As I write this in mid/late 2026, a Doomerism narrative around the outcomes of AI development has gripped the nation, and the majority of people are starting to fear that AI is on course for the dystopian, rather than utopian, outcomes. Bans on Data Centers are spreading, politicians are calling for halts on super intelligence, and even executives in the leading labs are demanding government enforcement to prevent smart models from proliferating.

It’s a complicated situation, and there are various buried motives at play no doubt (curious coincidence that a government body enforcing model limitations would have the side effect of killing open source, which is growing very rapidly, as well as codifying an enormous moat around the few large players and blocking anyone else from entering). But it centers around a concept called RSI. Recursive Self Improvement.

Current teams of human engineers are working hand-in-hand with LLM models to improve each iteration of model. Improvements come from training techniques, training data quality, and other architecture optimizations, discoveries, and concepts. At some point, it is believed, AI will become smart and capable enough that the human-llm partnership will ultimately be slower and less efficient that a pure llm-development cycle. At this point the machine will be more capable of improving each of the inputs: training techniques, training data, and architecture. The resulting system, by definition, will be even more capable of improving its successor, and so on.

Once this happens, a feedback loop kicks off that pushes intelligence to infinity, vastly eclipsing humans, and by such a gap we can never assert control over the resulting thing again.

In my humble opinion, this is likely bullshit. I’ll put this forth as my asymptote theory. So here’s the thing: The current LLM architecture is trained on entirely on human knowledge (or derivative synthetic content, that comes from human-knowledge built systems), which means the models we select during training are the ones that have the fullest and best internal representation of that human knowledge corpus.

Right off the bat, there is no reason to think that such a process could exceed human knowledge in any given discipline… with two exceptions.

First, it would be expected that a perfect LLM model would reach a level of expert knowledge equivalent to the very best expert in a given field; however, unlike humans who are typically restricted to a single field of maximum expertise, the system could reach that equivalent across ALL fields. There are very few frontier level mathematicians who are also frontier level biologists. This is a huge advantage, sure, but it does not represent a path to infinity. Instead, it represents a raising of the asymptote to the cumulative sum of acquired human knowledge and the connections that might exist between disciplines that are hitherto undiscovered.

The second exception is time. You can run thousands of agents and do the equivalent of thousands of years of human work without ever exceeding human knowledge, and still accomplish incredible things. This is the case for OpenAI’s recent Navier–Stokes equation solve. They ran 10,000 agents over a hundred billion tokens, which was the equivalent of something like ~5,000 human years of research. It had no mathematical skill above human mathematicians, it just had a lot of time. There was nothing super-human about this, and humans probably could have solved it in much less time, but they nonetheless are outperformed by a less-intelligent system that has a massive time advantage.

So AI has the advantage of time leverage and the advantage of cross connections between disciplines. This is an advantage to be sure, but it’s still represents a hard asymptote at some low multiple above the sum of present human knowledge. Recursive self improvement is therefore not a recursive takeoff to infinity, it is a recursive approach towards an asymptote, which will result in ever diminishing returns.

What would change this is a breakthrough in architecture, or the ability to gather new knowledge by running novel experiments and reviewing the data. The first of those remains in the real of possibility, but there’s frankly no obvious path for how such a system could hit the point of recursion without humans in the loop. Remember it all starts with training and training data. Those runs are setup by people, and are not connected to the LLMs. Separate servers, separate workflows, separate everything. And the second of those is far out of the realm of possibility. When we have research labs setup with robots run by LLM systems without oversight, then it’s time to be worried, but at present every single step of the way requires human approvals, meetings with board members, politics, liability conversations with lawyers, physically plugging and unplugging things… thousands of steps that MUST move at human-time speeds and presently MUST involve humans.

The realistic scenarios where a singularity of RSI takes off to infinity and the system is able to jump from the digital realm to the physical realm in a non-trivial way is not realistic. Many major critical infrastructure systems have hard offline backups built in (left overs from cold war prep) that are explicitly non-hackable. Human manipulation is a real thing, but not some magic bullet either. Human systems have plenty of security built in to protect against malign actors.

The doomerism narrative is the kind of vague fear that we see whenever our status quo is changing in a big way, and it certainly gains wind from the many scifi stories that imagine our downfall to the machines, but realistically speaking, given the present composition of the world and our infrastructure and datacenters and AI development processes, it’s just not going to happen.

LLM Intelligence with RSI is unlikely to explode to infinity and more likely to find ever-diminishing returns towards an asymptote somewhere slightly above human knowledge. Even if massive breakthroughs are visible from that vantage, there is no path that lets it bypass humans in any real way to take over the world or kill us. Not until some far and imagined future where people have forgotten all these worries and handed the keys to our major systems to AI models, but that’s just not in the zeitgeist and not anywhere in the cards for decades.

Why does this matter? Because AI is actually a lot more than LLMs, and I really worry that in our present society where nuance is a forgotten concept, this is getting buried. In the rush to declare chat bots are going to destroy the world, we risk abandoning some real AI tools that have already brought a number of benefits, including:

• AlphaFold solved protein folding, which evaded human researchers for years, and has already aided in malaria vaccine research and will undoubtedly lead to hundreds of other therapeutic interventions that will save lives. Built on Evoformer, which has similarities to LLM’s architecture.
• MIT’s Mirai has enabled prediction of breast cancer up to 5 years earlier, which vastly improves outcomes. This exceeds human abilities, and other AI models are showing similar results on other cancer types.
• Accessibility for deaf/hard-of-hearing is undergoing a massive expansion with STT AI systems. YouTube can auto-generate captions, as can Zoom/MSTeams, and it’s coming to everything as AI has pushed the cost of this process so low. Transformers + CNNs.
• TREWS is an AI-powered (classic ML in this case, not transformers/NN) sepsis detection system now in wide use in hospitals. A two-year, multi-site study of 600,000 patients across five hospitals associated TREWS with a ~20% reduction in sepsis-related mortality because it could identify sepsis hours before doctors could see it.
• GenCast for disaster preparedness / extreme weather event prediction outperforms the European Centre for Medium-Range Weather Forecasts’ (ECMWF) Ensemble Prediction System (ENS) — the previous gold standard for this — 97% of the time. That’s a pretty big improvement. This is built on diffusion + transformers + attention (so essentially the same as the image generators like MidJourney).

This list goes on and on, and the ceiling for these kinds of targeted AI tools is not currently in sight. This stuff has real world beneficial impacts and real life saving outcomes that we are already seeing (not some vague techno-optimism).

Don’t give in to the doomerism. It’s not real. Asymptotes exist in the current architecture and training techniques. Fundamental breakthroughs in approach are possible, but there’s no evidence for that presently. Jumping from digital to physical is not some trivial step easily accomplished by a super-intelligence, unless we hand over a huge number of keys that nobody’s even considering handing to anything at present.