3 remarkable things that happened in AI this week: money, Colorado, and the least socially-useful startup ever
Google sort of says it’s going to invest more than $100 billion in AI
What happened: Denis Hassabis, the CEO of DeepMind (Google’s AI business), said that Google’s investment in AI would likely be more than $100 billion, as initially reported by Bloomberg. The comment was made (in a somewhat offhand manner) in response to a question about a rumored $100 billion computer (“Stargate”) under development from Microsoft and OpenAI; Hassabis didn’t actually give a specific number, but said that he thought Google’s investment would exceed Stargate’s reported price tag.
Why it matters: Hassabis was simply answering a question he’d been asked, and I suspect he didn’t anticipate how much coverage his comment would receive. So I wouldn’t view this as Google putting down a marker in the AI investment wars, especially because we don’t know what time frame Hassabis was thinking about. But it’s still a revealing answer: $100 billion is greater than Google’s net profit last year, and equivalent to about one-third of its revenue. To put that number in perspective, by my back-of-the-envelope estimates, that would be the equivalent of Walmart saying that it needed to demolish, rebuild, and restock every single store they owned in the US (4600 stores x maybe $40–50m per store = c. one-third of $600b revenue). It’s also three times as much as the Manhattan Project cost in today’s dollars. If Google is really going to spend $100b on AI, and it expects a reasonable return on investment, then 1) it’s going to suppress Google’s profit and cashflow for years, and 2) the company expects a significant increase in scale and revenue on the other side of that.
In a way, this is the latest after-echo of Sam Altman’s rumored $5–7 trillion investment that I wrote about in February. It’s not clear if that number was ever real, but it created plenty of noise. The leaks about Stargate in March might have been a deliberate attempt to put a more realistic number in play (although $100b is just as significant a number for Microsoft as it is for Google), and now, intentionally or not, Google has suggested it will match or exceed that spend.
More broadly, with $100 billion here, $100 billion there, pretty soon you’re talking real money. If Microsoft, Google, and perhaps Meta are going to keep spending at this rate, then the future of AI looks like it will continue to be centralized in the big tech companies. This has implications for the future of the economy and even geopolitical implications, given the US-centricity of big tech.
Colorado gets ahead of a problem
What happened: Colorado amended the state’s 2021 Privacy Act to expand the definition of sensitive data to include biological data. The definition of biological data isn’t entirely clear, so most reporting just quoted the (long) text used in the bill, but it clearly covers all data that would be produced in any kind of medical testing or procedure, and it might include data that could be derived from video footage (such as facial-recognition data, height, or someone’s movements around a public space). The bill explicitly calls out neural data, which it probably doesn’t need to, but is clearly a focus for the bill’s sponsors, who seem to have devices like the Neuralink in mind (so to speak). The change means that companies seeking to collect biological data from consumers must first seek consent.
Why it matters: Historically, the costs of obtaining, storing, and processing information were so high that we only needed to protect personal data in a few specific instances, such as financial and medical records. As these costs have fallen dramatically over the last few decades, our ability to gather and use personal data has moved much faster than the regulatory and ethical frameworks governing its collection and use. Often, governments don’t take action until problems have already started to emerge; while Colorado’s bill only addresses one specific issue, it’s good to see a government addressing an emerging space with a simple solution.
At last, we can use AI to buy crap we don’t need
What happened: Diddo, a startup that the Internet tells me is based on a common misspelling of “ditto,” announced that it had raised $2.8m in seed funding. Diddo uses machine vision to identify products in television shows and let viewers buy them, solving the age-old problem Americans face when trying to buy stuff. Diddo the startup has no connection to the artist Diddo who made a life-size human skull out of cocaine.
A screenshot of the product suggests that the purchase links are created manually by Diddo (or by someone), which raises the question of how much effort the machine vision is saving you, if a human still needs to figure out where you can buy that Lululemon Wunder Train Tight With Pockets. But, I mean, AI.
Why it matters: Good Lord, it doesn’t.
3 remarkable things is a more-or-less weekly roundup of interesting events from the AI community over the past more-or-less week
