4 ms·

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

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.

TheBozzCL 3y ago
Many of the people hyping AI work on AI. Their livelihoods depend on the hype.
tibbar 3y ago
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.
- block_dagger 3y ago
  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.
  - tibbar 3y ago
    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.
    - sdenton4 3y ago
      Where was the bar five years ago?
      - packetlost 3y 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.
- freeone3000 3y ago
  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.
  - __loam 3y ago
    what bugs we're going to have to fix when these things inevitably fuck up because they're at their core stochastic machines.
j45 3y ago
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.
theonemind 3y ago
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.
aeonik 3y ago
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.
atleta 3y ago
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.)
sheeshkebab 3y ago
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.

TheBozzCL 3y ago

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

tibbar 3y ago

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.

block_dagger 3y ago
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.
- tibbar 3y ago
  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.
  - sdenton4 3y ago
    Where was the bar five years ago?
    - packetlost 3y 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.
freeone3000 3y ago
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.
- __loam 3y ago
  what bugs we're going to have to fix when these things inevitably fuck up because they're at their core stochastic machines.

j45 3y ago

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.

theonemind 3y ago

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.

aeonik 3y ago

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.

atleta 3y ago

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.)

sheeshkebab 3y ago

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.