Second ending by classic SF writer James White, nice and quick vintage SF read about the alignment problem and self-driven machines (not so negative)

Beyond computational equivalence: the behavioral inference principle for machine consciousness Abstract. Large Language Models (LLMs) have rapidly become a central topic in AI and cognitive science, due to their unprecedented performance in a vast ar

I'd also add that we tend to forget that literally all we know about the putative structures of the representations, it is inferred from behavioural observation, so we like to claim R then B, the "historical" reality of cognitive science is always B then R doi.org/10.1093/nc/niag002

I guess what Fred meant is that you can try to explain behaviour based on the environment without postulating the mediation of internal variables. But then the law linking E to B can scientific. The matching law is a good example. It explains B from E (avoiding R), but can be falsified.

Looking forward to read this then, and yes, it would be lovely to find an opportunity to pursue this conversation further, please let's stay in touch 🙂

I have just finished reading The Technology of Teaching by Skinner it confirmed two ideas for me.
First, the image of Skinner as an “evil” behaviorist is false. Reading him directly, I find a profoundly humanist thinker.
Second, he may be more useful than expected in our AI-powered era
Few takeaways

Skinner was wrong (and often exaggerated) on many points, and Beyond Freedom and Dignity is probably too complex (and provocative) to discuss properly within Bluesky’s constraints. But my own reading of the text is less dystopian, and I hope we’ll have a chance to discuss it at some point 🙂

The mythical revolutions of American psychology The history of experimental psychology in America is typically told as a series of two Kuhnian revolutions separating three periods of normal science dominated by the mentalist, then behaviorist, and ...

The more I study the issue and the original texts, the more the scientific arguments against behaviorism (and, conversely, in favor of the cognitive “revolution”) seem weaker to me. This is one of my favorite papers across all fields (and very convincing IMHO) www.academia.edu/75630727/The...

For that reason, Skinner seems to me essential reading for anyone interested in both the ethical and the technological dimensions of education. Not because he had all the answers, but because he asked, with unusual clarity, the good questions.

Freeing teachers from routine tasks, could help focus on the more “meta” level of teaching: guiding, structuring, interpreting, and accompanying learning in its most interesting and most human dimensions.

And if, like me, you sometimes worry about what will remain of the role of teachers and pedagogues in a world of LLM-based tutors, Skinner is, here too, surprisingly reassuring. His view was not that machines should replace teachers, but that they could relieve them of routine functions.

On the technological side, many of the principles Skinner wanted teaching machines to implement, active responding, immediate feedback, adaptive progression, individualized pacing, now look far more feasible in an era of LLM-based tutoring. The technology was not ready then, but it may be now.

Better teaching, for Skinner, meant less coercion, less failure, less dependence on the lucky match between a “good” student and a poorly designed system. It also meant taking seriously the idea that educational institutions should help more people learn well, rather than merely sort them.

He also saw very clearly that one of the greatest failures of education is its inability to accommodate inter-individual differences without turning them into mechanisms of humiliation or exclusion. In that sense, his project was not cold or mechanistic. It was ethical and democratic.

Many ideas now presented as contemporary were anticipated there. Skinner argued that learning should be individualized, that students should progress at their own pace, and that punishment should be banished from the classroom because it produces avoidance and distress rather than genuine learning.

Skinner’s central idea is that teaching should be treated as a domain in which we can deliberately design better learning environments. He develops the notions of a behavioral technology of education and of teaching machines as tools for organizing feedback, pacing, and practice more effectively.

I have just finished reading The Technology of Teaching by Skinner it confirmed two ideas for me.
First, the image of Skinner as an “evil” behaviorist is false. Reading him directly, I find a profoundly humanist thinker.
Second, he may be more useful than expected in our AI-powered era
Few takeaways

📢 New preprint📜 out 🚨:
"Biased processing of multiple outcomes in human reinforcement learning: evidence from computational modeling and eye-tracking"

How do we integrate multiple outcomes that stem for a single decision?

🔗 osf.io/preprints/ps...

Too lazy to read? I made a summary of the thread👇

New preprint (by @henrivdd.bsky.social et al.)

In everyday life, choices often lead to multiple simultaneous outcomes — some positive, some negative. Yet most reinforcement learning research has focused on situations where each choice produces only a single outcome 1/5

most co-authors (Henri, Charlotte, Antonis, Camille) are not on bsky, except @mael-lebreton.bsky.social
This is part of a new line of research in team team about the attentional mechanisms underlying learning biases, such as @romanececchi.bsky.social's paper europepmc.org/article/ppr/...

Eye-tracking revealed a matching attentional pattern: participants looked more at positive outcomes, even when those were not the ones ultimately selected for payoff. A second experiment with complete feedback replicated the main findings. actuall 5/6 👇

Computational modelling showed a double bias:
• asymmetric updating (learning more from positive than negative prediction errors)
• asymmetric integration (overweighting gains over losses when combining outcomes) 4/5

Behaviourally, participants used information from both outcomes. But when outcomes were discordant, positive outcomes had a stronger impact on subsequent choices than negative ones. 3/5

In this new study, we asked: how do humans learn when one decision produces several outcomes at once? To test this, we designed a novel RL task in which each choice could generate two outcomes, both informative, but only one counted for final payoff. 2/5

New preprint (by Vandendriessche et al.)

In everyday life, choices often lead to multiple simultaneous outcomes — some positive, some negative. Yet most reinforcement learning research has focused on situations where each choice produces only a single outcome 1/5

osf.io/preprints/ps...

Using simulations, we investigated whether Value-Shaping imitation was always a good thing. And it's not (or is it ?? *gasp*) 🧵

However, in complex or volatile environments, especially with reciprocal imitation, VS can become maladaptive.

It leads to rigidity, poor adaptation, and even polarization / echo-chamber-like dynamics.

We then propose conditional Value-Shaping, combining imitation with performance monitoring.

We focus on Value-Shaping (VS) and Decision-biasing (another plausible form of imitaiton): a form of imitation where others’ actions directly modify one’s internal value representations.

Result: in simple, stable environments, VS outperforms the others well, especially when others are skilled.

📄 New preprint: The blessing and curse of Value-Shaping imitation
(by @isabellehoxha.bsky.social)
osf.io/preprints/ps...
Imitation is central to human learning: but not all imitation processes are equally adaptive. We study their computational properties using reinforcement learning models.

📘 Excited to share that Decision Making: A Very Short Introduction (OUP) is now available online (PDF for subscribers):
doi.org/10.1093/9780...

What is it about?

Understanding how humans (and other agents) make choices.
Below a bit more information👇