Recursive Self-Improvement: Jack Clark Moves AI Forecast to 2027

The timeline for one of the most debated milestones in artificial intelligence—recursive self-improvement—has just shifted. Jack Clark, co-founder of Anthropic and author of the Import AI newsletter, now believes that AI systems capable of developing their own successors could emerge as early as 2027. This represents a significant acceleration from his previous forecast of 2028, a move triggered by empirical data suggesting that models are acquiring a quality previously thought to be the sole domain of human intuition: research taste.

The emergence of research taste

At the center of this revised timeline is a new benchmark called TasteVal, developed by P-Zero Research. Unlike traditional benchmarks that measure raw performance or knowledge retrieval, TasteVal focuses on the ability of an AI to navigate the scientific process. P-Zero Research defines research taste through three distinct capabilities: identifying which research questions are worth pursuing, designing the experiments to answer them, and correctly interpreting the resulting data.

While TasteVal does not yet test the first capability—the intuitive leap of deciding which question to ask—it provides a rigorous test for the latter two. In these trials, models are given fixed research goals within AI development, such as optimizing training data selection, fine-tuning language models, or defending against adversarial prompts. The AI must then decide which experiment to run next, using the results to guide its subsequent steps. The data reveals an unusually steep improvement curve, with this specific capability roughly doubling every three months since December 2025.

Opus 5.5 vs. Human Experts

The practical implications of this growth are evident when comparing AI performance against human specialists. The TasteVal benchmark utilized a comparison group of 24 experts who had recently worked at elite organizations, including OpenAI, Google DeepMind, Nvidia, and Microsoft Research. When pitted against these professionals, Anthropic's Opus 5.5 led the field in experimental compute efficiency.

The cost disparity is perhaps the most striking finding. According to the P-Zero Research paper, a single run of Opus 5.5 cost approximately $282, whereas the average human run cost $9,123. This calculation accounts for GPU time priced at $2.50 per hour and human compensation at $119 per hour. By achieving superior results at roughly one-thirtieth of the cost, the model demonstrates that AI is not just mimicking research but is doing so with a level of efficiency that could drastically shorten the development cycle of future models.

The mechanics of recursive self-improvement

Recursive self-improvement (RSI) describes a feedback loop where an AI system conducts the research necessary to train a stronger successor, which in turn repeats the process. For years, the primary skepticism regarding RSI centered on the lack of creativity and judgment in LLMs. Critics argued that while AI could code, it could not innovate. However, the results from TasteVal suggest that the gap in research judgment is closing faster than anticipated.

To ensure that coding proficiency did not skew the results, P-Zero Research used a fixed coding agent to run all experiments on a single Nvidia H100 GPU. This isolation proves that the gains are coming from the model's ability to strategize and interpret, not just its ability to write Python scripts. This is complemented by Anthropic's own internal data, which shows a massive surge in engineering productivity; code changes merged daily rose from 2,500 in May to over 30,000 in September 2026, suggesting that coding is no longer the primary bottleneck in AI development.

Bridging the gap between proposal and decision

Despite the optimism, some researchers urge caution. A study from Fudan indicates that while AI agents can generate up to 55 percent of methodology proposals for their own development, humans still make over 85 percent of the final decisions. This suggests a critical distinction between proposing a path and deciding to commit resources to it. The real work of scientific discovery often happens in that final decision-making gap, which current benchmarks may undercount.

The debate is shifting away from raw model performance and toward the question of how much independent research work language models can actually take on.

Furthermore, TasteVal's tasks have results that are relatively easy to verify, and the human baseline excluded the absolute top tier of global researchers. While the trend is clear, the transition from a controlled benchmark to a fully autonomous research loop remains a complex leap.

The societal cost of acceleration

As the technical timeline accelerates, the conversation is shifting toward the economic fallout. The disparity in expectations is stark: a survey by Anthropic found that its own employees expect to delegate 75 percent of their duties to algorithms by 2030, while the general American public expects less than a third of tasks to be automated. This gap is described by Andrej Karpathy as a funnel, where only a few thousand specialists see the true speed of AI growth, while the rest of the world sees only a preview.

In response to the potential for widespread technological unemployment, Anthropic has pledged to allocate 5% of its shares to a public benefit fund. This fund is intended to provide assistance to those displaced by automation, acknowledging that the transition to an AI-driven economy will not be seamless. This move signals that the companies building these systems are no longer treating automation as a distant theoretical risk, but as an imminent economic reality.

Global implications for business and regulation

For entrepreneurs and executives in the USA and UK, the shift of the RSI timeline to 2027 necessitates a strategic pivot. In these markets, where the regulatory approach is generally more flexible than in the EU, the primary risk is not legal compliance but operational obsolescence. If AI can begin to iterate on its own architecture, the rate of software depreciation will accelerate. Businesses relying on static AI implementations may find their tools outdated within months rather than years.

In the UK, where the government has positioned itself as a hub for AI safety and innovation, the focus will likely remain on the stability of these recursive loops to prevent unpredictable model drift. In the USA, the competitive drive between labs like Anthropic and OpenAI will likely push the boundaries of RSI even further, potentially creating a winner-take-all scenario for the first entity to achieve a stable self-improvement loop. For the global market, the key takeaway is the decoupling of productivity from human labor costs; as seen in the TasteVal results, the ability to conduct high-level research at 3% of the human cost will fundamentally rewrite the unit economics of R&D across every tech-driven industry.

FAQ

Sources: The-decoder, Aidailypost, Therundown ·