On the one hand this is interesting to me because it's closely aligned with my own ideas about what will be the general form of the first AGI systems.
On the other hand, just like with my own ideas, there's very little concrete evidence that they're correct. It looks like the author has been very careful about wording and avoids saying directly "there exists an implementation of these ideas that seems to be working well enough to expect it to scale".
This is so retro. I went through Stanford CS in the 1980s, just as the "expert system" boom was collapsing, and the "AI winter" was beginning. This is very close to the hype of that era.
Many of the Stanford faculty really believed that artificial general intelligence via expert systems was close. If you could just hammer the real world into predicate calculus...
It might be worth revisiting, though. Embedding-based AI may be reaching its limits. If you need to plan or subgoal or compute, it's not quite the right tool for the job. Some other representation may be needed.
(Consider EMACS "org mode". It probably isn't that hard to get an LLM to translate a question into "org mode" form. Then you need a strategy module to decide which subtasks to work on, notice when progress is being made and when it isn't, work on the problem as a tree of subgoals, and combine into a result.)
Can't blame them too much since every other paper points to the fact that more parameters and data is the way forward to new SOTA results.
We have probably reached the limit of scraping publicly available English training data from the Internet. OpenAI is betting on synthetic data. Let's see where that takes them. The sad part is companies like OpenAI only makes money if the answer to AGI is large models and large data. I don't know if they have unknowingly restricted their ideas on AGI based on that
I'm too dense to grasp the site's mathematical propositions (I don't have the mathematical grounding to understand Borel algebra not to speak of Borel Hierarchies). So it was no wonder I had a lot of trouble discerning how what it proposes was materially different from expert systems, inductive logic systems, and Cyc. The associative memory layer seems to be picking up some lessons learned from Generative AI, but I'm not clear on that. It will be interesting to keep an eye on this project and where it takes us, and see if it makes more sense when someone comes along and dumbs it down for Blubs like me.
I appreciate the author’s attempt at avoiding commodification of the model but at this point there’s little point of discussing the pretty web page. Maybe resubmit the paper when it’s released?
Seems like a good idea, but "the proof is in the puddling". Can the author make a working implementation that competes with Transformers in terms of quality?
"I am Alex Levy, an independent researcher and AGI enthusiast"
Background appears to be exclusively in UI/UX design which shows because the website is very nice. But forgive my skepticism on everything else the page talks about. Might get some VC $ for the enthusiasm but not seeing any firepower behind the claims.
> Our goal is to replace humans in any domain where data work is required: data collection, preparation, analysis, and interpretation.
Nice that they say up front what their goal is. Not to help humans or make humans more productive. They explicitly want to make humans obsolete. I hope they fail.
The site is ~3000 words (half of Attention Is All You Need, as a measuring stick) of Eye-Catching Summary Statement\nShort Explanatory Paragraph statements like the ones GPT-4 is so fond of generating, and has no formal design model to reason about, no justification on why the design is right, no information on training, no evaluation results, and no code.
I can’t help but feel it’s a shame to have published this draft (I hesitate to call it a preprint) before any of the factual content of the article was suitable for publication, since if these improvements are as effective and profound as the summary suggests, you’re doing yourself and your work a disservice by advertising it in a shallow way.
> It has been shown that LLMs are unable to learn concepts beyond the first level of the Borel Hierarchy, which imposes severe limits on the ability of LMs, both large and small, to capture many aspects of linguistic meaning. This means that LLMs will continue to operate without formal guarantees on tasks that require entailments and deep linguistic understanding.
I would tend to disagree with this excerpt and the paper it came from. I am not familiar with the "Borel Hierarchy", but from the paper it seems it is just that LLMs cannot make a guarantee that they interpret "every", "all", etc properly. I don't want to bother trying to decipher all of the math, but it seems highly questionable that they proved it "cannot be learned". The experimental section is greatly lacking in data points.
If you disagree and think the paper is good, please do explain.
29 comments
[ 3.6 ms ] story [ 83.3 ms ] threadOn the other hand, just like with my own ideas, there's very little concrete evidence that they're correct. It looks like the author has been very careful about wording and avoids saying directly "there exists an implementation of these ideas that seems to be working well enough to expect it to scale".
* Paper - coming soon.
* Github repository - coming soon.
* Details about how it works - coming soon.
* Hype - here now.
This is so retro. I went through Stanford CS in the 1980s, just as the "expert system" boom was collapsing, and the "AI winter" was beginning. This is very close to the hype of that era. Many of the Stanford faculty really believed that artificial general intelligence via expert systems was close. If you could just hammer the real world into predicate calculus...
It might be worth revisiting, though. Embedding-based AI may be reaching its limits. If you need to plan or subgoal or compute, it's not quite the right tool for the job. Some other representation may be needed.
(Consider EMACS "org mode". It probably isn't that hard to get an LLM to translate a question into "org mode" form. Then you need a strategy module to decide which subtasks to work on, notice when progress is being made and when it isn't, work on the problem as a tree of subgoals, and combine into a result.)
We have probably reached the limit of scraping publicly available English training data from the Internet. OpenAI is betting on synthetic data. Let's see where that takes them. The sad part is companies like OpenAI only makes money if the answer to AGI is large models and large data. I don't know if they have unknowingly restricted their ideas on AGI based on that
I'm too dense to grasp the site's mathematical propositions (I don't have the mathematical grounding to understand Borel algebra not to speak of Borel Hierarchies). So it was no wonder I had a lot of trouble discerning how what it proposes was materially different from expert systems, inductive logic systems, and Cyc. The associative memory layer seems to be picking up some lessons learned from Generative AI, but I'm not clear on that. It will be interesting to keep an eye on this project and where it takes us, and see if it makes more sense when someone comes along and dumbs it down for Blubs like me.
https://ogma.framer.website/use-cases
National Security & Fraud Management
Social Networks Monitoring
Surveillance of User's Activity on the Internet
Influencing People's Social Trajectories
Human Resources
I appreciate the author’s attempt at avoiding commodification of the model but at this point there’s little point of discussing the pretty web page. Maybe resubmit the paper when it’s released?
Background appears to be exclusively in UI/UX design which shows because the website is very nice. But forgive my skepticism on everything else the page talks about. Might get some VC $ for the enthusiasm but not seeing any firepower behind the claims.
Nice that they say up front what their goal is. Not to help humans or make humans more productive. They explicitly want to make humans obsolete. I hope they fail.
I can’t help but feel it’s a shame to have published this draft (I hesitate to call it a preprint) before any of the factual content of the article was suitable for publication, since if these improvements are as effective and profound as the summary suggests, you’re doing yourself and your work a disservice by advertising it in a shallow way.
I would tend to disagree with this excerpt and the paper it came from. I am not familiar with the "Borel Hierarchy", but from the paper it seems it is just that LLMs cannot make a guarantee that they interpret "every", "all", etc properly. I don't want to bother trying to decipher all of the math, but it seems highly questionable that they proved it "cannot be learned". The experimental section is greatly lacking in data points.
If you disagree and think the paper is good, please do explain.
I guess "coming soon"'s time horizon just got a lot longer