Why are the AI capabilities discussions so polarizing? People either see current AI as a net negative or net positive. But the gap is a chasm. Well ok maybe there are people who are uncertain too. Okay maybe it is task specific. But the edges are extremes. Why?
I suspect the split aligns, along whether believe in planning.
On the one hand, yes there are many AI aided workflows that produce very stable reliable outcomes. And that side is growing. But some people see the billion dollar single person start up around the corner.
Some capabilities
RAG search is great. It’s a great pattern that can be a good boost beyond just BM25. Plus you can better contextualize the result in English, say.
Then you can also use this concept for routing questions, tie them to tasks or skills, and boom you have an agent. The possibilities here can be endless perhaps. A solid general purpose paradigm shift to building software. You still need to build the skills. If you want them to be reliable and deterministic you should hand code them.
Code stack trace debugging . This one can be hit or miss depending on if the model has been fine tuned on the relevant issues . Maybe cosine similarity search helps too with RAG. Unclear.
Learning. Yes can really help.
Proof of concepts (POCs) and demos, to inspire you and them.
Code Gen? Here is our datacenter of geniuses? Or not. This feels very hit or miss!
Parallelizing everything with an army of agents? Maybe if the same task, like map reduce. But tasks with unknowns, this seems to break down without humans in the loop.
Nod to the mundane
I discussed the mundane here11. To sharpen this up, you might say there are a set of tasks like repetitive work and also they are laborious and the benefit in doing the manual labor if there were a better alternative is unclear. I remember there were a few times in my life, in high school and in college, where I would do research in a library by looking up materials using a computer terminal, and then physically walking to the various books from my results on the shelves and flipping through them, to judge their relevance. Apparently the physical wandering in a circle is embedded in the original Latin -circare12 meaning. Is there a benefit to the fiber, the slow pace of research? Search is so much faster we dropped the “re-”! I wouldn’t necessarily want to go back to that kind of pace, but it definitely feels more rewarding to find the thing you were looking for if you had to physically walk to it took more time among the false positives.
I also tend to think through the transition from spiral bound maps and Map Quest down to Google-mapping your way to a destination. When I was young, I used to often be a car navigator during family trips before learning to drive. I would follow along on the map, flipping pages as needed, to make sure we took the right turns, well unless they were left turns of course. There was some missing of exits on highways and even some talking to strangers when even maps were not clear enough. Today I admit I use GPS navigation for nearly all trips. If I’m driving alone, I enjoy how the road lets me completely empty my mind. Instead of focusing on not missing a turn, say, I get to just handle the steering wheel and really feel just basal reflexes in complete silence. Now when driving becomes illegal because manual driving is seen as wreckless, that will be quite a sad day.
The Brain Drain has already hit us before?
Maybe we have been fighting brain drain aka skill shift, since frameworks and tools were a thing? You no longer need heavy math skills or software skills to build ML systems. (A coworker mentioned the bootcamp effect). Some have said other than frameworks tue skills have more focus on domain knowledge, yet deep learning has taken that away too, with the bitter lesson.
Maybe systems thinking is the only thing tuats left? Clean modular design? Simplicty, entropy management?
Reality is the Bottleneck
I am sure someone has already written about this in complexity theory, or chaos theory, but I think humans are at a theoretical upper bound for intelligence and impact as constrained by the real world. Got reminded of this through the recursive self improvement topic [2] Carl discusses.
Back to the adage that information is not enough for success else everyone would be a six pack ab wielding millionaire. Reality is the bottleneck. Defining intelligence is squishy but I threw in impact there too because no matter how good a model is, your brain or a statistical model, it will be wrong.
All models are wrong. Some are useful.
Not really reasoning
There are not enough atoms in the universe to build compute to predict the correct path through a complex problem. And path is the operative word here. Reasoning models have a very strange name because they dont “think”, they do. They one shot write plans then run them7. They will course correct too of course but they are doing it in reality not in a self contained logic center.
They should really be called while -looped models, though code-harnessed , another recent term, is more fitting.
The Age of Reason
What is a better definition of reasoning? It is using logic and thought experiments and armchair testing ideas. It is thinking out loud. Actually that’s probably why reasoning models got their name. Because of the chain of thought. And maybe I should go back to the previous paragraph and distinguish between an active agentic,–touch and smell the world–, and chain of thought.
So thinking out loud helps people I think in a different way than the mimickry of what we see in LLM use. We have seen CoT not nedessarily having a benefit8,9.
P vs NP Problem
Everything may change if the P vs NP Problem is solved. If a problem can be easily verified, is there a way to solve it in general?
As an aside , LLMs feign solving P vs NP because they clearly produce answers in linear time, a consistent token rate. And that’s why jokingly Yann Lecun has compared LLMs to giant lookup tables. But yes the look up table cheat code is a cheat code to P vs NP.
References
https://youtu.be/AkadGXzDqBw Carl on recursive self improvement
https://michal.piekarczyk.xyz/note/2026-05-01-odsc-closing-notes/
halting problem
no free lunch
https://michal.piekarczyk.xyz/note/2026-03-04-prompt-driven-development/#reasoning
https://arxiv.org/pdf/2504.00294v1 inference time scaling
https://youtu.be/ShusuVq32hc , George Montañez , on reasoning
https://michal.piekarczyk.xyz/note/2026-05-12--merging-is-hard/
https://michal.piekarczyk.xyz/note/2026-05-12--merging-is-hard/#the-mundanity-vs-the-novelty