🚀 Rocket Is Incentive Compatible

I am losing the fight. If anyone has a space near a BART stop they could volunteer for a couple hours once a week, hmu

ChatGPT moment for robotics is not just general purpose model. It’s also ease of use, which is equally important but overlooked

You could consistently hold both that The longrun equilibrium for robotics is general purpose models + task-specific hardware AND The optimal strategy for a robotics company right now is off the shelf hardware + specialized model

^ some of you, I fear :/

I WENT TO SILICON VALLEY and all I got was this lousy intuition for when the bitter lesson applies

@michaelfroid 👀

@rocketalignment Machiavelli wrote the Prince in a pretty apocalyptic context, interesting stuff to look into if you’re interested

Ominous combo

But when it’s 1) not naturalistic and 2) poorly targeted it just gets in the way Especially annoying when you’re coming in with a bunch of interdependent questions and you already sequenced them in an intuitive way. Like I’m asking the questions here smh

Like think about how people manage this in normal conversations - “…which is something I can go more into if you’re curious” “there’s a lot more we could say about that” “do you already know about…”

hot take - I think this feature actually has potential to be rly useful Like I DO want conversation partners to surface my unknown unknowns and this is a natural way to do it The problem is that the quality is low, so 1/10 times it’s insightful and the other 9/10 it’s annoying

From the amicus brief they submitted in support of Anthropic today: theinformation.com/briefings/goog…

OAI/Google employees: "A child’s tricycle can physically be driven on an interstate, but we do not allow it because of the risks of using the technology in that environment. Mass domestic surveillance and autonomous lethal weapons systems are the equivalently reckless domain"

But the DoD strengthened its legal hand by narrowing the scope of the SCR designation and using 2 redundant legal authorities

@UrbanAstroFella Researchers in the military do mechinterp?

Remember, this is the ex-Uber "deal guy" with do/say anything to win. He knows researchers in the military perform mechinterp tests and there are all sorts of hoops to jump through for certification in classified mil environments. I feel a little sick at the disingenuity and how many people are gobbling it up. Anthropic is highly transparent when it comes to mechinterp.

This is... not a remotely accurate description of what the 2023 Al executive order did? Undersecretary Emil Michael: "If you remember the Biden executive order on Al, which was this crazy executive order that limited the amount of compute any model company could do and was essentially grandfathering in a small number of ai companies that they were gonna designate as the winners, and everyone else was out" Its not true that the EO limited the compute that AI companies could do. What it did do was require companies who were training models above a certain very high compute threshold (10^26 FLOP or 10^23 FLOP for models trained primarily on biological sequence data) to notify the government and share what testing and red teaming they were doing for certain national security risks. People are free to dislike the Biden AI EO! But it seems good to factually describe what the policy said.

SCOOP: Dario Amodei sent an explosive memo to Anthropic employees on Friday, calling OpenAI's deal "safety theater" and criticizing the Pentagon, Palantir, OpenAI and the Trump administration writ large. theinformation.com/articles/anthr… with @steph_palazzolo and @srimuppidi

Palantir could be one of the next dominos after some fed agencies dropped Claude yesterday

Gemini citing itself

Confused about the Anthropic / DoW developments? Here's your reading list: (1) @rocketalignment (@theinformation) - the guy has more scoops than someone at Baskin-Robbins. And, of course, @axios - similarly serving up exclusive analysis at a rapid pace. (2) Lt. Gen. Jack Shanahan - he's writing mainly on LinkedIn but I can personally testify to his thorough and thoughtful commentary on these issues. (3) Some mix of @ARozenshtein (@lawfare), @neil_chilson (@abundanceinst), and @CharlieBul58993 (@law_ai_). They were writing about the DPA when you thought it stood for Department of Personnel Authority or something like that. (4) @deanwball (@JoinFAI) because he's been in the room where it happens and is always willing to cut through hype and vibes to share cutting analysis. (5) If reading isn't your thing, then subscribe to @scaling_laws.

.@rocketalignment (and me) with the answers you want to know: what is a supply chain risk? what does this MEAN? is this LEGAL? come for the answers to those questions, stay for this banger kicker quote: theinformation.com/articles/next-…

@peterbarnett_ @DmitryRybin1 I would read its substack too tbh

@DmitryRybin1 Did 3 0 ask for a Substack? 0 2

@herbiebradley Definitely 👍

@rocketalignment I don't think we're anywhere close to saturating robotics pretraining though, it's very early See Nvidia releases this week

As robots move into deployment, quality of posttraining data > volume of pretraining data Similar to trajectory for LLMs and AVs