4 ms·

I am not convinced current AI is the approach to an AGI. I think it is at least feasible someone outside of this sphere has a reasonable shot at it. It feels li

I am not convinced current AI is the approach to an AGI. I think it is at least feasible someone outside of this sphere has a reasonable shot at it. It feels like many AI researchers get caught up in refining existing techniques that amount to fancy statistics algorithms and data crunching, but not AGI. Current AI techniques may be synthesized or used in part in some AGI but it’s clear there is a revolutionary step to be made. Kaggle is almost just an optimization fest, and not really advancing towards AGI.

TeamSlytherin 7y ago
The announcement today has the potential to cut the number of years estimated (to reach AGI) in half, and I'm sure VC funds are DM'ing him non-stop right now. But we don't know which path Carmack will take, and as you rightly point out, current trends in AI is mostly ML/xNN's with a goal of turning data into profit (heavy focus on products/markets). Even those talking about AGI are fractured into different groups. Oddly, many in ML are talking about future abilities that are mostly defined by AGI research (even if they see AGI as a distraction. From his post, I don't think "product development" is a goal. Not clear what challenges or milestones he will set for himself (just having a better testing suite has become an AGI issue of late, so maybe he will contribute to that first).
unityByFreedom 7y ago
> I am not convinced current AI is the approach to an AGI This is a non-sequitor. I didn't argue that current AI theory, or Kaggle, will lead us to AGI. > Kaggle is almost just an optimization fest Public machine learning competitions have produced a lot of innovative learning techniques. If Kaggle is so easy, and AGI so hard, it would follow that anyone tackling AGI would have some experience applying machine learning competitively in some public space. It doesn't necessarily need to be Kaggle. Kaggle just happens to be good at hosting such public competitions, and in fact has surfaced several state-of-the-art implementations. The difference between prize money (~top 5) and no-prize-money (5-10) may be an "optimization fest", but without the competition, the solutions presented would have entered the public sphere at a much slower rate. EDIT: Please be kind and explain your downvotes.
- jacobush 7y ago
  Ok, I'll take a shot: "it would follow that anyone tackling AGI would have some experience applying machine learning competitively in some public space". No, that would absolutely not follow. (I'm a pretty good devil's advocate, but I can't with this one.) And given AGI would come from some completely new breakthrough not related to the current practice of "machine learning", competitions may be completely moot. They may be great for finding the nice increments in the state of the art of machine learning, but they are unlikely to help much with AGI. I could even imagine a stumbling AGI being very stupid compared to just about any machine learning solution thrown at it - yet being undeniably AGI. Like a dog not being very good at DOTA, Star Craft or Chess, yet it undeniably possesses some kind of general intelligence.
  - unityByFreedom 7y ago
    Thank you for your reply. Turning a blind eye to existing knowledge may result in reinventing things that already exist. Nobody expects students to follow the same concepts as their teachers, the point is just to leverage existing knowledge. > I could even imagine a stumbling AGI being very stupid compared to just about any machine learning solution thrown at it - yet being undeniably AGI. Like a dog not being very good at DOTA, Star Craft or Chess, yet it undeniably possesses some kind of general intelligence. People have debated whether animals are intelligent for ages. This is another type of problem, how to define intelligence. The most famous attempt in recent times is the Turing test.
    - jacobush 7y ago
      Another type from what? If you can't define it, how can you optimize for it? (Certainly not in the online competitions of today.)
  - tarsinge 7y ago
    > And given AGI would come from some completely new breakthrough not related to the current practice of "machine learning" I’m not so sure of that. Intuitively AGI feels like being able to generalize and automatize what is already done in specialized problems, like having a meta program that that orchestrate and apply specialized subsystems, and adapt existing one. If playing Go, Starcraft, Speech recognition, Computer vision are already of the same building blocks, it feels like having a meta program that‘s just trained to recognize the type of problem and route it to the appropriate subsystem with some parameters tweaks is a path to AGI. In the dog example you don’t even need to have subsystem that are that better than humans individually. Edit: my point is I feel like AGI is the interface and orchestration between specialized subsystems we already know how to create. Trying to train a big network like generalizing Alpha Go is a dead-end, but having simpler sub networks ready to be trained at a specific problem seems feasible. Much like the brain is at first seen like a big network, but in practice there are specialized areas. The key is how are these networks interfaced and which information they exchange to self adapt. Maybe these interfaces themselves are sub networks specialized in the problem of interfacing and “tuning hyperparameters”. In short: I think when we’ll figure out how to automate Kaggle competitions (recognize the pattern of the problem, then instantiate and train the relevant subsystem) we’ll be a good step forward AGI. We don’t need better performance e.g. in image recognition, just how to figure orchestration.

