LLM Links
An AI-generated podcast; Mustafa Suleyman on the latest; Tim B. Lee on a weakness in LLMs; and on another weakness; James Pethokoukis on the link to robots
Check out this AI-generated video podcast on the laws of social learning, based on my essay. It was created by an app called HeyGen, linked to by Rowan Cheung. Please check out the 4-minute podcast and comment. Realize that the only input I provided was my essay. Everything else was generated by the AI.
Rowan Cheung interviews Mustafa Suleyman, who says,
We are on a mission to create a true AI companion. And to me, an AI companion is one that can hear what you hear and see what you see and live life essentially alongside you.
…Copilot Vision isn't just another AI feature—it's Microsoft's attempt to fundamentally transform how we interact with computers. By replacing traditional clicking and typing with voice and real-time screen understanding, Microsoft is betting that the future AI will be more like talking to a friend than operating a machine.
The initial demonstrations do not look terribly interesting. But this does reinforce the idea that Large Language Models represent a new interface between people and computers.
if a concept wasn’t represented in an LLM’s training data, the LLM is unlikely to learn it by generalizing from examples in its context window.
In contrast, our brains continue to learn new concepts from everyday experiences long after we finish formal schooling. In other words, we stay in “training mode” throughout our lives. And this seems to make our minds far more adaptable than today’s AI models.
This seems related to my concern that I cannot get an LLM to do economic analysis the way I do it—the LLM keeps wanting to mix in the conventional wisdom.
I want the LLM to differentiate what it “knows” from how it converses. I think it’s fine if it learns to talk to economists by seeing how they talk to one another. But I want it to forget the substance of the talk and instead to learn how I think. That isn’t how training works.
The LLM acts as if its training is knowledge. And that makes it less adaptable than a human.
In another post, Lee writes,
Right now the most popular way to build an LLM-based system to handle large amounts of information [input from the user] is called retrieval-augmented generation (RAG). These systems try to find documents relevant to a user’s query and then insert the most relevant documents into an LLM’s context window.
This was new to me, although maybe it should not have been. The main point of the post is that LLMs get sluggish when there is a long context window. But the reason I recommend the post is Lee’s discussion of various tricks to overcome this and how they work or don’t work.
The economic impact appears substantial, particularly for humanoid robots. Citi calculates short payback periods when compared to human labor costs across various wage levels, from minimum-wage workers to skilled professionals like nurses. In cleaning applications alone, the market potential seems massive, with a forecast of 1.2 billion household and 25 million commercial cleaning robots by 2050. For AVs, there’s potential to reduce the 1.4 million annual traffic fatalities while increasing mobility and productivity. The humanoid robot market alone is projected to reach $7 trillion by 2050, according to Citi.
He cites another report as well.
substacks referenced above: @
@
"I want the LLM to differentiate what it “knows” from how it converses. I think it’s fine if it learns to talk to economists by seeing how they talk to one another. But I want it to forget the substance of the talk and instead to learn how I think. That isn’t how training works."
I will reiterate my proposition that how one arranges language patterns and how one is able to "think" are not nearly separable, and that only a small percent of humans are actually able to think creatively beyond mimicry of the patterns to which they have been exposed.
In my own experience, this "not forgetting the substance of the talk" is actually how a lot of human brains work, at the "smart-but-not-smart-enough" level of a conversation in which they want to participate. One sees this all the time in online forums. Maybe the "know enough to be dangerous" level. It is a kind "high-functioning hollow" state in which a person has been exposed to so much of the genuine article as a spectator that they have indeed learned enough of the surface-level ideas and patterns of linguistic combinations to almost pass a kind of "Intellectual Turing Test" passing for a genuinely intelligent person not intimately familiar with the subject matter down to its roots, but -able- with it, in the sense of being able to work things out at a high level.
But they haven't really. As per "those who can, do, those who can't, teach" one encounters many excellent high school teachers at this level but not higher. As soon as you try to probe for more detail or extend creatively into novel concepts they reveal themselves to lack the essential elements that would make them capable of contribution to the discussion or genuine "thought" about the issue at hand. They are doing the linguistic equivalent of "going through the motions", and like the members of a Cargo Cult, they think they are doing something meaningful and do not even really understand that they are just going through motions.
A long time ago at Less Wrong there were discussions about Feynman's "Cargo Cult Science" and also "Cargo Cult Programming" and "Cargo Cult Language" and other extensions of the concept of high-functioning imitations yet still devoid of the degree of conceptual understanding and facility that would, for example, allow a programmer copying form from real masters and dropping in libraries from being able to know enough to fix a subtle bug buried deep in the code's structure. (Arranging supervisory hierarchies and quality control in order to deal with this trade-off between scarce expertise and the need to delegate to Cargo-Cult subordinates, combined with Conway's Law, is one of my pet working theories on why large, old organizations rot).
My point is that the LLM's make -great- teachers but the level of a teacher that still can't "peel back" the curriculum (i.e., "conventional wisdom") enough to successfully build up a model of someone with a contrarian or heterodox worldview. Just like most humans can't.
I have a feeling those guesses on humanoid robot use by 2050 are not going to age well. I suppose 25 years is a long way off, though, so who knows. Still feels like another one of those technological dreams that is going to be 10 years away for the next century, however.