Neurosymbolic is a good direction. People have been profitably mixing LLM with ontologies, etc for some time. Commercially. It's just not 'trendy' yet.
Systems like active inference (either message passing on bayesian nets, or Friston-esque free energy minimization, or other arer better, faster, cheaper than LLM goo. Also not quite trendy yet, but getting there.
Models without any discussion of underlying methods or principles feel like advertising. Which is fine. But makes for thinner discussion. Good luck.
Acknowledged. I’ll look harder. Perhaps just a short prose issue — not backprop, not knowledge graphs, not rube goldberg LLMs, hard to tell what it is.
But this is a drive-by hipshot, acknowledged, and I hope to see more.
All fundamental unsolved problems of AGI are concentrated in two elements of the scheme: 'Concept Formation' and 'Pattern Learning'.
For both, 'neuro-components' involvement in the design does not promise anything new. The search for patterns (and, in particular, the search for cause-and-effect relationships), essentially a combinatorial task, requires the generation of hypotheses with their subsequent testing. Finding a niche here for the useful use of a neural network is at least difficult.
In the task of forming concepts, neural networks can be used as a detector of those predetermined primitive basic features of the observed reality - but this is quite feasible in other ("classical") ways.
It is interesting that all the right words were said, but when it comes to implementation, it is proprietary.
I am guessing this is a specialized solution for certain problems, when all quantities involved can be well modeled. Likely we will need some of these, and each domain will have to have its own abstractions and calculus.
Not at all specialized. Quite the opposite. It is AGI, a system that can potentially learn any knowledge or skill that a human can -- using no more training material or time than a human. Like humans, it is also a great tool user.
No, not early stages. We've already commercialized early versions of it and are now doing the final scaling to full adult-level intelligence. Only about 100 man-years of work left.
It sounds great (except the part about Neural Nets and a few others). How does it compare with English? To be cruel, English hits it out of the park. You have described its abilities using English - wouled you care to describe the abilities of English using INSA? Would you recommend that people learn INSA instead of English? If notwhy not? (Sorry, I mean any natural language used in an advanced society to describe new technology). What are you doing with figurative language - "raised the bar", "a bridge too far"? English is used by a billion people - how many people will end up using INSA? Will we be back to the bad old days of people effewctively programming things they don't understand? When you talk about Neural Networks, are you talking about directed resistor networks, or eal neual networks, that can turn themselves inside out. Are you committed to handling mental states - the hardest part of AGI will be explaining to a person why something is a good idea when they can''t understand why, because of their limitations. A deep knowledge of mental states (including irrationality and psychosis) will be necessary. If INSA is intended for a small subset, that;'s fine, but you need to say so.
Neurosymbolic is a good direction. People have been profitably mixing LLM with ontologies, etc for some time. Commercially. It's just not 'trendy' yet.
Systems like active inference (either message passing on bayesian nets, or Friston-esque free energy minimization, or other arer better, faster, cheaper than LLM goo. Also not quite trendy yet, but getting there.
Models without any discussion of underlying methods or principles feel like advertising. Which is fine. But makes for thinner discussion. Good luck.
Not sure what you mean - lots of details in my various articles and whitepapers (links within articles)
https://aigo.ai/articles-white-papers/#
Acknowledged. I’ll look harder. Perhaps just a short prose issue — not backprop, not knowledge graphs, not rube goldberg LLMs, hard to tell what it is.
But this is a drive-by hipshot, acknowledged, and I hope to see more.
All fundamental unsolved problems of AGI are concentrated in two elements of the scheme: 'Concept Formation' and 'Pattern Learning'.
For both, 'neuro-components' involvement in the design does not promise anything new. The search for patterns (and, in particular, the search for cause-and-effect relationships), essentially a combinatorial task, requires the generation of hypotheses with their subsequent testing. Finding a niche here for the useful use of a neural network is at least difficult.
In the task of forming concepts, neural networks can be used as a detector of those predetermined primitive basic features of the observed reality - but this is quite feasible in other ("classical") ways.
Love this. I've thought about universal combinatorial primitives
It is interesting that all the right words were said, but when it comes to implementation, it is proprietary.
I am guessing this is a specialized solution for certain problems, when all quantities involved can be well modeled. Likely we will need some of these, and each domain will have to have its own abstractions and calculus.
Not at all specialized. Quite the opposite. It is AGI, a system that can potentially learn any knowledge or skill that a human can -- using no more training material or time than a human. Like humans, it is also a great tool user.
In that case, since it did not take the world by storm yet, it is likely in early stages.
No, not early stages. We've already commercialized early versions of it and are now doing the final scaling to full adult-level intelligence. Only about 100 man-years of work left.
How do we get our hands on it?
What does it cost?
Is there a service portal where anyone can interact with one of these?
Hi Peter, is your graph RDF based? Does it use OWL? Curious.
Bette yet, could you please share a link that does describe the graph framework?
No, not RDF or OWL - we needed something more flexible and fine-grained. Unfortunately we don't have more pubic details.
Interesting
It sounds great (except the part about Neural Nets and a few others). How does it compare with English? To be cruel, English hits it out of the park. You have described its abilities using English - wouled you care to describe the abilities of English using INSA? Would you recommend that people learn INSA instead of English? If notwhy not? (Sorry, I mean any natural language used in an advanced society to describe new technology). What are you doing with figurative language - "raised the bar", "a bridge too far"? English is used by a billion people - how many people will end up using INSA? Will we be back to the bad old days of people effewctively programming things they don't understand? When you talk about Neural Networks, are you talking about directed resistor networks, or eal neual networks, that can turn themselves inside out. Are you committed to handling mental states - the hardest part of AGI will be explaining to a person why something is a good idea when they can''t understand why, because of their limitations. A deep knowledge of mental states (including irrationality and psychosis) will be necessary. If INSA is intended for a small subset, that;'s fine, but you need to say so.
INSA is not a language. The system operated with visual and mouse, language input/ output