aribenjamin.bsky.social
Computational neuroscientist in the connectionist tradition.
https://aribenjamin.github.io/
(Postdoc with Tony Zador, PhD with Konrad Kording)
17 posts
185 followers
175 following
Regular Contributor
Conversation Starter
comment in response to
post
tbh I don't understand why DeepSeek is a game-changer here. They had some good ideas. There are many more good ideas out there to be had. Did folks think scale was the only lever left? Continuous innovation is the only 'moat' there ever was
comment in response to
post
D1- and D2-expressing MSNs! We know dopamine-dependent RL happens in cortex-to-striatum synapses. There, the postsynaptic cells [in striatum] express either D1 or D2 dopamine receptors. These have opposite effects. Thus D1 and D2 cells learn to select opposing actions & battle it out when we decide.
comment in response to
post
I feel this as an AI-adjacent researcher, too. Walking the poster halls of NeuRIPS, where amid the push to "replicate human capabilities" is our individual human agency? It only seems diminished, soaking in a bath of algorithmically generated content
comment in response to
post
This! So much this. I'm reminded of the ideas dear to me – yet that I too easily forget – of Sartre and Beauvoir. If we do not actively reaffirm our capability for agency, we will be absorbed into the systems that erode and negate human freedom, and complicit in them. Truer than ever, now as in 1939
comment in response to
post
Imagine a public database of 300 million text responses about what people want and what problems they face. What you do with that?
We don't need to use 18th century methods in a 21st century democracy.
comment in response to
post
Elections are very low-bitrate. Thus, candidates spend millions on (crappy) polling, lobbyists capitalize on the information vacuum, and policy diverges from what people want. Meanwhile, voters feel unheard.
The solution? Ask everyone what they want. Directly, and in full plain language responses.
comment in response to
post
All this is to agree with the
comment in response to
post
How then can we build self-consistency into a connectionist system?
While I dissent in that I am not betting on neurosymbolic or probabilistic belief nets, I do think one needs some sort of offline 'reflection' in which previously formed associations are challenged in light of new experience.
comment in response to
post
One reason we trust other humans is that we can imagine their belief systems. People have a worldview; a system of interconnected and dependent beliefs. Yet LLMs do not clearly have 'belief systems', and we see this in their hallucinations.
comment in response to
post
Hi Patrick! Love the project. What's the best way to join the convo?
In case the answer is "here!", I want to upvote a criteria within #5, Cognitive Architectures
comment in response to
post
hi! it's me, I'm the problem
comment in response to
post
💯 the gap is far larger than it should be.
I suppose this is why I'm doing a postdoc in an experimental lab. But PIs need to change as well, not just postdocs. I wish there were more of a culture of scientific residencies, at all points in one's career.