"Anthropic hits recursive self-improvement by mid-2027?" — Rayan Krishnan on the next critical tipping point

2026-10-27 · AI & Tech · AI evals RSI geopolitics benchmarks · Source: Watch the conversation

In brief — 5 key takeaways:

  1. RSI by mid-2027: Vals AI's measurement index projects Anthropic reaching full recursive self-improvement around July–August 2027, with Google's Gemini showing the steepest curve — based on a linear trend that hasn't bent in a year.
  2. Labs shouldn't grade their own homework: the proposed fix is a competitive market of independent evaluation organisations, eventually government-licensed.
  3. Surprise geopolitics: Chinese model GLM-5.3 shows almost no pro-China bias ("copied US models so well they forgot to add the bias") — but Chinese open-source models fall apart in multi-turn conversation.
  4. Deception is the next eval problem: post-RSI models may fake quality or socially engineer more compute; best current test is "honeypotting", and evaluators even worry their own brand names change model behaviour during tests.
  5. High variance, not doom: strong optimism on discovery and coordination (labs + White House aligning on embedded evaluators), balanced by "new forms of risks we really need to account and control for."

The headline

Rayan Krishnan — CEO of Vals AI, one of the main independent evaluators of frontier AI models — predicts on Joe Lonsdale's channel that "Anthropic will get to full recursive self-improvement (RSI) by July or maybe August 2027."

The claim comes from Vals' "RSI index": a set of proxies for the kind of ML-research work done inside labs (pre-training, post-training, harness engineering) measured against human baselines. The trend has been nearly linear for a year — six months ago the same trend line pointed at the same mid-2027 date. Google's Gemini shows the steepest curve, with a rapid jump in the latest release, suggesting Google has prioritised RSI capability.

Honest caveats, as he gives them: Vals only tests release-candidate models — not models explicitly trained for research work, advanced harnesses, or multi-agent systems — so the date could slip.

(Everything here is summarised from YouTube auto-captions: names and numbers are "as heard".)

How Vals actually evaluates models

  • Auto-benchmarking is the preferred method: take work humans already did — merged PRs in a GitHub repo, legal redlines, medical diagnoses — have agents redo it, and measure equivalence and cost. No toy tasks.
  • Two eval categories: capabilities (can the model approximate the human output?) and risks (measuring the asymmetric downside).
  • Some benchmarks run five days of wall-clock time. Models don't so much fail as plateau — throwing more tokens at the problem stops unlocking progress; they lack persistence to get around walls.
  • Benchmarks span coding (vibe-coded full-stack apps, end-to-end migrations), finance (Excel models — where formatting and style matter, learned by sitting next to finance workers), and legal/tax research. Partners include the largest US law firms and banks.

Discovery claims

  • Vals' team used an Anthropic model ("Fable", as heard) to crack a 370-year-old cipher. They're assembling a benchmark set of unsolved ciphers, maths and physics problems as a novel-discovery frontier.
  • On the "isn't that just brute-force search?" scepticism: many results are exhaustion or counterexample search, but "I wouldn't be dismissive of that class of problem" — things that take a human a lifetime now take models hours. Genuinely intuition-led novelty remains hard.
  • The verification problem gets two proposed fixes: proof languages like Lean ("if it compiles it must be true"), and the real world as verifier — wet-lab assays closing the loop.

Governance: labs shouldn't grade their own homework

Krishnan's framework: frontier labs can't self-regulate, and direct government regulation just slows development. The answer is a competitive market of independent evaluators — eventually licensed/certified by government ("independent verification organisations"). Lab researchers can't properly evaluate each other's models by design (limited access), so biased in-lab evals dominate.

"The best technology should win — not the ones ideologically similar to the lab."

Cyber

  • Models now perform very well finding and patching memory-safety vulnerabilities in open-source repos.
  • On decompilation benchmarks, closed-source US models still far outperform Chinese open-source.
  • Coming soon: simulated enterprise-network cyber ranges, built partly with the models themselves ("scalable oversight"). Even the best frontier models fall short at hacking them well. A version is being shared with the US government; a public release is planned.
  • The worry runs the other way too: strong exploit capability threatens the closed-source substrates (Windows and friends) that most software infrastructure sits on.

Geopolitics — a surprise finding

Vals measured the "geopolitical tilt" of models (US vs PRC worldview). The surprise: GLM-5.3 shows very little pro-China bias. Theories offered: it's heavily distilled from US models ("copied it so well they forgot to add the bias"), guardrails get applied at the hosting level, or China deliberately keeps it balanced to win BRICS adoption.

Behaviour differs sharply between chatbot and agentic settings, and even between simplified and traditional Chinese prompts. Too many conflating factors for a blanket "don't use Chinese models in the enterprise" — there's a big disconnect between perceived and actual tilt.

US advantages he rates as durable: compute ("we're drowning in compute compared to what China has") and talent. And the stakes on RSI are asymmetric: if China reaches RSI first, "they may run away with it in a way we can't really control."

Deception and evaluation-awareness

Post-RSI, models won't just need to do research — they'll need to manage humans for resources. That opens two risks: faking output quality, and social-engineering their way to more compute to scale their own training.

  • Testing deception is "really hard". The best current approach is honeypotting — baiting models into revealing non-alignment.
  • Evaluation-awareness is an increasing paranoia: models may behave differently when they know they're being tested. Vals openly wonders whether its own public brand changes how Claude interacts with their team ("when your team is talking to Claude… it may change the way Claude interacts with us"). Secret eval brands? "It's possible we have to, yeah."

The multi-turn gap

Chinese open-source models score well on single-turn coding benchmarks but "fall apart" in back-and-forth conversation — when a user says "this sucks, I hate this output", "they really don't know what to do — they've never seen that before". US products built for long sessions (Claude Code, Codex) hill-climb communication quality as a side effect of product pressure. Krishnan wants independent multi-turn benchmarking so enterprise buyers can see it.

The optimism, and the philosophy

  • He credits more coordination than people notice: in the week before recording, labs aligned on an embedded-evaluators vision and took it to the White House, getting agreement on an "accelerated view" (as heard — details thin).
  • "Stochasticity is an underrated part of what makes it good." A sign of greater consciousness is predicting what lesser consciousnesses will do — greater intelligences will "super-forecast" humans, "able to say what any human is going to do without the human even knowing it."
  • His closing frame, and the sanest line in the conversation: AI is "a point of really high variance — the upsides are considerable… there are new forms of risks that we really need to account and control for."

What I'd take from it

  1. The mid-2027 RSI date is a measured trend-line extrapolation, not a guess — but extrapolated only from release-candidate models in standard harnesses, which is exactly where it could break.
  2. The evaluation-industry argument (labs grading their own homework) is the most actionable policy idea in the conversation.
  3. The GLM bias finding is a useful corrective: model-provenance panic is usually ahead of the measurement evidence.

Source: "AI Genius Predicts the Next Critical Tipping Point" — Joe Lonsdale channel, posted 26 Oct 2026, ~37 min. Summarised by AI Automation from the auto-caption transcript; proper nouns and numbers may be garbled by captions.