My thoughts and feelings around AI are mixed. It has the power to become an amazing tool in all aspects of life, but society seem to be completely oblivious how to handle such a large hammer.

AI and society

AI can potentially automate most of the office jobs which require very little effort to get done, this will very likely result in mass layoffs and improve the efficiency at which companies operate. What once took 2-3 business days, now may take 10-15 minutes to percolate through all the pipes and systems. This affect all the paperwork related jobs, massive downsizing of the teams.

This is the ultimate finance operations (finops) dream for all the corporations and big businesses. We, as a society, don’t fully know how to cope with it. I think the governments around the world will have to enforce some form of an AI tax, as it is obvious that profit maximizing goes against the societal needs for people to work and sustain themselves somehow.

One such solution is to find a half-way point between firing 80% of your staff and paying extortionate taxes. That seems like a viable option. Another solution is to keep more people employed but turn 40 hour work week into something like a 25 hour work week (or even more flexible terms). Now this has impact on salaries, quality of life and so on, but I think if it is carefully thought through it can be actually beneficial to society.

The aim should be for the government to enforce a symbiosis between companies, people, and the AI. It will require some really deep rethinking about what “full time” actually means and whether the concept of “full time” in itself is not an outdated principle.

AI and Software engineering

The AI wave has also affected the tech profession and adjacent fields (e.g. graphics designers). Although I am not convinced the current state of the art LLM models can be called a replacement by any means, it can become a very handy extension for every professional. However, that doesn’t really translate into what a software engineer has to do outside of programming. I would say that a software engineer spends maybe 50% programming, but not more. In the spirit of the Amdahl’s law, we can optimize the 50% of the time a software engineer spends programming, the LLMs will not be able to help optimize the remaining 50% which are not related directly to programming.

Another trope which I is on the horizon is replacing the junior engineering roles with the LLMs, and I think that’s a horrible approach to maintaining the continuity within your company. If you don’t hire juniors who can stick around for years or even decades and you retire the old engineers, then you are burning the candle from both ends.

Companies like Google and Meta have been on a firing spree, replacing the capable seasoned veterans with new joiners and LLMs. There is no replacement for tribal knowledge and years of experience. Something that no LLM will be able to provide, ever. Even the best LLM of the future will not be able to upskill and mentor more junior engineers.

This process is untenable and makes corporations like Google just too prone to ultimate breakdown, and maybe that’s a good thing. I will drill deeper into AI in the software engineering another time.

AI and outside-of-the-box thinking

The current state of the art of the AI - the LLMs, are nothing more but some clever stochastic machines comprising of huge vectors and operations on them. As the joke goes “the I in LLM stands for intelligence”, there is nothing intelligent about LLMs.

During my times at university, one of my professors, a logician, who was teaching us mathematical logic has once said that no algorithm can increase the (information) entropy. So, although, presumably, any LLM model has a huge information entropy, any operation (algorithm) acting on the LLM will generate output which is less or equal in the entropy. Hence the LLM model is limited and behaves nicely as long as it plays within it’s own “boundary”, once the LLM leaves the boundary, hallucinations and weird behaviour (divergences) occur.

Hence, what we think is amazing, and so novel about AI, is just AI traversing the huge vector space and connecting the dots within it, so-to-speak, creating just ever-so-slight combinations of the same picture using the stochastic processes. Although from this fairly reductionist point-of-view on the LLMs, it describes the behaviour very nicely.

Another story about the same professor of mathematical logic was the computer-aided proofs. You can have the set of theorems and lemmas which have been proven by mathematicians for thousands of years and still not be able to prove anything new. All these theorems and lemmas spring up from axioms and observations on the completely fundamental level - many axioms are deemed to be so obviously true that we proclaim them as the gospel. Then we create a directed graph, from axioms through observations and trivial lemmas up to some very abstract concepts that are sometimes very hard to grasp for people.

Now, the LLMs have introduced a technique called “Chain-of-Thought”, or “given this, what we can derive from it”. If you have 1000 lemmas and theorems in the set T (theory), to prove something new means picking a minimum number of lemmas and theorems and prove something which is not already in the set T. Now you are getting into the problem of combinatorics and searching through the vast space of combinations of lemmas and theorems which will be suffice to prove a hypothesis.

Another problem with the “Chain of Thought” are the potential cycles in theorems and lemmas. These cycles might be trivial to spot, or maybe not! These cycles might be 40-50 edges long, who knows! The model of “Chain-of-Thought” may not be good enough for highly abstract thinking and theorem proving.

On top of it, if the statement is undecidable or even worse, unprovable, the AI may run out of memory before it delivers any output to you.

The proper out-of-the-box thinking seems to be fairly elusive and fairly non-trivial task to model with the current AI models. I think the LLM/AI will be able to connect the dots in the sea of the mathematical theorems and lemmas, but it will not be able to venture outside of the box. That needs more than just computational power.

For the foreseeable future, I don’t see mathematicians being outwitted by a machine, though I do expect some initial success in the computer-aided proofs, it will subside with the time.