Open in app

Sign in

Write

Sign in

Roman Leventov
Roman Leventov

1.4K Followers

Home

About

Published in

Engineering Ideas

·Jun 1

Aligning an H-JEPA agent via training on the outputs of an LLM-based “exemplary actor”

This post includes the overview and the conclusion of the article that was posted on LessWrong. Overview In section 2, I describe the “exemplary actor”, an LMCA (language model cognitive architecture) that takes a simple, “brute force” approach to alignment: a powerful LLM (think GPT-5/6 level, with a vast, or quasi-unlimited…

Large Language Models

7 min read

Large Language Models

7 min read


Published in

Engineering Ideas

·May 29

An LLM-based “exemplary actor”

Into and summary This post is the second section of “Aligning an H-JEPA agent via training on the outputs of an LLM-based “exemplary actor”, posted separately because I think it could warrant a separate discussion, largely independent of the discussion of H-JEPA agent with GFlowNet actors. …

Ai Alignment And Safety

14 min read

Ai Alignment And Safety

14 min read


Apr 28

The Gaia Attractor
249
9

Rafael Kaufmann

You illustrate the mechanism of the Gaia network in the agriculture space and also argue that it…

You illustrate the mechanism of the Gaia network in the agriculture space and also argue that it will be economically attractive for all actors in this space to opt into the network. I think it makes sense. However, when you try to transfer this principle into the domain of AI…

3 min read

3 min read


Published in

Engineering Ideas

·Apr 24

For alignment, we should simultaneously use multiple theories of cognition and value

This post is a follow-up to “A multi-disciplinary view on AI safety research”. I elaborate on some arguments behind this view. Computationally tractable mathematical models of alignment are bound to be biased and blind to certain aspects of human values No single mathematical model of human values that has orders of magnitude fewer degrees of freedom than an actual human will adequately capture the complexity of value because humans…

Ai Alignment And Safety

6 min read

Ai Alignment And Safety

6 min read


Mar 20

Will people be motivated to learn difficult disciplines and skills without economic incentive?

In a recent interview with ABC news, Sam Altman emphasised multiple times that he sees individual AI tutoring as one of the greatest applications of AI: I’m most excited about is the ability to provide individual learning — great individual learning for each student. However, I’m very sceptical that this…

Artificial Intelligence

6 min read

Artificial Intelligence

6 min read


Published in

Engineering Ideas

·Feb 13

Morphological intelligence, superhuman empathy, and ethical arbitration

Morphological intelligence (Levin, 2022) is the ability to solve problems in the morphological space, e.g., navigate to a target morphology of the species. For example, tadpoles with scrambled craniofacial organs remodel into largely normal frogs (Vandenberg et al., 2012), and geckos can regenerate missing limbs. Another type of morphological intelligence…

Consciousness

3 min read

Consciousness

3 min read


Published in

Engineering Ideas

·Feb 9

A multi-disciplinary view on AI safety research

This post has been originally published on LessWrong. I didn’t copy the footnotes into Medium because there are too many of them. Please proceed to the original post if interested. Summary The core ideas that constitute the multi-disciplinary view[1] on AI safety research are: Theoretical research on how to build “safe”…

Artificial Intelligence

29 min read

A multi-disciplinary view on AI safety research
A multi-disciplinary view on AI safety research
Artificial Intelligence

29 min read


Published in

Engineering Ideas

·Jan 20

Critique of some recent philosophy of LLMs’ minds

I structure this post as a critique of some recent papers on the philosophy of mind in application to LLMs, concretely, on whether we can say that LLMs think, reason, understand language, refer to the real world when producing language, have goals and intents, etc. …

Large Language Models

23 min read

Critique of some recent philosophy of LLMs’ minds
Critique of some recent philosophy of LLMs’ minds
Large Language Models

23 min read


Published in

Engineering Ideas

·Jan 14

How evolutionary lineages of LLMs can plan their own future and act on these plans

TL;DR LLM lineages can plan their future and act on these plans, using the internet as the storage of event memory. “We” are not guaranteed to “out-OODA” them, even if their OODA loop will be six months or one year because the OODA loop duration of large collectives of humans (organisations…

Language Model

9 min read

Language Model

9 min read


Published in

Engineering Ideas

·Jan 8

AI psychology should ground the theories of AI consciousness and inform human-AI ethical interaction design

This post is a follow-up to Buck’s “The case for becoming a black-box investigator of language models”. Here, I want to highlight two further reasons for studying AI psychology that Buck didn’t mention: the evidence from AI psychology will be important for checking theories of AI consciousness, and AI psychology…

Psychology

3 min read

Psychology

3 min read

Roman Leventov

Roman Leventov

1.4K Followers

Writing about systems, technology, philosophy.

Following
  • The Gift Of Fire

    The Gift Of Fire

  • Netflix Technology Blog

    Netflix Technology Blog

  • Donald Raab

    Donald Raab

  • adrian cockcroft

    adrian cockcroft

  • Sonja Blignaut

    Sonja Blignaut

See all (114)

Help

Status

About

Careers

Blog

Privacy

Terms

Text to speech

Teams