AGI's impact on Tech, SaaS valuations | Hacker News Reader

AGI's impact on Tech, SaaS valuations

nextword.substack.com

17 pointsnextworddev24 comments

Is it just me or the whole premise (the "super AGI thing") is just way too "good" to be a real thing in the short term and there's not much to gleam out of this?

There's been very little actual value being realized with the LLMs and enterprises will take forever to make use of localized models due to their own large information silos.

This whole thing reads like we've invented real artificial intelligence, stuff like:

"For example, if AGI can do taxes perfectly in 5 minutes, why need Intuit"

Most LLMs can't even do basic math. Sometimes I feel like I'm living in a completely different universe than the AI hype folks.

Many of the people hyping AI work on AI. Their livelihoods depend on the hype.

Yes, I think this article is positing 99th percentile AI upside in the next five years, to the extent that tech and SaaS valuations will be the least of our concerns. AGI that can replicate *Salesforce* nearly instantly from a video and *migrate all your data* within 20 minutes is probably society-ending as we know it. Forget human-level AI -- that's "100 human-years in 20 minutes" level AI.

As I was reading, my intuition said the human couldn’t take the video and explain it in 20 minutes. Also the data could not be migrated that fast for any kind of real scale. However, GPT-4 is fairly close to filling in the AI side of that equation. Five years is a long time in AI research these days.

GPT-4 is amazing in short context windows, and you can patch together more complex stuff out of that, but an app with the complexity of Salesforce, with the associated backend, is orders of magnitude more complex (the difference between a React component and hundreds of years of human effort.) If it's allowed to cut corners and not actually build a working, scalable backend, then maybe -- but that sort of belies the premise of being able to migrate all your data.

Teams/companies of humans are society's current way of producing superhuman structures. If AI can just perform at that level within 5 years, it's moved way, way past the current bar, is all I'm saying.

Where was the bar five years ago?

Not much lower than it is now? The things that have gotten dramatically better are producing passable creative works, text->speech/speech->text but not much in terms of actually automating anything that requires a modicum of logical reasoning.

I know several companies working on the latter and several companies working on the former. Jam ‘em together and suddenly humankind has reached 100 years of progress? No way; this is just the start. Imagine what we can accomplish once all the codewriting and data mungling has been taken off of our hands — what wonders we can dream.

what bugs we're going to have to fix when these things inevitably fuck up because they're at their core stochastic machines.

I don’t work in the industry.

LLM isn’t the only type of A.I.

AGI will be nothing and then very suddenly the real deal.

Wolfram can do math.

Valuations are a subjective and not a science anyways and swing with market influences.

LLMs seem like a neat trick to me. I see it as perhaps a language generating component of an AGI, and think our brains probably do a similar trick, but it's not any kind of general intelligence when an LLM does or when our brain does it. I really don't see it as a bridge or any significant stepping stone to AGI, aside from maybe a nice linguistic interface.

LLMs are great, but they still require someone to understand what the hell they output and edit the garage they produce.

I've been using ChatGPT endlessly every day, and it has improved my ability to self learn, and grow by leaps and bounds, but it is utterly terrible once you cross certain thresholds of unsolved problems or deep abstractions.

The next huge jump on value will be leveraging trained models that know your entire knowledge base or code base, but transformers don't do well adapting to large volumes of live updated context, so it's too expensive for now to really take that last jump.

We are stuck either improving the feedback loop of integrating existing LLM output, or need to wait for the next type of AI that is better at keeping and updating large contexts.

No, it's not just you, but a lot of people are downplaying the dangers of AI. The easiest to accept one is that it can cause mass unemployment and displacement of workforce. No, it doesn't matter that we have better jobs that what the luddite textile workers lost 200 years ago, because it's not guaranteed to be the same situation (indeed, I'd say it's guaranteed to be different) and those luddites ended up in way worse situation anyway.

So the thing is that nobody knows what the development curve of AI is going to be and what the exact economical and societal effects are going to be. Whether it's 5 years to AGI or 50. (Neither of these seem very likely BTW.) Now since we do expect that there can be problems and since we at least can't rule out that these will manifest in the foreseeable (near) future, it's better to assume that we will have (at least economical) problems soon. It doesn't matter what LLMs can do today.

The development curve is what matters. And even if I said we don't know it, we have pretty good reasons to think (see above) that it's going to be powerful enough soonish. Just remember: about 1.5-2 years ago basically nobody would have predicted that LLMs would be able to do what they can do today. And I mean most experts would have probably said that it's not possible for LLMs to do what they can do today at all . Definitely not that they would be doing it by mid 2023. Or even just that they would be so powerful that a lot of non-technical people would use them. (Though, sure, there is still very little practical use as of today but the capabilities did make a huge and unexpected jump. It even surprised researchers like Geoffrey Hinton.)

It feels like math issue is solvable, so are an all the other more precise engineering type things (even basic calculators can do all that). Yes, it’s early and but it seems like llms are on the right path this time, and what the article describes is achievable.

What the author I think is missing is if they can automate s/w engineering with this, by that time all other existing office work will largely be automated as well. Whether it really happens in our lifetimes is debatable.

AGI doesn't exist yet therefore current impact is zero.

And no... I've relinquished the AI term for these little statistical tricks instead of actual intelligence. We had to modify "artificial intelligence" to mean these pale mimics, and say "artificial general intelligence" to mean something that is actually aware and intelligent. No, marketers, you can't have that one, too.

AGI is R2D2, R. Daneel Olivaw, or V'Ger (minus the planetary destruction). It is not the thing that generated an ever-so-slightly off picture of a clown.

Totally agree, it's as if suddenly with LLMs everyone thinks they know 'stuff' or self declare experts because they can interact with an AI.

AI is overhyped because of how it's easy to use, and there is a bias tricking people into believing statistical tricks = consciousness.

Before being ready to create a conscious mind out of silicon, we should be ready to prevent any harm actual AI/LLM can cause to our society.

Ideally, we'd also be able to formally classify intelligence and consciousness. We currently can't. It's apparently as ineffable as pornography or art. I'm not even sure we'll know it when we see it.

That's a whole lotta text to say AGI would screw SaaS and still, IMO, misses the point completely in terms of the actual impact. What's defensible is BS in this model as the reality is it would push the overeall "system" in ways that are hard to predict (e.g. limit behaviors and bifurcations) and will likely make defensibility arguments random.

> non-technical person faithfully replicate an app as complex as Airtable or Salesforce via 1) taking a video of the app, 2) stating the requirements, 3) and having the AI produce and customize the app, while 4) migrating all existing data. All within say 20 minutes.

We already have this. RoR can do this. This is not the hard part of scaling a business.

HN is faithful to its “I could build it in a weekend” estimates, but even on that scale “a non-technical person could replicate Salesforce with RoR in 20 minutes” is extra brave.

Sorry, my point is making a crud +front end is not the hard part, which is what ChatGPT 5.0 will be able to do.

I like that his premise is basically pretending computers are magic.

He says:

> Extreme margin compression due to more supply of software. AGI simply enables vastly more creation of software, which hurts the pricing power of incumbents. Every company will do some double take on their IT spending, and try to rebundle.

I think this is actually a big misunderstanding about how software works. MORE software isn't necessarily a good thing. That's more systems to deploy and maintain and probably a shit ton of redundancy.

> has near infinite context length

Which has near infinite costs, or it's not the same thing as attention. So it'd be something different... or yeah, computer = magic.

Reply on news.ycombinator.com