TeamSlytherin 7y ago

The announcement today has the potential to cut the number of years estimated (to reach AGI) in half, and I'm sure VC funds are DM'ing him non-stop right now. But we don't know which path Carmack will take, and as you rightly point out, current trends in AI is mostly ML/xNN's with a goal of turning data into profit (heavy focus on products/markets). Even those talking about AGI are fractured into different groups. Oddly, many in ML are talking about future abilities that are mostly defined by AGI research (even if they see AGI as a distraction. From his post, I don't think "product development" is a goal. Not clear what challenges or milestones he will set for himself (just having a better testing suite has become an AGI issue of late, so maybe he will contribute to that first).

unityByFreedom 7y ago

> I am not convinced current AI is the approach to an AGI This is a non-sequitor. I didn't argue that current AI theory, or Kaggle, will lead us to AGI. > Kaggle is almost just an optimization fest Public machine learning competitions have produced a lot of innovative learning techniques. If Kaggle is so easy, and AGI so hard, it would follow that anyone tackling AGI would have some experience applying machine learning competitively in some public space. It doesn't necessarily need to be Kaggle. Kaggle just happens to be good at hosting such public competitions, and in fact has surfaced several state-of-the-art implementations. The difference between prize money (~top 5) and no-prize-money (5-10) may be an "optimization fest", but without the competition, the solutions presented would have entered the public sphere at a much slower rate. EDIT: Please be kind and explain your downvotes.

jacobush 7y ago
Ok, I'll take a shot: "it would follow that anyone tackling AGI would have some experience applying machine learning competitively in some public space". No, that would absolutely not follow. (I'm a pretty good devil's advocate, but I can't with this one.) And given AGI would come from some completely new breakthrough not related to the current practice of "machine learning", competitions may be completely moot. They may be great for finding the nice increments in the state of the art of machine learning, but they are unlikely to help much with AGI. I could even imagine a stumbling AGI being very stupid compared to just about any machine learning solution thrown at it - yet being undeniably AGI. Like a dog not being very good at DOTA, Star Craft or Chess, yet it undeniably possesses some kind of general intelligence.
- unityByFreedom 7y ago
  Thank you for your reply. Turning a blind eye to existing knowledge may result in reinventing things that already exist. Nobody expects students to follow the same concepts as their teachers, the point is just to leverage existing knowledge. > I could even imagine a stumbling AGI being very stupid compared to just about any machine learning solution thrown at it - yet being undeniably AGI. Like a dog not being very good at DOTA, Star Craft or Chess, yet it undeniably possesses some kind of general intelligence. People have debated whether animals are intelligent for ages. This is another type of problem, how to define intelligence. The most famous attempt in recent times is the Turing test.
  - jacobush 7y ago
    Another type from what? If you can't define it, how can you optimize for it? (Certainly not in the online competitions of today.)
- tarsinge 7y ago
  > And given AGI would come from some completely new breakthrough not related to the current practice of "machine learning" I’m not so sure of that. Intuitively AGI feels like being able to generalize and automatize what is already done in specialized problems, like having a meta program that that orchestrate and apply specialized subsystems, and adapt existing one. If playing Go, Starcraft, Speech recognition, Computer vision are already of the same building blocks, it feels like having a meta program that‘s just trained to recognize the type of problem and route it to the appropriate subsystem with some parameters tweaks is a path to AGI. In the dog example you don’t even need to have subsystem that are that better than humans individually. Edit: my point is I feel like AGI is the interface and orchestration between specialized subsystems we already know how to create. Trying to train a big network like generalizing Alpha Go is a dead-end, but having simpler sub networks ready to be trained at a specific problem seems feasible. Much like the brain is at first seen like a big network, but in practice there are specialized areas. The key is how are these networks interfaced and which information they exchange to self adapt. Maybe these interfaces themselves are sub networks specialized in the problem of interfacing and “tuning hyperparameters”. In short: I think when we’ll figure out how to automate Kaggle competitions (recognize the pattern of the problem, then instantiate and train the relevant subsystem) we’ll be a good step forward AGI. We don’t need better performance e.g. in image recognition, just how to figure orchestration.