Recursive self-improvement is likely lossy rather than an exponential explosion

Most of the hype around recursive self-improvement (RSI) ignores the fact that automatable research is too narrow to trigger a massive net acceleration in progress, especially when you factor in the exponential costs of scaling laws. While frontier labs like OpenAI and Anthropic are now deploying thousands of concurrent agents to optimize their internal processes, this doesn't necessarily mean we are hitting a singularity. The reality is that we are seeing diminishing returns as more AI agents run in parallel, and the actual bottlenecks for building strong LLMs are often resource-related or political—things AI can't simply "solve" through code.

Why the "singularity soon" narrative feels exaggerated

The culture in the San Francisco AI scene tends to amplify every concern. We saw this in 2023 and 2024 with loud AI safety debates and claims that open-source AI wasn't viable; those forecasted risks didn't actually arrive on those timelines. Now, as agents found stronger product-market fit at the start of 2026, the anxiety has shifted toward extinction risks.

Richard Ngo pointed out a critical nuance here: the AI safety community is orienting toward an intelligence explosion within a few years. He suggests that while they might be "directionally correct" compared to outsiders, they are likely "factually wrong." We probably won't see superintelligence within the next 8 years, but the pace of progress will be fast enough that it feels like the short-timeline advocates were right.

The constraints on true RSI

For me, the argument against a sudden intelligence explosion comes down to three specific frictions:

• Scaling Law Costs: The exponential cost of scaling makes it difficult for automatable research to provide a net acceleration.
• Parallel Diminishing Returns: Adding more agents in parallel does not result in a linear increase in breakthroughs.
• Physical and Political Bottlenecks: Building the next generation of models requires hardware and political navigation that software agents cannot bypass.

Unless we see a foundational, imagination-based breakthrough that fundamentally changes how AI evolves, I view RSI more as a tool to sustain progress against rising costs rather than an unpredictable or unstable explosion.

The impact of mass inference capacity

Looking at the current landscape, the real shift isn't necessarily "superintelligence" but the sheer scale of inference. Labs are throwing thousands of agents at measurable problems while continuing to scale their compute capacity. This creates a massive short-term acceleration in how problems are solved, which is often mistaken for a recursive loop leading to a singularity. The distinction is important: scaled agents working productively is a technical milestone, not a religious omen of extinction.