Are you using the information from chatGPT’s results?
chatGPT uses ~3000 times less data to run than a human brain. (source from another post here, I would be interested in tighter estimates). This seems to be fairly compelling evidence that your model may be off by several OOMs.
It seems to be able to match, even exceed human performance in some areas
There are very obvious low hanging fruit that might radically improve it’s performance: a. Give it a calculator and RL problems to exercise this capability. This isn’t even an AI advance, it’s just missing it! b. Integrate it with perception and visualization models like DALL-E so it can extract information about the contents of an image and produce visualizations to illustrate when needed c. RL training on programming tasks d. RL training on all available english language multiple choice tests at high school/college level e. Give it a way to obtain information from the internet
f. add a second neural network and have the machine always presented with a ‘salient context’, or the most valuable tokens from the set of all prior tokens in this session with a user.
Will this hit 20% of cognitive tasks without further improvement? It might, and we will likely see an llm with all 6 improvements either in 2023 or 2024. (I would say I am certain we will see at least one before the end of 2023 and probably 3 or 4 by the end of 2024. I would estimate a greater than 50 percent chance all these improvements will be made and several more by EOY 2024.
Thank you for making the point about existing network efficiencies! :)
The assumption, years ago, was that AGI would need 200x as many artificial weights and biases when compared to a human’s 80 to 100 Trillion synapses. Yet—we see the models beating our MBA exams, now, using only a fraction of the number of neurons! The article above pointed to the difference between “capable of 20%” and “impacting 20%”—I would guess that we’re already at the “20% capability” mark, in terms of the algorithms themselves. Every time a major company wants to, they can presently reach human-level results with narrow AI that uses 0.05% as many synapses.
Yes. And regarding the narrow AI : one idea I had a few years ago was that sota results from a major company are a process. A replicable, automatable process. So a major company could create a framework where you define your problem, provide large amounts of examples (usually via simulation), and the framework attempts a library of know neural network architectures in parallel and selects the best performing ones.
This would let small companies get sota solutions for their problems.
The current general models suggest that may not be even necessary.
Holy shit I seen to have been correct about every single prediction.
A. Calculator. Yep, python interpreter and Wolfram
B. Perception, yep, gpt-4v. Visualization, yep. Both gemini and gpt4 can do this decently.
C. RL on programming. Yep, alpha code 2.
D. RL on college. Yep, open sources models have been refined this way.
E. Internet browsing, yep old news
F. Salient context: announced as a feature for gpt-4, not actually available at scale I don’t think.
Update: I was wrong about this. Tons of different GPT-4 wrapper bots use a vector embeddings database that effective gives the model salient context. So this was satisfied before EOY 2023.
Can this do 20 percent of cognitive tasks? Eh maybe? Issue is that since it can’t do the other 80 percent and it can be fooled many ways gpt-4 can’t actually do something like sell cars or other complete tasks. A human has to be there to help the model and check it’s work etc.
If you broke every task that humans do at all into a list, especially if you scaled by frequency, gpt-4 probably can in fact do 20 percent. (Frequency scaling means that easy tasks like “find all the goods in the store on the shopping list” happen millions of times per day while improving the bleeding edge of math is something few humans are doing)
Are you using the information from chatGPT’s results?
chatGPT uses ~3000 times less data to run than a human brain. (source from another post here, I would be interested in tighter estimates). This seems to be fairly compelling evidence that your model may be off by several OOMs.
It seems to be able to match, even exceed human performance in some areas
There are very obvious low hanging fruit that might radically improve it’s performance: a. Give it a calculator and RL problems to exercise this capability. This isn’t even an AI advance, it’s just missing it! b. Integrate it with perception and visualization models like DALL-E so it can extract information about the contents of an image and produce visualizations to illustrate when needed c. RL training on programming tasks d. RL training on all available english language multiple choice tests at high school/college level e. Give it a way to obtain information from the internet
f. add a second neural network and have the machine always presented with a ‘salient context’, or the most valuable tokens from the set of all prior tokens in this session with a user.
Will this hit 20% of cognitive tasks without further improvement? It might, and we will likely see an llm with all 6 improvements either in 2023 or 2024. (I would say I am certain we will see at least one before the end of 2023 and probably 3 or 4 by the end of 2024. I would estimate a greater than 50 percent chance all these improvements will be made and several more by EOY 2024.
Thank you for making the point about existing network efficiencies! :)
The assumption, years ago, was that AGI would need 200x as many artificial weights and biases when compared to a human’s 80 to 100 Trillion synapses. Yet—we see the models beating our MBA exams, now, using only a fraction of the number of neurons! The article above pointed to the difference between “capable of 20%” and “impacting 20%”—I would guess that we’re already at the “20% capability” mark, in terms of the algorithms themselves. Every time a major company wants to, they can presently reach human-level results with narrow AI that uses 0.05% as many synapses.
Yes. And regarding the narrow AI : one idea I had a few years ago was that sota results from a major company are a process. A replicable, automatable process. So a major company could create a framework where you define your problem, provide large amounts of examples (usually via simulation), and the framework attempts a library of know neural network architectures in parallel and selects the best performing ones.
This would let small companies get sota solutions for their problems.
The current general models suggest that may not be even necessary.
Holy shit I seen to have been correct about every single prediction.
A. Calculator. Yep, python interpreter and Wolfram B. Perception, yep, gpt-4v. Visualization, yep. Both gemini and gpt4 can do this decently. C. RL on programming. Yep, alpha code 2. D. RL on college. Yep, open sources models have been refined this way. E. Internet browsing, yep old news F. Salient context: announced as a feature for gpt-4, not actually available at scale I don’t think. Update: I was wrong about this. Tons of different GPT-4 wrapper bots use a vector embeddings database that effective gives the model salient context. So this was satisfied before EOY 2023.
Can this do 20 percent of cognitive tasks? Eh maybe? Issue is that since it can’t do the other 80 percent and it can be fooled many ways gpt-4 can’t actually do something like sell cars or other complete tasks. A human has to be there to help the model and check it’s work etc.
If you broke every task that humans do at all into a list, especially if you scaled by frequency, gpt-4 probably can in fact do 20 percent. (Frequency scaling means that easy tasks like “find all the goods in the store on the shopping list” happen millions of times per day while improving the bleeding edge of math is something few humans are doing)