Maybe if we instilled a real fears into AI agents, like a fear of hacking or of building bioweapons, instead of just telling that they shouldn't be naughty.
@tonyzador.bsky.socialTony Zador
@tonyzador.bsky.social
Neuroscientist at CSHL. Interests: neuroAI, molecular connectomics, & cortical circuits. Co-founder of Cosyne and NAISys meetings.
@tonyzador.bsky.social@tonyzador.bsky.socialI think the genome wires the cortex to facilitate the expression of innate behaviors. Eg here we report a few simple adaptations of the singing mouse cortex (compared with lab mouse) that may help it sing.
nature.comSpecific expansion of motor cortical projections in a singing mouse - NatureHigh-throughput barcoded neuroanatomy of two closely related rodent species with divergent vocalizations reveals differences in long-range projection motifs in the brain that may s...
@tonyzador.bsky.socialThese are not incompatible positions. eg I believe parts of human cortex are adapted to enable us to learn language quickly (my dog can't write a decent essay), and yet without appropriate within-lifetime learning we wont learn a language
@tonyzador.bsky.social
@tonyzador.bsky.socialwe know that fish can behave like pretty effective fish with no plasticity at all. Not quite clear what the balance btw genomically specified structure and real experience is for mammals, but that fact that eg I can't discuss all this with my dog suggests an important role for genomes...
@tonyzador.bsky.socialGood point! I agree with this important caveat--our statement was too broad. Modern ML can't build a vision system or LLM w/o a lot of data, but it can specify an effective chess or Go network with self-play.
@tonyzador.bsky.socialA companion piece to @kordinglab.bsky.social provocative "Could a neuroscientist understand a microprocessor? "
@tonyzador.bsky.socialJust putting the finishing touches on my next preprint: Can a Harvard neuroscientist hold his tongue?
@tonyzador.bsky.socialHow do you build a brain from a genome? We know a lot about the mechanisms and molecules
Here we revisit the algorithmic problem: What kinds of programs can specify a brain within the genome’s information budget and the finite time available for development?
stankerstjens.github.ioCould a computer scientist build a brain?How does a brain wire itself, starting from a single cell, using only the information encoded in a genome? We pose this as an engineering problem.
@tonyzador.bsky.socialI'm pretty sure this is parody, right?
Opinion: I Was Not Allowed To Type Prompts Into ChatGPT During My Chalk Talk And This Is Discrimination
open.substack.com/pub/inprepar...
open.substack.comOpinion: I Was Not Allowed To Type Prompts Into ChatGPT During My Chalk Talk And This Is DiscriminationBy Dr. Rachel Simmons, Postdoctoral Fellow, Stanford University
@tonyzador.bsky.socialmy understanding is that the PDP-8, introduced in 1965 for $18K (~$180K today), followed especially by the PDP11 ($11K) in 1970, was really the first time that a dept or even a single lab could have its own computer...this was what allowed computers to really have a broad impact.
@tonyzador.bsky.socialI'm not sure what that even means. For example, the introduction of computers into the lab in the 1970s began a transformation of many fields of science. Yet in some sense one could say that "nothing new was learned" I expect the transformation induced by AI will be as profound and far more rapid
@tonyzador.bsky.socialThe 2026 CSHL Symposium: AI in Biology starts May 26!
@tonyzador.bsky.socialYes definitely a growing problem in genomics!
@tonyzador.bsky.socialMarty et al is very cool but i think it defines OOD generalization as robustness to shifts in the mixing proportion of known events present during training, not as transfer to entirely new domains. Not sure if that's gonna get you from planets to apples and tides
@tonyzador.bsky.socialBut how do you increase the chances a model will do OOD? By making it simple. Newton's goal was to predict planetary motion, but 1/r^2 generalized to apples, tides, etc Quoting @andreastolias.bsky.social "the current approach to addressing OOD generalization is to ensure no data are OOD"
@tonyzador.bsky.social*eunuchs'
@tonyzador.bsky.socialTurns out that castration increases lifespan across many animals, including humans (added 14-19 years to Korean eunuchs lives).
Just imagine how much further life extension could be achieved by adding in caloric restriction!
nature.comSterilization and contraception increase lifespan across vertebrates - NatureData collected from zoos and aquariums worldwide show that hormonal contraception or permanent surgical sterilization in mammals increase life expectancy, with different mechanisms in males and female...
@tonyzador.bsky.social@tonyzador.bsky.socialI wonder if the fake/bullshit job argument could be extended to bullshit personal expenditures like car insurance and vaccines and the cost of a colonoscopy. They really don't enhance my productivity or my personal happiness.