Ben Lansdell

Machine learning engineer. Also computational neuroscientist, mathematician, cyclist, dad.

[Ben Lansdell @benlansdell]

@ugisu You're welcome!

[Ben Lansdell @benlansdell]

@ID_AA_Carmack Indeed -- causal inference lies at the heart of that problem, and at the heart of understanding the world and our place in it: https://t.co/uAJjKIjSyi

[prof-g @prof_g]

https://t.co/tjd4WBMuHv

[Ben Lansdell @benlansdell]

@yudapearl @NandoDF While there's no structural causal model or anything, the claim is such agents would be performing causal inference. Perhaps not in a formal sense, but more in the sense studied by developmental psychologists and animal behaviorists.

[Ben Lansdell @benlansdell]

@yudapearl @NandoDF The Q it asks is how far can 'standard' RL frameworks and algorithms get at solving tasks in environments in which, while not known to the agent, there is a separation between observation and interaction, and solving the tasks does require transferring bw obs and interaction.

[Ben Lansdell @benlansdell]

@NandoDF @yudapearl An RL algorithm that can transfer what it learns from observation to interaction, and vice versa, can more reasonably be argued to be capable of causal reasoning. Even just at the level 2 rung. Shameless plug for a workshop paper I wrote on this: https://t.co/3kr1DDBpQ4

[Ben Lansdell @benlansdell]

@NandoDF @yudapearl On your second question, another way of phrasing it may be the need to be able to distinguish between actions that are in some sense passive observations of the world and actions that are interactions. In some way this may be separation of the environment into self and non-self.

[Ben Lansdell @benlansdell]

@littmath @2the_dark_byte_ I spent a long time trying to get it to give me the right answer. It refused to say Bob

[Ben Lansdell @benlansdell]

@Sara_Imari I think a paraphrase of the last clause is: we've evolved to have a sense for 'mathematical' patterns in reality and an ability to form abstractions. Depth in time = evolved long ago to the point of being subconscious abilities for us. But I'm not sure... What is depth in time?

[Ben Lansdell @benlansdell]

@Sara_Imari Crystals?

[Ben Lansdell @benlansdell]

@madeofmistak3 The belief you have agency.

[Ben Lansdell @benlansdell]

@lpachter @littmath Just as much as I would believe any other exact, ordered possible outcome.

[Ben Lansdell @benlansdell]

@Sara_Imari Computing existed before Turing. Computers are interesting because they represent (and can manipulate) quantities we're interested in. Biological systems very likely represent (and manipulate) features of their environment.

[Ben Lansdell @benlansdell]

@PhilosopherJoeC But no grounding in science renders philosophy useless. An evolving understanding of the world is the best we have.

[Il Memming Park @memming]

My brilliant trainee Piotr Sokol (@MannyDePresso) PhD is on the job market! He's an expert in machine learning, differential geometry, dynamical systems / stochastic processes, and lover of manifolds and algebra.

[Ben Lansdell @benlansdell]

@drmichaellevin Making things happen, by Woodward

[Melanie Mitchell @MelMitchell1]

Tour-de-force discussion from @raphaelmilliere on what it means for a video generator (like Sora) to have a "world model". Long, but definitely worth reading.

https://t.co/PxXEgqgj1X

[Ben Lansdell @benlansdell]

@bayesianboy @ylecun Indeed. Where did he get that idea from?

[Anil Seth @anilkseth]

1/ IIT can and should be critiqued, but this letter, despite many admirable signatories (& colleagues/friends) is disappointing. The accusation of pseudoscience is serious & clickbaity & IMO the letter doesn't justify it https://t.co/W9n6XkZz3S