This week in AI: Agents are buying your DDRAM, NVIDIA achieves AGI (again?), and Salesforce loses its head.
On to the main stories:
- Nvidia agrees to buy Hugging Face for $12.9 billion. The deal doesn't seem to be officially confirmed yet, but it's another huge ecosystem acquisition within a week (after OpenRouter by Stripe last week). This is an incredible outcome for the team, but I do wonder how the community will evolve post-acquisition. There's no reason to believe NVIDIA won't be a good steward, but even with the best intentions, incentives shift. Perhaps the best parallel is the acquisition of GitHub by Microsoft in the mid-2010s... open-source developers were aghast. However, the company turned out to be a very good steward of the community for many years and arguably made investments that otherwise would not have been made. Things have been a little unstable at times recently, but volumes have also shot through the roof. Also this week, Hugging Face released a line of microduck robots - perhaps this is what NVIDIA was really after?
- The full-length interview with Yuval Noah Harari | The Economist. Yuval Noah Harari makes some thought-provoking points in this interview about the big-picture impact of AI on humanity. I fundamentally disagree with the thesis that because "AI" could derive more complex financial systems or take control of particular processes, it will, at least not in the near future. This anthropomorphizes AI and suggests a very unlikely (in my view) "escape" scenario. His suggested solution also just seems naive: "just slow down." I agree that tech company inevitability narratives are rather disingenuous - i.e., "it's going to happen anyway, might as well be us"- who, if not the tech companies, does have control? The reason there will be no slowdown, however, is because the AI we have today (and have line of sight to) is clearly such a powerful tool that there is no putting it back in the bag. There will be open source; many teams worldwide will build models, and there will be competition for years to come to build strong models. There'll be no slowing down. This is really like suddenly having invented electricity (an electricity that is 100x more powerful and makes all devices more useful). Instead, I would interpret the challenge in another way: we must hold people accountable for the uses of AI. We will still have many decades where human agency will be upstream of AI actions. The biggest problem really is humans saying "looks good" to unverified solutions and executing them without checking. There should be consequences for these actions that create harm. Not an escape hatch saying "the AI suggested it".
- Uber faces fine of nearly $1B over automated driver suspensions. A case in point of responsibility for automated decisioning is Uber's huge fine in Europe over driver account deactivations. The complaint alleges that the company shuts down driver accounts without warning based on certain triggers. The resulting economic impact can be significant. The company counters that these are primarily due to detections of fraud. It's hard to say who exactly is right or wrong here, and there are (in my view) cases where decisions need to be quick and can't involve human review (for safety reasons, for example). However, in general, with impactful decisions like this, decision systems need significant safety mechanisms and review. I'd put them into a similar budget as credit approval algorithms used by banks. The impact may be a bit less severe than being rejected for a bank loan, but being banned from working without review is quite harmful.
- Writing with AI, a contrarian view (by Tim O'Reilly). This is a great piece by Tim O'Reilly on what writing with an LLM is (he argues - an art and expression of its own). Worth reading in full.
- Your agent can sign up for Mintlify fully on its own. Mintlify is a popular documentation platform, and this small announcement caught my eye: agent sign-up. The feature makes it possible for an agent to sign up and get access to Mintlify without human intervention (and presumably without violating the Ts&Cs). The flow still involves clicking on an email link, but the assumption is your agent has access to email. I think it's going to become more and more common to enable these flows - the web is becoming agent-driven. What's interesting, though, is that behind Mintlify's setup there is still a core assumption that your agent is signing up on your behalf, with your email address. Things would be much cleaner with something like AAuth.
Wishing you a great week.
