4 ms·

The most comforting AI news I have read this year.

by fergie 4y ago

jackmott42 4y ago
If you are worried about AI, this shouldn't make you feel a ton better. GPT4 is just trained to predict the next word, a very simple but crude approach and look what it can do! Imagine when a dozen models are wired together and giving each other feedback with more clever training and algorithms on future faster hardware. It is still going to get wild
- ignoramous 4y ago
  > Imagine when a dozen models are wired together... Wouldn't these models hallucinate more than normal, then?
- causi 4y ago
  I worry that the hardware requirements are only going to accelerate the cloud-OS integration. Imagine a PC that's entirely unusable offline.
  - cj 4y ago
    > Imagine a PC that's entirely unusable offline. FWIW we had thin clients in computer labs in middle school / high school 15 years ago (and still today these are common in enterprise environments, e.g. Citrix). Biggest issue is network latency which is limited by the speed of light, so I imagine if computers in 10 years require resources not available locally it would likely be a local/cloud hybrid model.
- ShamelessC 4y ago
  Machine learning is actually premised on being “simple” to implement. The more priors you hardcode with clever algorithms, the closer you get to what we already have. The point is to automate the process of learning. We do this now with relatively simple loss functions and models containing relatively simple parameters. The main stipulation is that they are all defined to be continuous so that you can use the chain rule from calculus to calculate the error with respect to every parameter without taking so long that it would never finish. I agree that your suggested approach of applying cleverness to what we have now will probably produce better results. But that’s not going to stop better architectures, hardware and even entire regimes from being developed until we approach AGI. My suspicion is that there’s still a few breakthroughs waiting to be made. I also suspect that sufficiently advanced models will make such breakthroughs easier to discover.
- xwdv 4y ago
  People think something magical happens when AI are wired together and give each other feedback. Really you’re still just predicting the next word, but with extra steps.
  - Teever 4y ago
    People think that something magical happens when transistors are wired together and give each other feedback. Really you're just switching switches on and off, but with extra steps.
- ryneandal 4y ago
  Personally, I'm less worried about AI than I am about what people using these models can do to others. Misinformation/disinformation, more believable scams, stuff like that.
- quonn 4y ago
  I have repeatedly argued against this notion of „just predicting the next word“. No. It‘s completing a conversation. It‘s true that it is doing this word by word, but it‘s kind of like saying a CNN is just predicting a label. Sure, but how? It‘s not doing it directly. It‘s doing it by recovering a lot of structure and in the end boiling that down to a label. Likewise a network trained to predict the next word may very well have worked out the whole sentence (implicitly, not as a text) in order to generate the next word.
- DesiLurker 4y ago
  I actually have high hopes for the hybrid architecture Ben Goertzel has been working on at OpenCog. I think the LLMs are soon going to hit a S curve w/o introduction of additional scientific knowledge like physics and notion of energy (wrt AGI development, they'll still be good for tonnes of other jobs displacing things).
Freire_Herval 4y ago
[dead]
famouswaffles 4y ago
Title is misleading lol. Plenty of scale room left.
DesiLurker 4y ago
I agree, probably for a different reason. IMO the fact that Sam is saying this likely means that the LLMs are close to the upper knee of the S-Curve and after GPT5 they likely dont have many new fundamental ideas (additional SCurves) to throw at it. since they are ahead of the rest, it likely means we might be headed for an AI improvement pause for couple of years after GPT5. Still good enough to upset the balance in search/ad market. Interesting times.
- dragonwriter 4y ago
  > it likely means we might be headed for an AI improvement pause for couple of years after GPT5. I suspect that a pause in base LLM performance won’t be an AI improvement pause; there’s a whole lot of space to improve the parts of AI systems around the core “brain in a jar” model.
  - DesiLurker 4y ago
    I agree, there will be other things to be improved in AI system, but IMHO (tea leaves reading really) it would only lead to incremental improvements in overall systems. Also there is a lot of 'interfacing' work that needs to happen & i suspect that would end up filling the pause, which really is LLM productization loosely speaking. far as AGI is concerned I dont believe LLMs are really the right architecture for it, AGI likely needs some symbolic logic and a notion of physicality (ie.. physical laws & energy/power).
    - dragonwriter 4y ago
      > but IMHO (tea leaves reading really) it would only lead to incremental improvements in overall systems. It will reach a point where that is the case, sure; it is not there now, and if we are within one model generation of exhausting (for now) major core model improvements, I don’t think we’ll have reached the point of gradual incremental improvement from rest-of-system improvements yet.

