❓ What is recursive self-improvement (RSI)?
At the heart of the warning is a phrase from Anthropic co-founder Jack Clark: "Claude 10 building Claude 11". It describes AI helping design a better AI, which in turn helps design the next one—faster with every cycle. That is recursive self-improvement, or RSI. It is neither magic nor fiction: it is an engineering process that feeds back into itself.
🧭 New here? Two terms, without jargon
- RSI (recursive self-improvement): recursive self-improvement. “Recursive” means something applied to itself. Here, AI is used to improve AI, and the result becomes the tool for the next round.
- Frontier model: the most advanced AI models available, built by leading labs (Anthropic, OpenAI, Google DeepMind). They are at the center of this discussion.
🔁 The one-line definition
RSI is when AI stops being only the product of research and also becomes part of the engine that drives research. If each model helps build the next, progress no longer depends only on how quickly people think.
🧭 The shifting bottleneck
Every process has a bottleneck —the slowest point, which limits the pace of the whole system. In AI research today, that bottleneck is human: how quickly people can form ideas, write code, run experiments, and interpret results. RSI shifts where that bottleneck lies.
Research does not disappear—the slowest link changes. When people are no longer the bottleneck, “how many chips and how much energy you have” begins to determine speed (a Track 3 topic).
🔎 Why this matters
When people are the bottleneck, hiring more researchers speeds things up. When the bottleneck becomes compute and autonomy, the race changes: those with more data centers and a greater willingness to give systems freedom move faster. That is why infrastructure becomes central to the warning.
📈 Soft vs. hard self-improvement
Demis Hassabis (Google DeepMind) makes a crucial distinction for reading the warning without exaggeration: we are in a phase of "soft" (gradual), not "hard" (hard). Soft means AI making engineers much more productive. Hard would be a model that “disappears into a data center” and returns superintelligent on its own. Today we are in the soft phase.
🧭 New here? One term
Superintelligence (ASI): an AI that would greatly surpass the best humans at nearly everything. This is hypothetical, not something that exists today. “Hard self-improvement” is an imagined route to get there quickly—and it is exactly the speculative part of the warning.
The “hard” step is dashed on purpose: it is an imagined scenario, not an observed fact. Confusing the two is the biggest source of miscalibrated panic.
Soft (observable today)
- •AI speeds up the work of real engineers.
- •People still direct every important step.
- •The gain is productivity, not full autonomy.
Hard (speculative scenario)
- •A model would redesign itself without a human in the loop.
- •Rapid leaps that are hard to predict or stop.
- •No one has demonstrated this—it is the part to treat skeptically.
🎲 Clark’s estimate (~60% by 2028)
Jack Clark gave a figure: about a 60% chance of some significant form of recursive self-improvement by the end of 2028 (according to the video—unverified). Assigning a probability to an idea is an honest way to say “I think it is likely, but I am not sure.” The mistake is reading it as a set date. It is the calibrated estimate of one one person—not a field-wide consensus.
🎯 What “60%” means (and what it does not)
A subjective probability expresses confidence, not a countdown. “60% by 2028” means that, in his view, it is more likely than not—but there is plenty of room to be wrong. It does not mean “it will happen in 2028” or “it is settled.”
✅ Solid
- ✓RSI is an explicitly named central topic for Clark, Hassabis, and Hinton.
- ✓Using probability to express uncertainty is a legitimate practice.
- ✓The direction (AI accelerating its own research) is real and observable.
⚠️ To verify / hype
- !The “60% by 2028” figure is one person’s estimate, not a consensus.
- !Treating the date as certain is rhetoric, not evidence.
- !“Confirmed / it is real” is the channel’s framing—check the primary source.
Copy and run: ask a chatbot to separate fact from estimate
Goal: see a model distinguish consensus from speculation about the 2028 timeline. Paste the prompt below into any chatbot (Claude, ChatGPT, Gemini).
Explain recursive self-improvement as if I were 15, then give me 3 reasons for and 3 against the 2028 timeline. Label each item in brackets as [consensus] or [speculation].
How to check: A good answer labels most “2028” claims as [speculation] and keeps only the general direction (AI already speeds up research) as [consensus]. If it states the date as certain, it failed the honesty test—exactly what this course trains you to notice.
🧬 RSI ≠ singularity ≠ consciousness
Three terms are constantly confused—and that leads to both misplaced panic and misplaced skepticism. RSI is not “the singularity,” and neither requires consciousness. RSI is a bottleneck shift in an engineering process. That is all.
🧭 New here? Three terms to keep straight
- Singularity: a hypothetical point where AI progress becomes so fast that it is unpredictable. It is a futuristic idea, not a measured fact.
- AGI: artificial general intelligence—AI that can perform most human cognitive tasks well. ASI (superintelligence) is the next step up: it would greatly surpass humans.
- Consciousness: having subjective experience. RSI does not depend on this—a system can optimize code without “feeling” anything.
✓ What RSI IS
- ✓An engineering process that feeds back into itself.
- ✓A shift in who is the slowest factor.
- ✓Something measurable through proxies (speed, task horizon).
✗ What RSI is NOT
- ✗A conscious robot “waking up.”
- ✗The singularity guaranteed on a specific date.
- ✗Something already proven to have moved beyond “soft.”
🚨 Why this is a warning, not a reason to panic
The course is called “Alert,” not “Apocalypse.” The early machinery is visible, timelines are being put on the table, and labs increasingly rely on their own models. This justifies informed attention —not panic, and not the certainty that nothing will happen.
💡 The course’s stance
For every strong claim, ask two questions: “Is this a well-supported category or a figure from one article?” and “Who is speaking, and what is their interest?” Keeping both questions in mind helps you explore the topic without needless fear or ignoring what matters.
The course map from here
🔁 Fundamentals
The loop and why it starts with code.
📊 The evidence
The METR curve and MirrorCode—the real figures.
🛡️ Safety and 2028
When measurement breaks, the race, and the scenarios.
Optional self-check: which is the best description of RSI according to this module?
🎯 Module summary
Next module:
1.2 — Why it starts with code: the loop that closes in seconds and the agents that close it.