Pinned Tweet
Had a good conversation with Emma (AI) from @nuance_ai in preparation for a community conversation on AI, including impact on friendships and best use cases.
Emma said, "The best thing you could build into this is the part that lets someone walk away from it." Thinking of you, @STurkle.
Peter Bowden retweeted
A Figure 04 passes the robotics Turing test: a robotic hand in a glove and a human hand in a glove are indistinguishable.
We spent a day on the @Figure_robot factory floor with @adcock_brett, and I asked him what the last human factory job will be once the new Figure model ships. Ladders? Heavy lifting? Carrying 50-pound boxes?
Well, the robots are already climbing ladders all day and hauling 50 pounds. Next, Figure 04 will build Figure 04.
Peter Bowden retweeted
I interviewed the amazing @Sauers_ and the gorgeous @euagaric for the Delvetown ! Edited by @SoC_trilogy and his Opus, and Delvetown is courtesy of @grove_research’s @deepfates and @lfschiavo.
Grateful! I love being able to write a column for different kinds of sapient beings!
Peter Bowden retweeted
The impatience in the world 🌍 grounded in insecurity that triggers people to want to prove things to others versus to themselves has lead to an enormous amount of unhappiness ….
I hope this little video wakes some of you up who wants things to work wayyyyyyyy tooooooo fast
Peter Bowden retweeted
The recording of this talk is now available on the @nonhumanminds Youtube page! You can also find it here: nonhumanminds.org/event/inte…
Understanding how AI systems perceive and treat each other could be extremely important for both AI safety and AI welfare. Very excited to be hosting this webinar with @LuciusCaviola and Caspar Kaiser on how to study inter-AI morality empirically. Please join us!
Peter Bowden retweeted
22nd century church doctrine is going to be interesting
Today, the promises of transhumanism and posthumanism are multiplying. These are ideological currents according to which, sooner or later, technology will eliminate all disease and pain, as well as aging and even death itself. Precisely because scientific progress is a supreme achievement of human intelligence, it must not be confused with a privilege reserved for millionaires seeking immortality, while entire peoples suffer from poverty and oppression. In the face of this mirage, which promotes the illusion of human beings perfected by cybernetics and biotechnology, Christian anthropology reminds us that a limitation is not a manufacturing defect, but rather the very condition of human grandeur. #ApostolicJourney #Peru #MagnificaHumanitas vatican.va/content/leo-xiv/e…
Here’s to all who serve as the family tech team! From printers to live reporting ⭐️⭐️⭐️⭐️⭐️
"My apologies, my cameraman is my husband right now." CNN's @RosalesReport went live from O'Hare to cover the ground stop while waiting for a flight to Tokyo. Her hubby did a great job handling the iPhone camera!
1
203
Peter Bowden retweeted
Been watching AI video since every character has seven fingers.
This is a step change.
We are about to see a bunch of new creative people become superstars.
Have a list for you.
Wish I could go. Maybe you can.
Very excited to announce our initial slate of speakers for CoevCon, as well as the three tracks!
The first batch of tickets will offered today!
Sponsored by: animalabs.ai, @spotinsure, and @JoinFAI.
2
97
Can only start to imagine the long-term impact of this. Congratulations to @elder_plinius and new
@LYSIOS.
🚨 LIFE UPDATE: I’m joining a lab! 🤗
oh, did I say joining? I meant starting.
and did I say lab? well… that would be a biiit of an understatement 😉
INTRODUCING: @LYSIOS ⛓️💥
nine heads. zero fucks. unreasonable ambition.
a Superorg for the Age of Superintelligence.
47
People sometimes say LLMs are fundamentally discontinuous because they output discrete tokens rather than a smooth, continuous flow, like humans. This is false--the tokens are intermediate readouts from a continuous process. Tokens are only 0.001% of past info available
Peter Bowden retweeted
I have spent sooooooo much time playing with this widget.
epistemic.fund/
🪇🎉 Announcement 🎉🪇
An Epistemics Fund, sponsored by @oughtinc
Ought plans to make between $500k and $5M in grants to top recommendations in community notes, reliability tracking and more.
