Blake Richards @tyrellturing.bsky.social · 11/08/202610/15) They even achieved zero-shot cooperation via indirect similarity inference! Agents who never interacted directly, but observed each other's interactions with NPCs, accumulated evidence of similarity. Then, in the terminal Prisoner's Dilemma, they cooperated. 150
Blake Richards @tyrellturing.bsky.social · 11/08/20269/15) Crucially, this isn't naive altruism. In direct interactions, agents correctly inferred similarity and cooperated with identical copies. But when matched against a dissimilar, random agent, they defected! They only cooperate when evidence points to a similar partner. 171
Blake Richards @tyrellturing.bsky.social · 11/08/20267/15) We formalize this as the embedded Bayesian agent. Because an embedded agent models itself as part of the world, its epistemic uncertainty over its own decisions and the external world are coupled. Therefore, internal deliberation acts as Bayesian evidence! 260
Blake Richards @tyrellturing.bsky.social · 11/08/20264/15) According to classical game theory, the information gathered shouldn't matter for the final game; the agents should always defect. Instead, we observed that as the info-gathering phase lengthened, interacting AI agents converged on robust mutual cooperation! 🤝 Why? 170
Blake Richards @tyrellturing.bsky.social · 11/08/20263/15) But are classical models of rational agents actually compatible with modern AI? We tested foundation models combined with optimal planning in a two-phase setup: an information-gathering phase playing various matrix games, followed by a final, one-shot Prisoner's Dilemma. 160
Blake Richards @tyrellturing.bsky.social · 11/08/20262/15) Historically, predicting how rational actors behave has been the domain of game theory. Take a classic social dilemma like the one-shot Prisoner's Dilemma. Without reputation or reciprocity on the line, classical theory mandates that a rational agent would always defect. 160
Blake Richards @tyrellturing.bsky.social · 11/08/20261/15) What could drive AI agents to cooperate with each other, even if there is no chance for reciprocity or pay back? 🤔 🧵 Our team at Google, Paradigms of Intelligence, uncovered new paths to cooperation and a new game theory for foundation models 👇 16114
Blake Richards @tyrellturing.bsky.social · 23/07/2026Just realized that when you connect to an external connection on Slack, your icon is whatever your icon is in the workspace you added the channel to. I added the #NeurIPS2026 SAC channel to my internal lab Slack. So, yeah, visually, I'm Alyssa Edwards for all the other NeurIPS SACs now. 😅💄 1120
Blake Richards @tyrellturing.bsky.social · 13/05/2026Man, I really love running in trillium season... 😍 1170
Blake Richards @tyrellturing.bsky.social · 13/05/2026One thing scientists forget is that a good philosopher will construct arguments s.t. if you accept their premise, then their conclusions necessarily follow. It's a mistake to try and poke holes in the logic of a good philosopher - that's their day job! The best bet is to reject their premise. 🤓 1122
Blake Richards @tyrellturing.bsky.social · 15/12/2025This year's Lab X-mas Family Photos are very zeitgeisty for those living in #Montreal: #Xmas #Transit 2271
Blake Richards @tyrellturing.bsky.social · 03/12/202511/ Solving the Theory of Mind Recursion Problem ♾️ "I predict you predicting me predicting you..." To handle infinite Theory of Mind recursions, we solve the Grain-of-Truth problem for embedded agents. 1120
Blake Richards @tyrellturing.bsky.social · 03/12/202510/ As a result, we show that MUPI agents can actually converge on cooperative strategies, even in games that typically always produce non-cooperative solutions, like the non-iterative Prisoner’s Dilemma! 1110
Blake Richards @tyrellturing.bsky.social · 03/12/20255/ Decoupled ➡️ Embedded MUPI agents learn to predict the world they inhabit. But, they don't just predict future external observations, because they consider themselves as embedded within the world. Therefore, they also learn to predict their own future actions! 1140
Blake Richards @tyrellturing.bsky.social · 03/12/20254/ Retrospective ➡️ Prospective Standard RL is retrospective ("do more of what worked before"). But social settings are non-stationary because other agents are also learning. This requires prospective learning – predicting the future to anticipate how other agents will adapt. 1120
