Hacker News Reader

Comment by brucethemoose2 | Hacker News Reader

Well, there is lots of low hanging fruit in AI land that, IMO, isn't being taken advantage of.

I'm thinking of stuff like ML compiler integration or other smaller features (including safety features) in the already-popular frameworks being championed by lone hero devs. 2-3 paid ML devs could move mountains right now.

Somehow I'm skeptical, and suspect Mozilla will try to create yet-another-standalone-framework that no one really cares to use, among the pile of 10,000 other from-scratch ML frameworks.

View on HN

Those low hanging fruit are useless in the face of much more advanced models and even agi coming in the next year or 2. Every effort spent implementing workflows for existing models is pointless when you can just make a better model that has the intelligence to do it automatically.

I dunno about that.

Stable Diffusion is a good example. Some straight up better models, or breaking variants of SD, have come out, but the sheer amount of tooling and augmentations clamped onto Stable Diffusion has kept it dominant.

Llama is another. The intertia and tooling has made MPT and Falcon, and all other non Llama architectures, basically irrelevant.

Not really pointless. I was messing with PlanAndExecute agents for a while, then OpenAI released Functions. Now with Functions I can implement an OpenAI MultiFunction agent that is significantly more reliable, more stable, qualitatively better...

I don't think the efforts that are being expended now to make v0, PoC systems that leverage these newer, more powerful tools is "pointless" - if anything it just helps all participants gauge what is useful, what isn't.

If you want to rely on optimism that AGI is "here" or "around the corner" then sure, go for it... but there's a lot to do in this new paradigm

>much more advanced models and even agi coming in the next year or 2

Are you certain this is going go happen? Especially in such a short timeframe.

Reply on news.ycombinator.com