This Week’s Science News from SWTG

The AI Bubble & A European Ice Age

Did DeepSeek burst the AI bubble or will the Stargate Project make it worse?

In the past two weeks, the global AI playground has completely changed, and some are saying that this is the long-awaited bursting of the AI bubble. I have a brief summary of all you need to know, and why I don’t think the bubble is bursting quite yet.

On January 21st, Donald Trump announced the Stargate Project. Named after the 1994 movie Stargate, it’s supposed to symbolize a wormhole into an ancient pyramid. I guess.

It’s not a governmental project, but an umbrella for private investors to pour money into AI. The investors currently include OpenAI, SoftBank, Oracle, and an Abu Dhabi-based investment fund named MGX. The initial investment is $100 billion and by 2029 it’s supposed to reach $500 billion, at least that’s the plan. Elon Musk promptly rained on Trump’s parade and claimed that they don’t actually have even the $100 billion.

Just two days later, the Chinese Firm Deepseek released another open source AI model called r-1, a reasoning model similar to GPT-4o, and briefly afterwards an image generation model called Janus Pro.

Neither is the absolute best in its category, but among the best. This made it clear that China is not behind on AI, and if we factor in the possibility that they’re not telling us the whole story, maybe China is actually ahead of the rest of the world.

What really upset the Americans, though, is that the Chinese say they can do AI equally good but at vastly lower cost and with less powerful chips.

The numbers that have been circulating is that they did with $6 million what cost OpenAI more than $100 million, and made do with 2,000 GPUs where OpenAI used 100,000.

To add insult to injury Deepseek is built on Meta’s open-sourced Llama, and was most likely trained on GPT output, which absolutely no one could have seen coming.

The Deepseek people said they did it with a lot of small efficiency improvements that add up to a major difference. There’s been some discussion about whether they’re just faking numbers, and some have pointed out that the approximately $6 million was for the v3 training, not for the more powerful models, so it’s an underestimate. But my impression from expert commentaries is that while we can quibble about the exact numbers, the efficiency gain and cost reduction is broadly speaking plausible.

In response, Nvidia stocks plunged 17% wiping out more than half a trillion dollarsin value and Gary Marcus declared that “OpenAI may well become the WeWork of AI”.

WeWork, as a reminder, was an American company whose business model was to lease office spaces. That’s not much of a business model you might say, but investors put a lot of money into it nevertheless. They went bankruptcy at the end of 2023. Gary Marcus has also repeatedly said that OpenAI gives him “Theranos vibes”.

Yes, we have an AI bubble in the sense that many if not most of the existing AI companies are hugely overvalued. But I don’t think the AI bubble is yet ready to burst. The most likely reason Nvidia stocks plunged is that the DeepSeek announcement further devalued the American AI companies which buy their chips from Nvidia. But if the AI training costs come down, this will open the market for many more players. I expect the stock market to recover because ultimately this is good news for technology development and the economy, not bad news.