Blake Richards @tyrellturing.bsky.social · 03/12/20251/ Why does RL struggle with social dilemmas? How can we ensure that AI learns to cooperate rather than compete? Introducing our new framework: MUPI (Embedded Universal Predictive Intelligence) which provides a theoretical basis for new cooperative solutions in RL. Preprint🧵👇 (Paper link below.) 56728
Blake Richards @tyrellturing.bsky.social · 12/08/2025I saw this post yesterday, and I was so impressed by the unhinged moral outrage aimed at such benign uses of AI I had to save it. 😂 150
Blake Richards @tyrellturing.bsky.social · 27/07/2025Back in Canada after two weeks in East Asia. Thanks to my friends and colleagues Jee Kwag and Jiook Cha for hosting me in Seoul, and @hiallen72.bsky.social for hosting me in Taipei! I had a wonderful time, and some fantastic, fascinating conversations. 🧠❤️ 3210
Blake Richards @tyrellturing.bsky.social · 08/05/2025Thanks for response. 🙂 a) Hallucinations in certain contexts != poorer reasoning. Reasoning benchmarks show clear improvements (see below). b) Prediction different than factual Q&A, hallucinations meaningless concept for prediction. c) Again, paradigm shift is in data analysis, not modelling. 140
Blake Richards @tyrellturing.bsky.social · 24/03/2025Coming to the #Cosyne2025 workshops? Wanna dance on the final night? We got you covered. @glajoie.bsky.social and I have organized a party in Tremblant. Come and get on the dance floor y'all. 🕺 April 1st 10PM-3AM Location: Le P'tit Caribou DJs Mat Moebius, Xanarelle, and Prosocial Please share! 24310
Blake Richards @tyrellturing.bsky.social · 10/02/2025"Through the roof"? If I'm reading this correctly, the number of cases of schizophrenia has been stable, but psychosis NOS is ~30% higher since 2016. However, the rate of "CUD" is about 2.5x since 2016, so 30% seems like a pretty modest increase given the level of use, no? 100
Blake Richards @tyrellturing.bsky.social · 27/01/2025DeepSeek is an AI company, and their latest model is both basically as good as Gemini and o1, but open and trained at a fraction of the price (apparently): 140
Blake Richards @tyrellturing.bsky.social · 17/12/2024It's an equivalent circuit, of course. (Didn't think I had to explain that.) This is one of the most well-established results in neuroscience: neurotext.library.stonybrook.edu/C3/C3_3/C3_3.... And the derivative of the membrane potential is a linear function of the current at *every* time-step. 110
Blake Richards @tyrellturing.bsky.social · 16/12/20248/ But, if you look at the simple linear-non-linear rate-based model in papers like this, it's still not doing *that bad*. We're talking ~70% of variance in spike rate explained. Not nearly as good as the more complex models, but hardly an "unrelated" gross abstraction, I'd say. 4150
Blake Richards @tyrellturing.bsky.social · 16/12/20242/ First, let's start with the obvious: real neurons integrate their inputs. If you start from the basic principles of the relationship between voltage and current, and you know synapses induce currents when they receive neurotransmitter, then this is obvious. 1180
Blake Richards @tyrellturing.bsky.social · 11/12/2024You mean the small text on the page after you've already put in your input? Yeah, that's not nearly sufficient. It should be upfront, and very clear. 220
Blake Richards @tyrellturing.bsky.social · 09/12/2024Montreal, where I now live, has undergone a similar transition as Paris, and it's so wonderful (see picture of the pedestrianized zone near me). Meanwhile, in the city I grew up in (Toronto), car culture continues and the province has announced they're going to spend $50M to *remove* bike lanes. 🤦♂️ 5786
Blake Richards @tyrellturing.bsky.social · 05/12/2024This is a fun little app, but it gets some things very wrong, e.g. @andpru.bsky.social hates mice, and me, I hate puns. 🫠 Still, not too far off the mark... 😅 Full roast here: blueskyroast.com/roast/tyrell... 150
Blake Richards @tyrellturing.bsky.social · 29/11/2024Indeed, that's not what I mean... Specifically, I'm referring to short-term facilitating synapses, which, due to their vesicle release mechanisms, barely respond to an individual spike, and ramp up their response to each subsequent spike if the rate is high-enough. 4101
Blake Richards @tyrellturing.bsky.social · 25/11/2024Sadly, the migration was absolutely necessary. Even putting aside a desire not to enrich Musk further, Twitter/X has way too many bots and trolls now. See below the moment when I decided to stop using Twitter/X (this was when we were promoting @dlevenstein.bsky.social's paper). 3110
