stoniejohnson

Great post! I appreciate the diagrams.

If you're always looking up but never looking down it becomes hard to appreciate things.

Revisited this, and while it was tongue-in-cheek, it feels a bit ungrateful.

A lot of tech roles are intellectually stimulating and have a wage in the upper distribution of salaries. I think we have a very nice situation going for us.

While those are good reasons to keep the weights static from a business perspective, they are not the only reasons, especially when serving SOTA models at the scale of some of the major shops today.

Continual/online learning is still an area of active research.

They were high in 2013.

I get where you are coming from, and it is definitely an interesting thought!

I do think it is an extremely inefficient way to have a swarm (e.g. across time through training data) and it would make more sense to solve the pretraining problem (to connect them to the external world as you pointed out) and actually have multiple LLMs in a swarm at the same time.

I was responding to the idea that an LLM would believe (regurgitate) untrue things if you pretrained them on untrue things. I wasn't making a claim about SOTA models with gigantic training corpora.

I was responding to the back and forth of:

> If you pretrained an LLM with data saying Moscow is the capital of Connecticut it would think that is true.

> Well so would a human!

But humans aren't static weights, we update continuously, and we arrive at consensus via communication as we all experience different perspectives. You can fool an entire group through propaganda, but there are boundless historical examples of information making its way in through human communication to overcome said propaganda.

Humans are not isolated nodes, we are more like a swarm, understanding reality via consensus.

The situation you described is possible, but would require something like a subverting effort of propaganda by the state.

Inferring truth about a social event in a social situation, for example, requires a nuanced set of thought processes and attention mechanisms.

If we had a swarm of LLMs collecting a variety of data from a variety of disparate sources, where the swarm communicates for consensus, it would be very hard to convince them that Moscow is in Connecticut.

Unfortunately we are still stuck in monolithic training run land.

They have a dark pattern around annual subscriptions, i.e. I had a monthly subscription I wanted to downgrade, and with no confirmation they charged me for the entire year after I selected the lower tier and hit next.

This is probably why.

I issued a chargeback.

I'm not really familiar with that technology space, but if you take that as true, is your argument something like:

- We don't have limitless CPU cycles

- Thus we need to split things into sub-problems

If so that might still be amenable to the bitter lesson, where Sutton is saying human heuristics will always lose out to computational methods at scale.

Meaning something like:

- We split up the thought to vision problem into N sub-problems based on some heuristic.

- We develop a method which works with our CPU cycle constraint (it isn't some probe -> CPU interface). Perhaps it uses our voice or something as a proxy for our thoughts, and some composition of models.

Sutton would say:

Yeah that's fine, but if we had the limitless CPU cycles/adequate technology, the solution of probe -> CPU would be better than what we develop.

I think the bitter lesson implies that if we could study/implement "how a machine with limitless cpu cycles would make our eyes see something we are currently thinking of" then it would likely lead to a better result than us using hominid heuristics to split things into sub-problems that we hand over to the machine.