I expect rapid progress but not towards general superintelligence

Interconnects published an analysis arguing that rapid AI progress is coming over the next few years, but that it will be concentrated in engineering and infrastructure rather than in a leap toward general superintelligence. The author notes repeated surprise at hearing top industry researchers say models will outperform them at their own jobs within a few years, and traces that expectation to the acceleration researchers observe in the infrastructure and engineering capabilities of models. The piece agrees models will be superhuman distributed GPU engineers before long, and argues that this makes experimentation and tinkering with model formulations far easier without making the models dramatically different in nature.

The framing the author uses is that research and engineering have traded places before. Before deep learning took off, AI work was more of a research endeavor; today, the strongest researchers are judged largely on implementing and scaling ideas in complex infrastructure. The claim is that the field is entering another such transition, in which engineering stops being the binding constraint. The author offers the deliberately unglamorous label "parallelized, AI-assisted language modeling" for this, positioned as a more grounded cousin of RSI, or recursive self-improvement — the idea of a system improving itself in a compounding loop. The point is that readers do not need to accept takeoff scenarios to see that a large amount of infrastructure improvement is coming and that it will change how AI work is done.

The mechanism matters. Large parts of the training and inference stack are highly verifiable, meaning success or failure is measurable with a clean numeric signal. Training metrics include tokens per second per GPU, effectively training speed; inference metrics include tokens per prompt, FLOPs per token, and cost per answer. Those quantities have established sub-problems and architecture trade-offs, which makes them amenable to automated search and optimization. The author expects AI agents to help optimize this process end-to-end within a few years, pushing inference capability close to the maximum compute available on accelerators such as GPUs.

The near-term result is compounding efficiency. Recent years already produced large inference gains, with examples of saving roughly 10 to 30 percent on the cost of serving a model after it had been announced at a given price point. Stacked together, the author expects the effective cost of model intelligence to decline near-exponentially in coming years, possibly faster than recent trends, with the accessible efficiency wins captured within a few years. Beyond that, longer-horizon co-design of accelerators and models is expected to add further orders of magnitude on top of a flexible GPU platform, and GPU flexibility stays valuable while the architecture exploration space remains wide.

One concrete prediction: pretraining research, at least architecture and data selection for the current class of models, looks automatable in two to three years on this view. The author also expects a Jevons paradox dynamic for agentic models — cheaper intelligence driving more demand, not less — with the industry largely bottlenecked on figuring out how to orient and deliver agents. Meta's Muse agent is cited as an early indicator, and more Muse-like experiences aimed at different audiences and use cases are expected. The value there comes from understanding how agents work rather than from pushing the performance frontier, which is a subtly different engineering problem than training a better base model.

A second low-hanging fruit is reinforcement learning environments. The author notes that you do not have to look far to find new RL data companies that have crossed 100 million dollars or 1 billion dollars in revenue, more than expected. The quality of average outputs from this sector is described as remarkably low, with many researchers agreeing that much of what they buy is poor, even as leading labs see clear return on investment from the purchases. The problems with many RL environments are framed as clearly fixable, which is why the author treats them as an industrial-scale opportunity rather than a research mystery.

The broader scientific picture is where the author sets limits. In fields like biology and chemistry, the equivalent phase would look like models finding numerous cross-sub-field connections, because they are superhuman at crawling literature and linking sparse networks that are currently maintained by small communities of scientists who rarely interact. Whether that produces genuinely new eras of discovery, such as cures for most cancers, or simply accelerates the trajectory science was already on, is described as a fine line rather than a settled outcome.

The analysis also touches on what the shift means for how AI work feels. Working with coding agents is easier than before, and the form of the work is flexible and engaging, which the author reads as the start of an era where good ideas can be much more valuable than good execution in software. The argument is about where the bottleneck sits — engineering effort versus taste, framing and choosing the right problems — rather than a claim that models suddenly become qualitatively different.

Why it matters: for AI developers and infrastructure teams, this reading points to a few years in which automated optimization of serving cost, training throughput and RL environments becomes a primary competitive lever. That favors teams that can measure cost per answer, tokens per prompt and similar metrics cleanly, since those are exactly the signals agents can act on. Product teams should expect cheaper and more plentiful agent experiences rather than only step-change in raw model quality, though this remains the author's forecast, not a demonstrated result.