The success of LLM is undeniable, and does not match my earlier predictions. I already admitted that on several occasions over this year, and I'll do it here too now. My mathematical analysis of the capabilities (or lack thereof) of machine learning and of the transformer algorithm, is correct. What is mathematically impossible, remains so. However there are two things that I didn't take into account. One is the huge computational resources invested in LLM: if you take a bad algorithm (i.e. with very low complexity-theoretic power and even then with far-from-optimal implementation) yet give it so much compute time and space, it will give nice results. The second is a real shame: again I made the classical mistake of overestimating humans, forgetting that doing better than humans, is not that hard, and this is pretty much where current AI stands. It works similarly to humans, i.e. broad but not very deep. Logical solvers, on the other hand, can go exponentially deeper but on narrower scopes. Transformers proved an important theoretical point: a cap on what does it take to be fluent in natural language. Now that we know that, we can try to do better by many other means. This is an effort I started a few months ago, taking core lessons from transformers and studying them in frameworks that spend less on the unknown, and by that I mean: transformers give a very abstract skeleton of how things should work (quadratic forms over the history, then MLP, and repeat), then using SGD to tune all its many parameters. The part to learn (in the machine learning sense) is huge. I'm working under the guideline of "solve more, learn less" by having a skeleton that leaves much less degrees of freedom, using lessons learned from transformers, and using the expressive power of the logical frameworks I'm dealing with over the last years. To this end I also formulated the basis of theoretical machine learning in those frameworks, suggesting new aspects and giving rise to a fruitful combonation of the approaches. Empirically, some of the new settings managed to beat small transformers on lower wall-time even though they ran on CPU while the transformers ran on GPU.