Technical Context
I love moments like these, when hype suddenly collides with a very sobering assessment. In his conversation with Dwarkesh, Terence Tao essentially said what I see in engineering projects: LLMs are excellent at expanding the research perimeter, but they don't make the main intellectual leap for a person.
If you translate this into the language of AI automation, the picture is very clear. The model helps dig deeper into literature, gather more cross-references, write code for numerical checks faster, build graphs, and format material in a way no one would have had time for before.
Tao even formulated it almost perfectly: papers with AI become richer and broader, but not deeper. I like this phrase precisely because it removes the magic. Productivity grows, but the source of discovery is still in a person's head, often with a pen and paper.
And there's an important technical detail here that many miss. LLMs work quite well on local tasks with a fast cycle: summarization, draft code generation, finding similar ideas, formatting, numerical experiments, and iterating through options.
But when a task requires a long, cumulative trajectory, the model starts to struggle. Tao pointed this out very accurately: AI doesn't build on partial progress the way a human or a strong research pair does. It jumps, makes mistakes, jumps again, and the context and evolution of the idea often fall apart.
I would also add from my experience: even when a long context is formally available, it doesn't equate to a stable research state. The model lacks a genuine internal mechanism for scientific persistence. It doesn't hold onto a hypothesis for weeks or live inside a problem.
What This Changes for Business and Automation
This is where it gets interesting. If you're in charge of R&D, an analytics team, or AI integration into the research process, you shouldn't sell yourself the fairy tale of 'automated discovery.' The ROI is currently elsewhere.
It lies in throughput. A single team can test more hypotheses, review more literature, close off weak directions faster, package results more neatly, and not waste valuable researcher hours on mechanical work.
That's why the talk of 'LLMs as cheap research labor' sounds crude but points to a real shift. People are already using models as ultra-cheap research workforce: to generate options, run code, check wording, create additional graphs, find related papers, and prepare draft arguments.
The winners are those with strong verification. The losers are those who confuse volume with quality and start measuring progress by the number of generated hypotheses.
This, by the way, almost mirrors what I see in business outside of science. As soon as the cost of generating ideas and artifacts drops to near zero, the bottleneck shifts to verification, filtering, prioritization, and accountability for decisions.
That is precisely why I would build an AI solution for research teams not around a single 'smart chatbot' but around a pipeline. Separate agents for literature search, separate ones for code and numerical experiments, separate ones for normalizing notes, and a layer of quality control and logging on top.
Without this, it's very easy to create a beautiful factory of pseudo-knowledge. More papers, neater graphs, longer tables, but no real increase in understanding.
At Nahornyi AI Lab, we solve these kinds of problems for clients: we don't just bolt on an LLM, but we assemble a process so that automation with AI genuinely removes routine tasks without diluting scientific or product discipline. If your team is drowning in hypothesis testing, reviews, drafts, and repetitive research steps, let's analyze your workflow and build an AI architecture that accelerates people instead of creating another source of noise.