
@TauLogicAI Founder & CTO. The logical way to AI! -- I rarely check my DMs, so better DM @TauLogicAI
Joined March 2009
- Tweets143
- Following2
- Followers958
- Likes20
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.
New paper
arxiv.org/abs/2605.12539
This tweet is +1y old, but I've said it for much longer. ChatGPT came years ago, and "AI" is still more or less the same. Against all "expert" predictions, it didn't cure cancer, it didn't colonize Mars, it didn't discover new physics, it didn't really improve much. Because I was right all along: machine learning is around its peak, and it is a huge bubble. Nowadays it's obvious, but back then I was a single voice against the whole world. Not only that machine learning is fundamentally incapable of logical reasoning, but the architecture of those giant data centers, is also incapable of logical reasoning. And I keep telling you: the real deal is Logical AI, and we are the leaders of this segment. Machine learning is only for translation.
1. Yes. Machine learning models are more or less at their peak, and the bubble is about to pop.
2. Yes and no. Although it can be mathematically proven that there are things impossible to do, it is also proven that for any given intelligence ability, there's another one above. But the truly exceptional abilities got nothing to do with machine learning, but with logic.
3. Easy. Humans are not that smart at all. The only thing we understand better is human nature.
The path to AGI is in new methods regarding Atomless Boolean Algebras. And those are all patented. So you are stuck with me, sorry.
Really really funny. Not even close. Maybe his analogy holds for extremely simple logical systems, far from enough for software specification. He thinks he's the first to find the connection between Datalog (polynomial time) and matrix multiplication (while seemingly he missed expressing recursion using matrix inverse). Can kill his argument using undergrad complexity theory!
I've found the path to AGI:
arxiv.org/abs/2510.12269
Ohad Asor retweeted
Machine learning models / LLMs excel at patterns but will never offer logical correctness for non-trivial/complex problems.
I'm excited about formal software synthesis from logical requirements, where correctness is guaranteed by construction rather than hoped for.
Light years away, not even close and doesn't even say anything. LLM can't even start helping to code the Tau language.
teapawt.org/i/grok/share/t7KF7b3DY…
Quantum mechanics works very well (except when it doesn't), but whenever a quantum theory is illogical, they say "find a better logic". LLM works very well (except when it doesn't), but whenever it is illogical, they say "what is logic? there isn't such a thing". Coincidence? No. Human nature, yes. Same human nature that made humanity susceptible to religion.
Show me one blockchain project which is not just more of the same thing. Show me one AI project which is not just more of the same thing. I'll show you one that ticks both boxes: @TauLogicAI $AGRS. Read my two recent articles here to see why it's the case
Done. In the next few weeks we'll make some bug fixes and performance improvements, and most importantly: pointwise revision. This will demonstrate an interactive machine that can fulfill commands, in a way never seen before.
In the slightly longer term: additional language features (esp. tables) to allow a more realistic swdev language.
Asked ChatGPT "why can't machine learning perform logical reasoning?" and got a good answer! Too long to paste here, so try it yourself
I bet that currently no field is full of misconceptions, fake news, and outright lies, as much as the fields of AI and Blockchain are, each. And I live in both those worlds. But I will never succumb to falsity.
"The results show that a four-way interaction is difficult even for experienced adults to process without external aids."
Yes, we humans suck. But LLMs suck even more: you can do better than LLM with only some minutes of work. The only way to surpass the human capabilities is logical AI.
researchgate.net/publication…
At least they don't go backward, but they still go to the left instead of forward:
1. If first order logic theorem provers could practically verify nontrivial programs, they'd revolutionize the swdev world long ago. In fact the only reasonable swspec lang, not only for verification but also synthesis, is our Tau language.
2. The way to combine ML and logic is the other way around: ML should be expressed in logic, rather than logical engines embedded in ML.







