@koraykv and I just had a wide ranging conversation at
@theinformation AI Summit in his first interview since running all of DeepMind and I can’t wait to share the video with all of you.
What struck me (and these are based on scribbles while interviewing him)
- Gemini 4 coming soon and body language suggested def before end of year. He is positive and says Google is “certainly at the frontier.”
- Doesn’t see a slowdown as necessary because says you always build safety in lockstep. I pushed him on this especially with new techniques like looped transformers and he went into the chain of thought work on some depth
- believes big in Google’s continued hardware advantage with TPUs and discussed how new generations are designed alongside his team. Says not a problem to give those to competitors like Anthropic because it helps with scale and improving
- talked a bit about the role of agents in coding and what has changed in last six months and how he thinks about RSI
- pushed back on AGI as the goal and called it not the right conversation. The best framework is continually improving intelligence.
- seemed nonplussed about distillation and happy to call out that Google invented it (he said)
- Didn’t get a clear sense how compute could or could not be a challenge. He rejected my personal hypothesis that Gemini in gmail isn’t good enough because they need more data centers :).
- a lot is made of the arms race but on narrative, I am really struck by how google, meta and Microsoft all are striking one tone and Anthropic and OpenAI a very different one. In many examples, he rejected the narrative out there in favor of a different one. People can decide whether that is good or bad for google in the long run but struck by the difference.
More to come