Minh Nhat Nguyen

I think AI will split into two economically distinct layers. Cheap, good-enough models will handle routine work, while expensive frontier systems may be most valuable as autonomous research machinery. If those systems can run experiments, evaluate results, write code, and iterate with limited supervision, frontier labs may increasingly resemble automated research labs rather than ordinary software companies. That does not mean more generated work automatically becomes meaningful progress. AI makes it very easy to produce code, papers, experiments, and polished-looking activity. It can increase useful output, but it can also make pointless work feel productive. The scarce resource shifts from producing things to choosing worthwhile goals and judging whether the apparent progress is real. We already see reasons to worry about agent-training instability and systems whose confidence outruns their reliability. There are also less glamorous failure modes. AI-generated insecure software can create attack surfaces, while stronger models can assist motivated attackers, making theft of valuable lab secrets a serious risk. Scientific communication can similarly be polluted by cheap, low-quality papers repeatedly resubmitted across venues. Finally, I would resist collapsing all of this into AGI, ASI, or RSI branding. Those terms should name distinct claims, not serve as interchangeable corporate labels. AI’s future will be easier to reason about if we describe concrete capabilities, incentives, and failure modes instead of letting grand terminology do the thinking for us.

Overall, I expect AI to increase useful output substantially, especially in research, but not to translate cleanly into equivalent social or scientific progress. The upside is real: frontier systems could accelerate experimentation, coding, and iterative discovery, while cheaper models make routine capabilities broadly available. The harms are not merely hypothetical catastrophe. They include insecure generated software, stronger intrusion capabilities, theft of valuable research secrets, polluted publication channels, and enormous volumes of polished but pointless work. AI lowers the cost of producing both useful artifacts and convincing junk. So my expectation is mixed but transformative. The central bottleneck becomes judgment: selecting worthwhile goals, designing reliable evaluations, and distinguishing genuine progress from activity that only looks productive. I would not compress that into a precise net-positive or net-negative forecast, because the balance depends heavily on where these systems are deployed and what incentives shape their use.

The biggest update would come from evidence about whether autonomous research systems can reliably produce genuine discoveries rather than merely more plausible-looking work. If frontier agents could sustain long research loops—choosing useful questions, running experiments, detecting their own mistakes, and generating results that survive independent scrutiny—that would push me toward a much larger positive impact. The opposite finding would matter just as much: if scaling and improved training still leave agents unstable, overconfident, reward-hacking, or unable to distinguish meaningful progress from polished noise, I would downgrade the automated-research-lab picture substantially. Likewise, a major AI-enabled theft or security failure could show that deployment risks are arriving faster than the research benefits. So I would update most on measured outcomes in real research environments, not on another model launch, benchmark jump, or freshly diluted “superintelligence” slogan.