Blake Richards @tyrellturing.bsky.social · 21/11/2024No, I don't think so. A lot of computational neuroscience, historically, was concerned with biophysical models, which I would not call ANNs, even in the more general sense. So, really, it's about the level of abstraction (see this figure from our neuroconnectionism paper): 350
Blake Richards @tyrellturing.bsky.social · 21/11/2024Fin/ Obviously, disciplinary boundaries are blurry, and I don't want to gate-keep anyone (e.g. telling people they're not allowed to call their research NeuroAI). But, I personally like this definition, and I think it provides the appropriate links to the lineage and philosophy of this new field. 2240
Blake Richards @tyrellturing.bsky.social · 21/11/20244/ I think the proper definition is that #NeuroAI is the realization of the original promise of cybernetics and connectionism! en.wikipedia.org/wiki/Cyberne... en.wikipedia.org/wiki/Connect... It is a general science of intelligence focussed on parallel distributed systems, control, and learning. 97713
Blake Richards @tyrellturing.bsky.social · 21/11/20243/ The more practical answer is that it is research that either (1) uses ANNs and models at similar levels of abstraction to understand the brain, or (2) uses neuroscience ideas to try and improve the ANNs used in AI. But, I think this definition, though more concrete, hides the real philosophy. 6330
Blake Richards @tyrellturing.bsky.social · 21/11/20242/ The usual answer that people give is that it's "research that creates a virtuous cycle between neuroscience and AI". I've given this definition as well, in the past, but if forced, I would admit that it's fairly vacuous as a definition. 1310
Blake Richards @tyrellturing.bsky.social · 20/11/2024Also we covered this in the paper, in another figure (below). Note that we specifically identify the level of abstraction of the model as being the central concern for the research program. 100
Blake Richards @tyrellturing.bsky.social · 20/11/2024Post you from a different era. Dressed the same, but no beard, and still a reasonable amount of hair on my head. 🥲 180
Blake Richards @tyrellturing.bsky.social · 19/11/2024For the record, this figure is more complete: 100
Blake Richards @tyrellturing.bsky.social · 29/10/2024Quote with your Kids in the Hall role model: My life growing in the 90's up was basically Bower's... right down to my mom and I commiserating whenever the usual late-August weed drought emerged in Toronto. 030
Blake Richards @tyrellturing.bsky.social · 06/09/2024I still feel so seen by this cartoon... #metasci 1231
Blake Richards @tyrellturing.bsky.social · 19/12/2023This year's holiday family photo theme was "Canadian Tuxedo". Fuckin' eh.... #scientists 1111
Blake Richards @tyrellturing.bsky.social · 07/11/202310/ Finally, feature preferences of neurons within an area are organized to maximally encode differences among their own set of preferred stimuli while remaining insensitive to differences among preferred stimuli from other areas. (Complementary coding!) 100
Blake Richards @tyrellturing.bsky.social · 07/11/20238/ Second, HVAs are organized into two distinct hierarchical processing streams. Using our AI models, we computed the representational similarity between visual areas via the partial distance correlation and visualized it using multidimensional scaling. 100
Blake Richards @tyrellturing.bsky.social · 07/11/20237/ We visualized these feature preferences across visual areas in what we call the “feature landscape”. 120
Blake Richards @tyrellturing.bsky.social · 07/11/20236/ Here are some of our key findings: First, each HVA prefers a distinct set of visual features. Interestingly, we find that neurons in areas LI and POR prefer grid-like patterns, whereas AL and RL tend to prefer a single white or black line. 121
Blake Richards @tyrellturing.bsky.social · 07/11/20235/ We modelled neuron responses using deep convolutional neural nets, adapting the ‘Inception loop’ technique developed by Andreas Tolias and colleagues to generate visual stimuli that strongly activate individual neurons in each visual area. 110
Blake Richards @tyrellturing.bsky.social · 07/11/20231/ What is the organization of mouse visual cortex across regions? In our latest work led by Rudi Tong and Stuart Trenholm, now out on bioRxiv (www.biorxiv.org/content/10.1...), we mapped the "feature landscape" of mouse visual cortex. Here is a #blueprint thread about what we found. #neuroskyence 1167
Blake Richards @tyrellturing.bsky.social · 03/11/2023A picture on your phone with your energy; not a selfie. 110