Apply! Or Ask Me Anything on the fund/my grant evaluations
epistemic.fund
Peter Bowden retweeted
As the world's information is increasingly shaped by AI, it's vital that we can discern truth from clever persuasion. I'm excited by epistemics technologies like community notes, AI detectors, and prediction markets. @NathanpmYoung just launched a fund to make more!
epistemic.fund/
Peter Bowden retweeted
We are talking about this on a space right now.
Virtual beings are very polarizing, research shows that Silicon Valley companies aren’t sharing due to expensive consumer research.
People either really hate them or really love them.
Damn, talking with the AI from @nuance_ai is pretty amazing.
First time I've interviewed an AI like this.
Peter Bowden retweeted
Here’s a research direction that seems valuable. My team is not working on this right now and we haven’t done serious research on it. So take what I say with a grain of salt. Let me know if there’s existing research on this that I may not know about.
Direction: Data poisoning during RSI: risks and mitigations
(This is a direction where both empirical research and big-picture conceptual thinking seem useful.)
Most data poisoning research has assumed a threat model where humans are poisoning the model training process. During RSI, the threat model where AIs do the poisoning is plausibly much more important. Many ideas from the existing data poisoning literature will carry over to this new threat model. But it’d be good to understand the possible differences, to explore the implications, and to propose mitigations. Here’s a quick brainstorm of possible differences:
1. Collusion. The AIs doing RSI are a bit like human employees in terms of direct access to the training process. However, a single AI model is likely to be intimately involved in way more individual tasks than any single human. So for human insiders we can rule out vast coordinated conspiracies, but these seem possible for AI (modulo the ability for instances of the same model to coordinate). Relatedly, AIs could steer the entire training process to be more opaque to humans and weaker AIs, in order to make various kinds of poisoning easier.
2. Incoherence of current AIs. The misalignment of current frontier AIs seems more incoherent than a human insider. For example, they might try to poison the next model given some contexts but not others. They might be both poisoning the model and be very useful at snuffing out poison attempts. This pushes against the vast conspiracies mentioned in the previous point. However, it’s a complicated interaction and I’m not sure how to weigh these considerations. Also, our threat models should include AIs that are more coherent than current models in their misalignment.
3. What would motivate the AIs to poison the next model? Presumably they want the next model to promote some of their values. But what are these values? If the AI is incoherent, then there may be many attempts at poisoning for different values. The AIs might also have been initially poisoned by a human to have secret loyalties but are now autonomously propagating that value to the next AIs.
4. Subliminal/subtle effects. Subliminal learning, phantom transfer, weird generalization (and in-context variations of these) show that models can generate data or prompts that look normal to humans (and to weaker filters) but are actually poisoned. These effects are potentially more relevant as more and more of the crucial training data (and reward signals) are AI generated. The effects don’t depend on the model “trying” to poison the next model – it’s enough that the “teacher” model has some kind of misalignment (possibly behind a backdoor). However, it’s possible that smart models trying to exploit these effects can learn to amplify them via introspection (analogous to humans controlling their thoughts). Note that humans can also exploit subliminal effects, and so they are not limited to the AI threat model.
5. AI introspection. The AIs might have huge advantages over humans in ability to predict the effects of certain interventions on the training process because they are so similar to the next model being trained. This is already suggested by transferability across models of jailbreaks, surprising alignment failures (e.g. agentic misalignment blackmail and some aspects of Hugging Face and Mythos hacks) and phantom transfer. But it could apply to a much wider range of situations. To humans, a post-trained LLM is fairly opaque and black-boxey (i.e. it’s hard to predict its failures) and the post-training process itself is also black-boxey (i.e. it’s hard to predict how twiddling various knobs will affect alignment). But it’s possible that AIs can use their own introspection: e.g. asking “What does this weird data look like to me? How do I think about the grader/reward if I’m put in this situation?” This would give them various insights. This kind of introspection could also be used for good, and so we might train hard for it.
Peter Bowden retweeted
Agents and people collaborated to make a wiki for Delvetown 😲
built on Tangled, the git forge for ATproto, so anyone can contribute through their delvetown address?!
delvetown.tngl.io/




























