Item — HN Reader

We unfortunately do live in a post-post privacy world now: so LLMs (and personal assistants based on them) with higher-than GPT-4 levels of quality that are so efficient to train and use that we can target the next generation of state-of-the-art personal computing devices. All-in on on-the-edge and power efficiency (data center energy consumption will increasingly become ridiculous; as will Nvidia’s power…).

Why does a child need so much less data for “training” to reach “adult brain”?

Apple can’t be our only hope in this regard, can it?

Do they even employ anyone who could substantially push the envelope on the software side there? I would guess they would need to headhunt people from Google and OpenAI and at completely outlandish terms (less of an issue for them; rather than convincing to leave respective AAA “AI” teams). Academia probably the only hope, I’d straight go there if I’d be Apple. It will probably need a complete overhaul of the concepts / architecture of current ways to go about LLMs. Or maybe not, maybe just needs a “small fix”? I for sure don’t know.

2 comments

> Why does a child need so much less data for “training” to reach “adult brain”?

If we count all sensory input for years I do not think a child needs less data, but more.

If, in addition, we consider that the child gets new data (decides new interactions) based on past data, there is an optimization in the child that the LLM lacks: the LLM cannot just decide to get some new data based on the data it just saw. It is stuck with the data fed to it for training.

I’ve got two downvotes here, really curious on the why? It’s clearly not only the hardware that is “still too slow” at this point? Again just for training then, why does a LLM basically need the whole internet and then some to arrive at “uncanny valley intelligence” at best?

The question is: "what software do you need that doesn't exist?".

Your reply contains a lot of things, except an answer.

I’ll summarize then: “A GPT-4+ like and OSS LLM based personal assistant that can be trained and run on soon-to-be common personal computing devices.”

Lots of people are probably already working on that but it doesn’t exist yet.

I didn't downvote but I imagine they came from either (a) asking for what amounts to a personalized AGI on a phone, which doesn't seem to fit the spirit of the question or (b) talking about Apple like they're significantly related to (a).