Researcher Uplift from Code Output

Sources#

• Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x

Summary#

METR's Thomas Kwa (2026-07-08, practitioner-opinion) asks what Anthropic's reported 8× code merged per engineer per day (AI Accelerating AI Development) implies for serial researcher uplift — the factor by which a researcher's total effective research output has risen. Feeding the single number M = 8 through standard production functions, Kwa estimates the answer is robustly over 2× from coding agents alone (central estimate ~2.5×), even assuming zero AI speedup on non-code work. The result is notable less for the number than for its robustness: the estimate barely moves whether coding and non-coding work are complements or substitutes, thanks to a mathematical coincidence — 8 ≈ e².

The note is explicitly framed as opinion: "others at METR disagree," and "the math was checked by Claude but not a second human." Treat the 2.3–2.9× band as one careful modeler's estimate derived from a single vendor-reported lines-of-code figure, not an established measurement.

The production-function translation#

A researcher's output depends on two inputs — code produced (q_n) and non-code produced (q_o: thinking, writing, experiments, communication). A production function maps input quantities to total research value; researcher uplift U is the post-AI / pre-AI output ratio. The observable is M = 8, the measured code-output multiplier — not the per-hour coding speedup g_n, which is unobserved because researchers reallocate time when coding gets cheaper. β is the pre-AI time share on coding, fixed at 0.5 throughout ("roughly the median of what people I ask find reasonable"; Kwa flags substantial uncertainty but holds it constant for simplicity).

Cobb-Douglas: U = M^β#

The simplest model. Assuming pessimistically no non-code uplift (q_o constant), uplift collapses to U = M^β = √8 ≈ 2.83.

CES and the 8 ≈ e² robustness#

CES generalizes Cobb-Douglas by adding the elasticity of substitution σ between coding and non-coding output (σ→0 = strict complements, left-and-right shoes; σ→∞ = perfect substitutes, butter-and-margarine; σ=1 recovers Cobb-Douglas). Observing M = 8 pins U down almost independently of σ:

σ	implied coding speedup g_n	post-AI coding share	U
0.5	23.3	17%	2.75
1.0	8.0	50%	2.83
2.0	4.8	83%	2.91
3.0	4.2	95%	3.07

The intuition: 8× output constrains the two free parameters (coding speedup and marginal value of code). If both were high, output would exceed 8×; if both low, less than 8×. Only the intermediate trades — high speedup + low marginal value (σ<1, researchers flee coding for other bottlenecks) or low speedup + high marginal value (σ>1, researchers pile into coding) — are consistent with exactly 8×, and both land U near 2.83.

The coincidence: the first-order σ-sensitivity of ln U is proportional to [−ln M + (ln M)²/2], which vanishes at M = e² ≈ 7.39. At M = 8 it is ≈ 0.08 (near zero), so U stays within ±3% of 2.83 for σ ∈ [0.5, 2]. Kwa's punchline: "Because 8 ≈ e², the estimate is robust to σ."

M = 8 also lower-bounds σ ≥ 0.5: at σ = 0 (Leontief / perfect complements) output can rise at most 1/(1−β) = 2×, so 8× is flatly impossible; small σ would require implausible coding speedups (g_n ≈ 114 at σ = 0.3) and an implausibly low post-AI coding time-share (3.5%). This rules out strong complementarity between coding and non-coding research — i.e. non-code work is not a strict bottleneck that caps how far code speedup can carry total output.

Code heterogeneity: the lower band#

If AI disproportionately speeds up low-stakes code (which also tends to be lower value per line), the 8× is padded with cheap lines and value-uplift is lower. Modeling code as an inner CES between low- and high-stakes work (high-stakes getting 1/3 the log-uplift) gives U ∈ [2.33, 2.66] across the plausible low-stakes-share range α ∈ [0.3, 0.9] — lower than the homogeneous estimate, but still over 2×. This model is less σ-robust than the homogeneous one (±11% at α = 0.9 vs ±3%).

