While Silicon Valley's giants slam the brakes — Anthropic and OpenAI both signaled slowing frontier-AI pace, and tech and semiconductor stocks sold off — a long read published by TMTPost on September 20, "The AI giants hit the brakes, but true RSI is still far away," offers the opposite observation: recursive self-improvement, the concept making those giants nervous, is still several hard steps from real implementation, and the most hidden step lives inside "how AI digests its own experience."
The piece dismantles the RSI narrative in three layers. Firs
[1]t, the market layer: the key assumption underpinning AI-infrastructure growth expectations just gained a new variable — when people who have been pressing the AI accelerator start talking about braking, the logic of "models keep getting stronger, capex keeps rising" is no longer self-evident; Nvidia fell over 3%, AMD, Intel and Micron dropped around 5% intraday, and the Philadelphia Semiconductor Index slid nearly 6%. Second, the debate layer: the popular claim that "the ceiling of harness evolution may exceed model evolution" meets the author's reservation — a harness, however strong, can only tell a model what it should do; it cannot make the judgment. How a task gets planned, how sub-agents get split — that still rests on the model itself. And the comparison is unfair: today's model evolution often builds on open-source models only, while harness evolution stands directly on the strongest closed models. The author argues for model-harness co-evolution, not either-or.
The third layer is where the piece earns its keep: three hidden traps of self-improvement. First, self-distillation drops reflection: solving a problem by trying A (wrong), then B (also wrong), finally C — if you feed only "do C" back into the model, the answer is learned but the whole negation process of why A and B failed is lost, and much good judgment comes precisely from that process. Second, the experience library "stores but fails to retrieve": as memory grows, at the moment of need the right experience may not come back. Third, path dependence: an agent sampling continuously gravitates to high-probability paths; the longer it walks one road, the more the context reinforces it and the harder it is to jump off. A system may grow stronger by learning — and it may also become increasingly constrained by its own past.
Read against the week's other texts, a full spectrum emerges: Dario says RSI has accelerated and needs braking; the Neuron commentary says, from neuroscience, that continual learning and self-monitoring are gaps; this TMTPost piece says, from engineering practice, that "being able to learn" is only the first step — AI must learn to judge which experiences deserve trust, which deserve forgetting, and when to deviate from its most familiar paths. The consensus hides inside the disagreement: whether braking or accelerating, self-improvement is not yet "turning experience into capability" — and that is RSI's real threshold.
The stance should be stated: this piece relays and interprets the TMTPost article's views, not an endorsement of any judgment within it; the author speaks from a researcher/practitioner perspective in interview form, and claims like "co-evolution is worth exploring" are personal judgments rather than lab conclusions. What deserves remembering is the paradox about learning itself: a system may of course grow stronger by learning, but it may also, by learning, become increasingly trapped by its own past. That paradox is closer to the essence of AI self-improvement than either "brake" or "accelerate".
[1]