The most comforting AI news I have read this year.

by fergie 4y ago

The most comforting AI news I have read this year.

If you are worried about AI, this shouldn't make you feel a ton better. GPT4 is just trained to predict the next word, a very simple but crude approach and look what it can do! Imagine when a dozen models are wired together and giving each other feedback with more clever training and algorithms on future faster hardware. It is still going to get wild

ignoramous 4y ago
> Imagine when a dozen models are wired together... Wouldn't these models hallucinate more than normal, then?
causi 4y ago
I worry that the hardware requirements are only going to accelerate the cloud-OS integration. Imagine a PC that's entirely unusable offline.
- cj 4y ago
  > Imagine a PC that's entirely unusable offline. FWIW we had thin clients in computer labs in middle school / high school 15 years ago (and still today these are common in enterprise environments, e.g. Citrix). Biggest issue is network latency which is limited by the speed of light, so I imagine if computers in 10 years require resources not available locally it would likely be a local/cloud hybrid model.
ShamelessC 4y ago
Machine learning is actually premised on being “simple” to implement. The more priors you hardcode with clever algorithms, the closer you get to what we already have. The point is to automate the process of learning. We do this now with relatively simple loss functions and models containing relatively simple parameters. The main stipulation is that they are all defined to be continuous so that you can use the chain rule from calculus to calculate the error with respect to every parameter without taking so long that it would never finish. I agree that your suggested approach of applying cleverness to what we have now will probably produce better results. But that’s not going to stop better architectures, hardware and even entire regimes from being developed until we approach AGI. My suspicion is that there’s still a few breakthroughs waiting to be made. I also suspect that sufficiently advanced models will make such breakthroughs easier to discover.
xwdv 4y ago
People think something magical happens when AI are wired together and give each other feedback. Really you’re still just predicting the next word, but with extra steps.
- Teever 4y ago
  People think that something magical happens when transistors are wired together and give each other feedback. Really you're just switching switches on and off, but with extra steps.
ryneandal 4y ago
Personally, I'm less worried about AI than I am about what people using these models can do to others. Misinformation/disinformation, more believable scams, stuff like that.
quonn 4y ago
I have repeatedly argued against this notion of „just predicting the next word“. No. It‘s completing a conversation. It‘s true that it is doing this word by word, but it‘s kind of like saying a CNN is just predicting a label. Sure, but how? It‘s not doing it directly. It‘s doing it by recovering a lot of structure and in the end boiling that down to a label. Likewise a network trained to predict the next word may very well have worked out the whole sentence (implicitly, not as a text) in order to generate the next word.
DesiLurker 4y ago
I actually have high hopes for the hybrid architecture Ben Goertzel has been working on at OpenCog. I think the LLMs are soon going to hit a S curve w/o introduction of additional scientific knowledge like physics and notion of energy (wrt AGI development, they'll still be good for tonnes of other jobs displacing things).

Freire_Herval 4y ago

[dead]

famouswaffles 4y ago

Title is misleading lol. Plenty of scale room left.

DesiLurker 4y ago

I agree, probably for a different reason. IMO the fact that Sam is saying this likely means that the LLMs are close to the upper knee of the S-Curve and after GPT5 they likely dont have many new fundamental ideas (additional SCurves) to throw at it. since they are ahead of the rest, it likely means we might be headed for an AI improvement pause for couple of years after GPT5. Still good enough to upset the balance in search/ad market. Interesting times.

dragonwriter 4y ago
> it likely means we might be headed for an AI improvement pause for couple of years after GPT5. I suspect that a pause in base LLM performance won’t be an AI improvement pause; there’s a whole lot of space to improve the parts of AI systems around the core “brain in a jar” model.
- DesiLurker 4y ago
  I agree, there will be other things to be improved in AI system, but IMHO (tea leaves reading really) it would only lead to incremental improvements in overall systems. Also there is a lot of 'interfacing' work that needs to happen & i suspect that would end up filling the pause, which really is LLM productization loosely speaking. far as AGI is concerned I dont believe LLMs are really the right architecture for it, AGI likely needs some symbolic logic and a notion of physicality (ie.. physical laws & energy/power).
  - dragonwriter 4y ago
    > but IMHO (tea leaves reading really) it would only lead to incremental improvements in overall systems. It will reach a point where that is the case, sure; it is not there now, and if we are within one model generation of exhausting (for now) major core model improvements, I don’t think we’ll have reached the point of gradual incremental improvement from rest-of-system improvements yet.