Gabriel Dulac-Arnold

@chris_j_paxton @xiao_ted The hardest part with ML (and particularly RL) systems is that the system is too smart for its own good, it will learn around broken contracts and missing features! I am very agressive now about typing, preflight checks, and deep observability. Good tooling==good results!

I really wish I had been there for this one with @xiao_ted -- full of insights on the insane trajectory robotics has been on over the last ~decade. What really stood out to me: The only reason something succeeds is if everything goes right. Behavior cloning, for example, seemed stuck at 60-70% success rate on key tasks until Ted's team rewrote their learning stack — at which point it hit 95-99%+ success rates. For most of those eight years, something was wrong. The stack wasn’t quite right, the learning algorithms were wrong, the data didn’t exist. Hardware and operations are not mature enough. But they kept working on these problems, over and over, until finally they have arrived at amazing breakthrough. There are probably lots of these little things that still need to be solved.

Really enjoying working with @WilliamVerst !

Like a lot of people, I have a soft spot for DeepMind. A Proper British company in roots; doing the stuff that made AI a positive aspiration (Alphafold etc) rather than the slop; Hassabis calling for international cooperation on safety even when the other frontier CEOs had stopped. Lots of very flattering media coverage e.g. the Mallaby book. And Hassabis' counterpart at Brain Jeff Dean is clearly a person of principle. But the actual actions and outcomes matter. DeepMind's commitments to their tech not being used for military and surveillance, which many of us firmly believed them on, are gone. Their leaders haven't faced the world, or even their own employees from the sounds of things, over these decisions and how they were navigated. 'Shameful' is the term used by their own employees, justifiably, from mid-level people taking real risks to do so. If you railed against OpenAI for this, GDM can't be given a soft pass. If anything it's more concerning - with Google's resources and projects their surveillance reach is potentially far greater than OAI's. We know how their leaders feel or once felt about these uses of AI. Either (a) they won't fight for it when push comes to shove, (b) fought and lost against the greater Google entity, or (c) think this is a battle worth sacrificing for other red lines to be held to. As someone worried about AGI risk, I can see the argument for (c). I see even scarier paths and decisions up ahead. But if you don't fight for your principles in 'lesser' battles, will you know how to fight, and win, when it's the time? There will always be an argument for going along this time to preserve capital for later. There might never be a clean 'now or never' line. I get the feeling a lot of good folk at GDM will feel deep unease from today forward, and will have some hard choices and sleepless nights ahead. The same is true across all of the frontier companies, in different ways and to different extents. I have no black and white answers to offer, and don't envy you your decisions. Thank you to everyone who has tried to nudge the companies in the directions you believe in, and my sympathies - at every level it is made very difficult to remain true to all your principles, and the tradeoffs are brutal. It will become more intense from here. The stakes will become higher. Maybe now is the time to go. But those of you most clearly conflicted are also the best of you - and these companies are too far along to stop without you. They will do business without you, but the how of it concerns me. We're in the late game: no easy answers from here on.

So much for futureoflife.org/open-letter/le… 🙄

I'm speechless at Google signing a deal to use our AI models for classified tasks. Frankly, it is shameful. For HR, I'm not speaking on behalf of Google but in my personal capacity, quoting public information from a well-sourced article of a reputable publication

@lgraesser3 Thanks for being part of the amazing people I got to hang out with at work!

@gabepsilon Congrats Gabe! UMA is lucky to have you. Looking forward to seeing what you build there.

Tl;dr After 10+ years pushing the boundaries of RL at @GoogleDeepMind , @Google Brain and then Gemini, I’m looping back to startup mode and joining @UMA_Robots to push the frontier of human-centric robotics. (1/4)🧵

@YohanHadji Congrats Yohan, very cool!

Meteogliders are now flying daily in Greenland!! DMI becomes the world’s first national weather service to use it operationally.

@albericlajarte @GoogleDeepMind @Google @UMA_Robots Thrilled :)

@gabepsilon @GoogleDeepMind @Google @UMA_Robots Welcome back to robotics Gabriel !

@marine_caous @GoogleDeepMind @Google @UMA_Robots Thanks!

@gabepsilon @GoogleDeepMind @Google @UMA_Robots welcome to the team!

@xiao_ted @chris_j_paxton @GoogleDeepMind @Google @UMA_Robots Thanks! Had to take the leap at some point :)

@gabepsilon @chris_j_paxton @GoogleDeepMind @Google @UMA_Robots Congrats Gabriel!

@RobKnight160836 @GoogleDeepMind @Google @UMA_Robots Indeed, childhood amazement re-lived every day!

@gabepsilon @GoogleDeepMind @Google @UMA_Robots Love those it's alive! moments too, looking forwards to changing the world with you

@psermanet @GoogleDeepMind @Google @UMA_Robots Already having lots of fun!

@gabepsilon @GoogleDeepMind @Google @UMA_Robots Amazing to have you on board @gabepsilon !

@Mickael_Chen @GoogleDeepMind @Google @UMA_Robots Gradient explosion 🤯

@gabepsilon @GoogleDeepMind @Google @UMA_Robots Welcome! Amazing journey, let's make this next step even bigger :D

@AractingiMichel @GoogleDeepMind @Google @UMA_Robots great to be working with all this new experience as well, we're all learning so much!

@gabepsilon @GoogleDeepMind @Google @UMA_Robots Welcome on board @gabepsilon ! It's quite a privilege to have someone with your experience in the team

@WilliamVerst @GoogleDeepMind @Google @UMA_Robots it's only going to get more legendary!

@gabepsilon @GoogleDeepMind @Google @UMA_Robots Legendary journey! 🫡

(4/4) Now the field has grown, the frontiers have been pushed back, and in a lot of ways, the singularity is here. But it's held back by being locked in the ether of our virtual systems. This is why I'm excited to share that I've chosen to 'leave the mothership' and jump into a new adventure at @UMA_Robots . We're building the frontier of robotics, taking on all the most challenging aspects with an excellent team, both technically but also humanly, and a clear mission: to build robots that enhance life for everyone. The world is changing quickly, and it's important that technical advances are also driven with social implications in mind. I am convinced (or I wouldn't be here) that the team at UMA, from the founders down, are strongly culturally aligned to push forward these new directions. We are rethinking fundamentally what this means for us as a society, and how we can best shape this new frontier for everyone.

(3/4) My biggest takeaway from this adventure: the people! @GoogleDeepMind was AGI-pilled before that was even a meme, and there was such an intense shared belief in the mission, and such a strong culture filter around open-minded, genuinely curious and nice people. The early days was my personal Hogwarts, with crazy adventures, 'oh-my-god-it's-alive' moments, and life-long friendships forged in breaking what humans considered was the boundaries of the possible at the time (Go was taught as 'unsolvable by computers' up until 2015!).