Part of: this research — view all 8 rows
As of November 2026, the baseline uplift in valuable research output at Anthropic is roughly 1.7× to 2.0×. Anthropic's internal capability index shows an acceleration trend of roughly 1.88×, and METR independently estimates progress on well-specified tasks at around twice the March 2024–October 2025 rate, attributing approximately 80% of this acceleration to AI Independent Assessment of AI R&D Acceleration ... Independent Assessment of AI R&D Acceleration .... Furthermore, Anthropic recently determined it met its Responsible Scaling Policy’s AI R&D threshold based on expected progress, even though no individual current model crosses it. Extrapolating this current acceleration to late 2029 requires modeling the tension between recursive self-improvement and structural R&D bottlenecks.
Crucially, task throughput does not equal research value. While Anthropic has reported a 45× increase in daily code volume and heavy delegation of implementation tasks, task substitution inflates these observed speedups well beyond actual value uplift Task Substitution and Uplift Task Substitution and Uplift. Compute limitations, experiment latency, and human research judgment act as severe rate-limiters. Under models of complementary inputs https://blog.redwoodresearch.org/p/if-mythos-actually-made-anthropic Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x, accelerating even 85–90% of a workflow by a factor of 25 yields only a roughly 5- to 7-fold end-to-end improvement. This structural friction constrains the expected growth ratio, rather than independently fixing the total uplift. I place the median near 8.8× as the product of the current baseline uplift and the expected growth ratio through 2029: a baseline of 1.7× to 2.0× implies roughly 4.4- to 5.2-fold further growth. This assumes broad automation of coding, evaluation, and experiment execution by 2029, but persistent constraints from serial training cycles and the need for human steering.
The lower tail of the distribution (p10 of ~2.2× and p25 of ~4.3×) accounts for the risk that current acceleration is overstated, heavily reliant on one-off task substitutions, or vulnerable to plateauing. Bounded task benchmarks do not easily generalize to discovering novel algorithms, and independent AI innovations remain underwhelming in practice TasteVal: Measuring the Experimental Research Taste of ... Can AI automate AI R&D yet?. Furthermore, safety incidents and governance pressures establish strong upper limits on realized pace. Anthropic has actively discussed modulating its development speed and keeping humans strictly in control of frontier models Anthropic says it met its AI research threshold Founders' letter, which could artificially cap autonomous R&D output even if the underlying capabilities exist.
Conversely, the distribution is heavily right-skewed (p75 ~17.5×, p90 ~45.0×) to capture the explosive potential of recursive self-improvement if coordination and judgment bottlenecks break. If scalable AI research taste matures—enabling reliable, fully automated AI researchers capable of designing, executing, and evaluating large programs—human-attention constraints would effectively vanish. Major labs, including OpenAI, are explicitly targeting full research automation within the 2027–2028 timeframe. If these milestones are met and compute scales accordingly, the counterfactual baseline of human-only research would appear remarkably slow, pushing the uplift multiple into the tens or hundreds by the end of 2029.
This estimate was explicitly aligned as the product of the current baseline uplift and the expected growth ratio through 2029, landing near 8.8x.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited