Why I still haven’t bought into true RSI

What it claims

The author argues that true recursive self-improvement (RSI) is not the right baseline for the next few years. Their alternative, "lossy self-improvement," rests on three points. First, automatable research is too narrow to produce a large net acceleration against the exponential costs implied by scaling laws. Second, running more AI agents in parallel has real diminishing returns. Third, resource bottlenecks and politics limit how strong LLMs get, and AI can do little to speed those up.

On this view, near-term RSI mostly makes models cheaper and more efficient rather than raising peak intelligence. Inference serving has clear, measurable metrics that agent swarms can optimise. Peak intelligence still needs exponential compute for linear gains. The author also reads the current spike in extinction-risk discussion, especially inside OpenAI and Anthropic, as amplification driven by lab culture, and calls it "very misplaced" absent new evidence. They allow that labs may have seen breakthroughs that are not yet public.

Evidence

This is an opinion essay. Its support is citations and quotes, not experiments.

• Anthropic's own assessment. The Claude Fable 5.1 & Mythos 5.1 System Card says internal model use has been "a key factor in maintaining the current rate of progress, but we do not yet see clear signs of dramatic acceleration beyond that rate." The author reads published lab measurements as showing that automation gains so far are concentrated in software engineering, log monitoring and running planned experiments.
• John Schulman on post-training, from a Dwarkesh podcast. Post-training needs many people because "someone has to think about how the model should behave" in each area. Also: "It's really easy to screw up post-training in some way that doesn't show up in benchmarks."
• Timeline estimates from the same podcast, relative to the interview date. The page does not give that date, and the answers were summarised by GPT-6-Astra:

• Richard Ngo's forecast, which the author endorses. Superintelligence will not arrive within 8 years, but progress will feel fast enough that short-timeline advocates seem right.

Practical implications stated

• Expect agent swarms to work well on clear, verifiable problems: experiment design and testing, inference efficiency, multi-agent systems. Do not expect them to do as well at hypothesis generation, intuition building or scientific communication.
• Do not mistake inference-time scaling, which the author calls "fairly predictable," for evidence of RSI.
• Expect the cost of LLMs at a fixed capability level to fall faster. The author expects Jevons paradox to sustain lab businesses despite price competition.
• Building RL environments hard enough to challenge leading models may become exponentially harder, and that could cap RL-driven gains.

Limits

• The argument is qualitative, with no new data. The author explicitly holds "high levels of uncertainty" and names what would change their mind: foundational, "imagination-based" breakthroughs inside labs that are not yet public.
• The timeline table is an LLM-generated summary of a podcast, not a direct transcript.
• Doubts about labs keeping a constant share of compute for internal R&D, given IPO scrutiny, are stated as speculation.
• The thresholds discussed (remote worker, researcher) are human-shaped. The author notes that AI capability is jagged, so these thresholds will be crossed gradually with a persistent long tail rather than at a clear point.