpenAI & co have come up with a funny argument for why to accept these externalities: Yes unfortunately there are trade-offs in the short term, they say, but they're necessary to reach so-called AGI; AGI will then help us solve all these problems! So trade offs = worth it. 11/

This line of thinking has been so effective in capturing the mindshare of powerful decision makers that president Trump just last week announced the Stargate Project, a $500B private investment into data centers & other computing infrastructure for OpenAI. 12/

Should Stargate follow through with $500B, it'll be the largest infrastructure spend in history and significantly accelerate the already startling pace of data center development post-ChatGPT. In turn that will significantly accelerate the aforementioned consequences. 13/

Now on top of the negative externalities on people’s power, water, air, and global climate, we have another: the ceding of more & more control over critical energy & water infrastructure to Silicon Valley. I wrote about this last week in @TheAtlantic. 14/ theatlantic.com/technology/arc…

But DeepSeek shows - in the same moment as the Stargate announcement - that the trade off that OpenAI & co frame as wholly necessary is actually not. 15/

Wait a minute. You mean to say that we don’t need to blanket the earth with data centers and coal & gas plants to maybe arrive at a future where we can wave a magical AGI wand to make all of the consequences of that go away?

Yes. This is a false trade off. Let that sink in. 16/

As I said earlier, scaling was always more about business than science. Scientifically, there’s no law of physics that says AI advancements must come from scaling rather than approaches using the same or fewer resources. Scaling is just an incredibly easy-to-follow formula. 17/

DeepSeek now also undermines the business case for scaling. OpenAI has been burning through staggering sums of cash to keep up its scaling paradigm and has yet to figure out how to balance its checkbooks - and it turns out it didn’t need to spend so much cash. 18/

So it doesn’t matter if you're a company in the US, China, or elsewhere. DeepSeek should be a cue to pivot HARD toward investing in far more efficient methods of AI development. Even if you care nothing about community & climate impacts, it’s just better business. 19/

And if a company does not make this pivot, that should be a major red flag for its capacity to innovate. And I don’t mean product innovation but innovation on how to develop AI models. The base of the stack. 20/

When I worked at an SV startup, we used to call an investment from SoftBank the kiss of death. SoftBank would pump so much cash into startups that it would completely kill that startup’s need to innovate or develop a financially sustainable business. 21/

They could mask all of the shakiness in their foundations with the hilariously large infusions of cash for quite a while — but eventually everything would topple. Necessity is the mother of innovation. Without necessity, the existential drive to innovate disappears. 22/

We’re basically seeing this again. OpenAI with Sam Altman is so good at raising capital (incl. now from Softbank) that it's covering up the company’s technical & commercial weaknesses. See also this excellent @AINowInstitute report from @bcmerchant. 23/ ainowinstitute.org/general/ai-gen…

DeepSeek is the flip side of the same coin: It innovated because of, not in spite of, its constraints. And now that it has upended the assumptions of the dominant AI paradigm, we should reject its costly trade-offs and seek new ways to develop AI without so many harms. 24/