On Tue, 29 Sept 2026, 02:02 Matt Mahoney, <[email protected]> wrote:
> Hi YKY! Do you have any experimental results for your TicTacToe or Tiny > Stories tests? I think a paper focused just on that would have a greater > chance of acceptance than a paper with lots of ideas but no supporting > evidence. How does your system compare with LLMs asked to solve the > same problems? > I have open-source code and the superiority of Logic Transformer has been verified several times by different tests. (But the repos and documentation are currently rather disorganized). I'm very confident of the breakthrough and will try to publish better-written papers. But that's not what I'm complaining about, which is racism. The reason I'm so slow to reply is that I *know* I'm talking to a bunch of stubborn people who will refuse to face the problem of racism till the bitter end. If Singularity is next Monday, they will remain racist till 11:59:59 on Sunday. We can add a clause to any company's constitution or policies, that all business processes are open and transparent, and eliminates discrimination, which can be subject to AI-assisted retroactive adjudication if disputes arise. I will do that in my current DAO, but other AI companies may not adopt the same. In my vision for the near future, the companies that refuse to add the anti-racist clause will get very bad reputation and they will eventually succumb to public pressure. Someone needs to make a start in this direction.... I see you are using Copilot to develop your code. An obvious application of > a hybrid symbolic neural approach would be to ask Copilot to solve a > problem and let it write the code to solve the part that can be done > more efficiently with a symbolic approach. Simple math is like that. Can > you do the same with logic? Can you demonstrate the advantages of your > approach on existing benchmarks or non toy problems? > > -- Matt Mahoney > It succeeded on Tic-tac-toe and outperformed traditional Transformers, that was when I realized I was on to something. Later tests with natural language corpuses demonstrate beyond simple games. I think I'm ready to train a large model from scratch, but I'm still looking for helpers and busy with the political and DAO stuff. https://github.com/Cybernetic1/RL-TTT-experiments https://github.com/Cybernetic1/AGI-experiments https://github.com/Cybernetic1/AGI-prototype ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/T8c5b5a041e72e8cd-M98cba54c18833956a577eef0 Delivery options: https://agi.topicbox.com/groups/agi/subscription
