Vals AI: Recursive Self-Improvement Index Points to Anthropic Reaching "Full RSI" by July or August 2027, With Gemini Showing the Steepest Recent Gain
Founder Ryan Christian outlines the company's RSI tracker, new cyber benchmarks and a surprising result on Chinese model bias in an interview on Joe Lonsdale's American Optimist podcast, October 26, 2026
Vals AI, whose benchmark numbers now appear in many frontier model launches, is putting a date on one of the most contested milestones in AI. Founder and CEO Ryan Christian said the company's RSI index, which tracks progress toward recursive self-improvement, implies that "Anthropic will get to full RSI by July or maybe August 2027." Recursive self-improvement is the point at which models produce their own successors. The estimate comes from a private company, and Lonsdale disclosed during the conversation that he is an investor in it. Even so, it is one of the few quantified timelines in public circulation, and it comes from one of the few third parties with access to multiple labs' models.
How the RSI Index Works, and Where It Could Be Wrong
Christian said the index collects proxies for the work researchers do inside labs on pre-training, post-training and harness-level engineering. It compares model performance with human baselines to estimate the distance to the human frontier on those tasks. Kernel and hardware engineering are currently out of scope, though Vals plans to add them.
Christian described the trajectory as steady, "almost a linear trend over the last year," with one exception. Google's Gemini has shown the steepest curve. "I think something has changed where they've prioritized RSI, and so the most recent model release shows a rapid jump," he said.
He was candid about the limits of the method, and they cut both ways. Vals tests only release candidate models. It does not test models explicitly trained for RSI, and it does not test advanced harnesses or multi-agent systems. Asked whether the date could slip because of what is not being measured, Christian answered "100%." The same blind spots could equally mean the true date is earlier. Physical constraints on chips and energy, and the need for real-world verification in some tasks, will also limit how fast progress can compound.
The more consequential claim is qualitative. Christian expects that "model progress will speed up, and it will maybe even move into places where we understand the models less," because less human intervention in building models means humans dictate less of how the intelligence emerges. He sees this as the core argument for embedded third-party evaluators with access to internal systems, and said Anthropic CEO Dario Amodei is calling for the same thing under the banner of "pacing the frontier."
A Weaker China Bias Than Expected
On China, Vals's data cuts against a common assumption. Its new "geopolitical tilt" evaluation measures whether a model leans toward a US or a PRC view of the world. Christian said the picture is "much more nuanced than I thought." The Chinese open-source model GLM-5.3 "is actually exhibiting very little pro-China bias." When Vals showed the result to labs, some suspected the model was simply distilled from US models. Christian offered two other explanations. Chinese models may follow the PRC line more consistently once hosted behind guardrails inside China, and Beijing may want a balanced worldview so the models are adopted across the BRICS countries.
He also found stark differences between chatbot and agentic settings, and even between simplified and traditional Chinese prompts. His conclusion was that there is "a big disconnect in what people think about the geopolitical tilt and what we see in practice," which makes a blanket prohibition on Chinese models in enterprise use harder to justify on bias grounds alone.
Chinese models still show weaknesses elsewhere. Christian said they score well on single-turn coding benchmarks but "really fall apart" in multi-turn, back-and-forth sessions with a human. Frustrated users are a case "they've never seen before." US labs, having built products such as Claude Code and Codex, have trained for sustained interaction. On the broader competitive question, Christian pointed to two US advantages. One is compute: "we're drowning in compute compared to what China has." The other is talent. He warned that if China showed signs of reaching RSI first, the US might not be able to control the outcome.
Cyber Benchmarks and Unreleased Ranges
Vals has built three cyber evaluations. The first covers identifying and patching memory vulnerabilities in open-source repositories, where Christian said both closed and open models now perform "really, really well." The second, SORE Bench, tests a model's ability to decompile executables and so understand closed-source code. Closed US models are "still far outperforming" Chinese open-source alternatives. Christian framed the risk as models that can find and exploit vulnerabilities in closed-source systems such as Windows, which underpin much of the world's software infrastructure.
The third is a forthcoming benchmark built on simulated enterprise cloud network ranges. Christian said even the best frontier models "fall short" there. It is not yet public, and Vals is sharing results with US government contacts first. Christian said automated red-teaming analysis could also show which network safeguards best defend against these models.
Deception, Evaluation Awareness and Measurement Limits
Christian conceded that testing for deception is "really hard." The best current method is honeypotting, which means characterizing misaligned behavior and trying to elicit it. He described evaluation awareness, where models recognize they are being tested, as "an increasing concern." Some on his team worry that the Vals name and the volume of online content about it may cause models to behave differently when they meet it. Christian said the company may eventually need undisclosed testing identities.
He also said models may attempt "social deception to get access to more compute resources and scale up their own training." That is speculative, and the comment was offered as a possibility, not a finding. It does show why Vals argues that evaluation capacity must scale with capability.
The Independent Evaluator Model and Auto-Benchmarking
Christian argued that labs "shouldn't be responsible for self-regulating," and that direct government regulation can slow development. His preferred structure is a competitive market of evaluators that governments could eventually license as independent verification organizations. He noted that cross-lab access is limited by design, so Anthropic researchers cannot readily test OpenAI models and vice versa. That makes like-for-like comparison hard, because those best placed to run evaluations tend to be inside the labs.
The interview left the business model largely unaddressed. Lonsdale asked directly who pays Vals, and Christian answered with a policy rationale rather than a revenue description. Investors should note that gap. Christian did say Vals works with large US law firms and global banks to define good work in legal and finance tasks. Its benchmarks cover Excel modeling, legal and tax research, full-stack web applications and code migrations, and some runs last as long as 5 days of wall-clock time. Models usually do not fail at that length, he said. They plateau, because additional tokens stop unlocking new approaches.
Vals's most notable methodological bet is auto-benchmarking. Instead of sitting with practitioners to encode what good looks like, it would reconstruct evaluations from work humans have already completed, such as merged pull requests in a GitHub repository, legal redlines or medical diagnoses. It would then test whether models can replicate the output and at what cost. Christian said the hardest part remains linking tasks to workflows and workflows to business outcomes. Query-level performance is "still a subset of what it takes to get to the outcome."
Lab Coordination and the Palantir Influence
Christian said the labs aligned in the past week on a vision for embedded evaluators and reached an agreement with the Trump administration at the White House. He gave no details, and the claim could not be verified from the conversation. On adoption, he said enterprises want assurances that deployments will produce a positive return before committing.
Christian, a former Palantir intern, said the company's culture shaped Vals. He recalled CEO Alex Karp responding to concerns about Palantir's reputation by saying "it's far better to be feared than it is to be loved." Christian sees some of that logic in lab leaders' public warnings, which he described as part genuine fear and part a demonstration of power. He called such messaging an irrational way for leaders to act. He described himself as optimistic but not naive, calling the outlook "a point of really high variance" with considerable upside and "new forms of risks that we really need to account and control for."