An opinion piece expressing skepticism about Discovery Loop, an AI research startup founded by former Google AI leaders, arguing that automating ML research differs fundamentally from automating scientific/engineering tasks and questioning their unique advantage while acknowledging they deservedly raised funding.
I got curious to read what exactly is being offered by these leading minds behind Google's AI development who recently left the company to build their own startup, Discovery Loop. And what I saw surprised me a little. In ML and AI they're simply the industry's top people. Of course, there are others at that level, meaning they're not unique, but they're broadly a one-off product, roughly speaking. But the idea they're formulating looks to me like two parts. One part is – we automate research in AI (totally clear). The second is those 10-ish tasks from science and engineering that they outline for the future. The difference between solving the first problem and the second class of problems is, in my opinion, incredibly critical. It's fundamentally different. In one case you really can formulate the task of ML research/training/loop and automate it within your own big infrastructure, but there are nuances. This is exactly the same task that Anthropic, and OpenAI, and many others are trying to automate for themselves. And I honestly don't see this team's advantage in this particular direction. As their expertise they listed something like 50 different top projects inside AI and ML, but nothing outside of that. And for each of the 10 tasks they outline for the future, you need access to real instruments, labs, robots, planning measurements, processing sensor data, safety and reproducibility checks, domain-specific success criteria, expertise and so on. I'm not saying it's impossible. I mean that the very structure of closing the loop and the bottleneck in that loop – that's usually what industry startups present when they talk about automation. One example I recently came across is the London startup Automata, which automates lab experiments. And in fact the main bottleneck in this kind of work is hardware, lab instrumentation, robotization in regards to automation. Closing the AI loop there actually turns out to be fairly primitive. That's why I'm a little skeptical that such loud claims will in the end turn into something close to their expectations. I could of course be wrong, but that's roughly how it all looks to me. On the other hand, do these people deservedly get to raise hundreds of millions of dollars (judging by the leaks I found)? Absolutely deservedly. And I think they can bring something new to the industry overall. Among other things because, as far as I remember, Google has had some problems lately related to exactly how teams are managed there, and Gemini isn't doing great in agentic coding. It's quite possible they'll deliver some cool milestone. A couple of screenshots in comments
Top Google AI researchers Jeff Dean, Sanjay Ghemawat, Oriol Vinyals, and Quoc Le are leaving Google to found Discovery Loop, a startup focused on automating scientific and engineering loops to advance AI and other fields.
Jeff Dean and several other top Google AI researchers are leaving to launch Discovery Loop, a public benefit startup aimed at using AI to automate and massively scale scientific experimentation, with backing from Alphabet and major VC firms.
Jeff Dean and longtime collaborators announce the founding of Discovery Loop, a Public Benefit Corporation aimed at automating scientific and engineering discovery with AI, after leaving Google.
Jeff Dean, Google's chief scientist, is leaving after 27 years to co-found Discovery Loop, an AI research company aiming to automate the research cycle. The startup has seed funding from Radical Ventures and Khosla Ventures, with Alphabet as a founding investor and cloud partner.