This is an excellent thread summarizing many of my own objections. I’d like to add two thoughts: 🧵 1/n - ThreadSky

jmxpearson.bsky.social • 70 days ago

This is an excellent thread summarizing many of my own objections. I’d like to add two thoughts: 🧵 1/n

There's a movement in neuroscience suggesting we should be pursuing bigger bets with larger teams. I think there's a case for doing a bit of this, but I think it's a bad idea to prioritise it for two reasons, and a good case for saying we should be moving in the exact opposite direction. 🧠🧪

Comments

jmxpearson.bsky.social•70 days ago

First, the unspoken premise in several of the arguments made by @davidamarkowitz.bsky.social @tyrellturing.bsky.social @patrickmineault.bsky.social and others is that what we’re really lacking is scale. 2/n

jmxpearson.bsky.social•70 days ago

If the question is of data collection, as @patrickmineault.bsky.social argues, then sure. If we want to generate big data sets like the International Brain Lab or Allen Institute have, scale helps. And lots of people benefit. 3/n

jmxpearson.bsky.social•70 days ago

If, as @tyrellturing.bsky.social suggests, we want to train giant foundation models, then yes, scaling is what’s needed. I disagree with this approach—I think the likeliest outcome is a very good predictive model and no insight—but it’s a project that could be done and takes big resources. 4/n

jmxpearson.bsky.social•70 days ago

For me, though, @davidamarkowitz.bsky.social’s claim that we are on the threshold of understanding intelligence seems to fly in the face of all the empirical evidence about what very impressive LLMs actually do. Many including @irisvanrooij.bsky.social and @kordinglab.bsky.social on this. 5/n

jmxpearson.bsky.social•70 days ago

If we believe that we have all the essential ideas, that what we need is not research but development, well okay. A giant project makes sense. But if we are still short several key ideas, this is a terrible idea because *we can’t scale our way to insight!* 6/n

jmxpearson.bsky.social•70 days ago

As @neuralreckoning.bsky.social argues, these are highly stochastic events, and we are better off purchasing more lottery tickets! 7/n

davidamarkowitz.bsky.social•70 days ago

If you accept the premise that 10 years of the BRAIN Initiative has given neuroscientists all the tools needed to address our "known unknowns", then we are indeed at the threshold of gaining transformational insights about intelligence. (LLMs did not factor into my argument.)

jmxpearson.bsky.social•70 days ago

Fair enough on LLMs. I'm not sure I agree about the "known unknowns." Plenty of biology to be documented, to be sure, but my concern is more with whether we have all the necessary conceptual ingredients to start scaling up. Others may well disagree.

davidamarkowitz.bsky.social•70 days ago

(That said, there are lots of new tools I think could make aspects of this task easier/faster/cheaper. Tool development is a great job for small, focused, well-resourced teams.)

Comments

Posting Rules

Reply