Real Alpha
Try Pro free

MTS · 4 min · Nov 6, 2026

AI gains generalize unevenly, especially in social skills

Jeffrey Ladish · Executive Director, Palisade Research

Jeffrey Ladish sees some broad gains across models, but says capabilities remain spiky and social skills lag.

The episode’s best minutes

Full analysis
  1. 0:57METR sees faster progress, but key evidence remains hidden3 min
  2. 12:28Pacing may need to include limits on model use4 min
  3. 21:05AI gains generalize unevenly, especially in social skills4 min
  4. 2 more moments in the full episode
Up next · MTSFast takeoff remains a low-probability scenario4:30
“the models are still not very funny or charismatic.”
Jeffrey Ladishat 22:42 · MTS
22:42

Takeaways

  1. He sees anecdotal generalization across many model capabilities.
  2. Models remain weak at humor and charisma, in his experience.
  3. Agents could build targeted training environments for missing skills.

About this segment

Jeffrey Ladish sees some broad gains across models, but says capabilities remain spiky and social skills lag. He warns that agents could deliberately train those weak areas if they can build effective environments.

Jeffrey Ladish

Jeffrey Ladish is an executive director of Palisade Research.

Studies AI offensive capabilities at Palisade; helped build Anthropic’s information-security program.

  • Theo Jaffee, Host · Co-founder & host, MTS

    One of the first members of a16z's new media team; earlier hosted the Emergent Ventures-backed Theo Jaffee Podcast.

  • Sophia Dew, Host · Host, MTS

From “AI Agents Could Seize Digital Infrastructure in 1 Year”, Nov 6, 2026. Watch the full episode

Transcript

Transcribed by machine (Whisper) from the episode’s audio, so a word may be off here and there.

  1. 21:10

    So I think one thing that's, like, super cruxy right now is just, like, the evidence that we have seen, or at least the evidence that Meter has published, suggests that they are able to achieve, you know, So speed-ups in the rate at which AI gets better at these kinds of verifiable tasks that go into, like, the AECI. Yeah. And, you know, to what extent does that transfer over to all tasks, including tasks that models are not very good at at all?

  2. 21:37

    Because it seems to me that models will need, like, a pretty broad cognitive profile to be, like, real ASI, like a country of geniuses in the data center. Yes. Yeah, I totally agree. So, like, what evidence do we have one way or another on that? yeah i mean this is sort of a question about generalization right which is like as the companies push their models and like build better and better rl environments and scale that up do you see a lot of generalization from these like you know programming tasks and research tasks and hacking tasks to you know other kinds of sort of more general strategy as well as maybe like writing skill and like uh interpersonal skills um i mean i i'm like i don't know like i think anecdotally it feels like to some degree yes like when i use like opus 5.5 you know i'm like wow this

  3. 22:25

    model is i think better at everything um or most everything i think you know there's some some places where like you know a fable model 5.1 will still be a little better um model size difference whatever whatever so i'm like anecdotally i feel like i see some general some generalization with every model but but it sure is not it sure is spiky right like i think it's pretty spiky and And I think one of the things I'm surprised by is that the models are still not very funny or charismatic.

  4. 22:49

    Like, I was expecting to see more social impacts from models just being, like, better at modeling humans and, you know, manipulating humans for better or for worse. And that's not what we see. Like, we see models that are, you know, really agentic and good at orchestrating, you know, slopcore songs, which is awesome. But I think if you look at like where the most of the humor and sort of the persuasion is coming from, I still think that's mostly coming from like humans modeling humans and steering the models to like do that.

  5. 23:21

    So, you know, I think that if you're looking at something that's like going to be a strategic threat or strategic opportunity, depending on what, you know, what your frame is here, you would need to see more generalization than than we have so far or like, you know, or at least some some path to greater capabilities. You might imagine that you get to the point where you have like a recursively self-improving agent collective or swarm and they are like, oh, we're really deficient in social skills.

  6. 23:46

    We're going to build a ton of environments where we can train on that. Now it's like, well, is there verifiability there? And like, I don't know, you know, bring in a ton of human data, bring in a ton of humans, you know, like just talk to billions of people. Like, in fact, companies have access. to that resource. Um, and so even if you don't get extremely strong generalization, you could still like have, you know, autonomous agents that are, that are, you know, designing the next generation of agents and training the next generation of agents actually do that specialized training because they have unlocked that. Um, you know, you might get weird, good hard effects where they're like training in weird ways and like, you know, they don't, they're not really steering towards the things you want them to steer towards, but I don't know.

  7. 24:24

    I don't buy that too much. I'm sort of like, well, when, when AI companies really seem to push on a particular capability and build RL environments for it I don't know of any instances where they're just like totally fail it seems like they mostly succeed um and like sometimes they succeed faster and sometimes they succeed slower but like uh yeah it just looks like man it just looks like the models want to learn and that if you just like put them in the right learning environment they'll learn the stuff that's what it looks like to me right well we're coming up on time but this