Prefer code output over code uplift#

A methodological takeaway: to estimate researcher uplift, the measured code output multiplier (M) beats the per-hour coding speedup (g_n). Output already prices in the marginal value of code through the time-reallocation researchers actually do; the raw per-hour speedup does not, and its uplift estimate is not robust to σ. (Robustness peaks near M ≈ e² but is always better for output than uplift.) The residual weakness: output is still corrupted by "irrational" time-allocation shifts (below), so the metric it really points to is quality-adjusted code output.

The three ways it could fall below 2×#

Kwa lists five caveats, three plausible:

• Verbosity — AI writes more lines for the same functionality. METR's own early-2025 uplift RCT found developers wrote 1.22–2.57× more LoC (95% CI) on AI-allowed issues. If Anthropic's verbosity factor were ~1.83×, true code output is 8/1.83 ≈ 4.4× and the heterogeneous model gives U ∈ [1.84, 2.08] — right at the 2× threshold. Kwa argues verbosity is bounded: LoC/person jumped 2.5×→5.8×→8.0× across recent quarters, and most of that rise came after code was already majority-AI-written, so it cannot be mostly verbosity.
• Barely-useful ("Cadillac") code — new low-stakes code that would never have been written by hand and adds little research value (a project-DAG visualizer, a web UI for agent runs, a redundant stats-methodology redo). CES already discounts this via diminishing returns; extra discounting is warranted only if irrationality or sheer volume outruns what CES predicts.
• Irrational time allocation — researchers coding more because it is fun, not because it is valuable. Cited support: METR's prior finding that developers felt ~20% sped up while actually being ~20% slower. Unlike verbosity (roughly a constant factor), this bias likely worsens as uplift grows.

Reconciling with Anthropic's "well short of 2×"#

Anthropic's Mythos Preview system card said overall R&D acceleration is "well short of a sustained, AI-attributable doubling… concentrated in engineering execution rather than research judgment." This is not a contradiction of >2× researcher uplift — the two estimate different quantities:

• Kwa estimates serial researcher uplift (labor only).
• Anthropic estimates overall R&D speedup, which also depends on compute — and compute growth does not count toward researcher uplift.

Ryan Greenblatt's decomposition: R&D speedup ≈ (serial labor acceleration)^0.55 × (compute)^0.45. So a 2.83× researcher uplift implies only ~1.77× R&D speedup, and Anthropic's 2× R&D threshold would require ~3.52× researcher uplift — "which basically requires coding to be >90% automated" and "could happen in the next year or so." Kwa finds even the "reassuring" 1.77× alarming: with compute tripling yearly (~1.6× research-input growth/yr on its own via the 0.45 exponent) and labor uplift rising as models improve, both R&D inputs grow exponentially — which under semi-endogenous growth models sustains exponential research-output gains. (See Intelligence Explosion Dynamics, Effective Compute Scaling.)

Connections#

• AI Accelerating AI Development — supplies the input this note models: Anthropic's ~8× code/engineer/day figure (which that page flags as "almost certainly an overstatement of the true productivity gain"); this page is the economic translation of that raw multiplier into a research-value estimate
• AI R&D Autonomy Evaluation (AECI) — Anthropic announced a shift to "direct measurement of AI R&D acceleration and researcher uplift" but had not operationalized it; this is a third-party (METR) attempt at exactly that, from the outside, using only the public LoC figure
• Intelligence Explosion Dynamics — the Greenblatt labor^0.55 × compute^0.45 decomposition and the "both inputs exponential → sustained exponential output" claim are growth-dynamics arguments; this note supplies their labor-side coefficient
• Effective Compute Scaling — the "compute tripling yearly" term in the R&D-speedup reconciliation is the compute-side input that researcher uplift multiplies against
• Research Taste as the Human Bottleneck — the model isolates code/engineering uplift and treats research judgment as the unmodeled residual (g_o = 1, zero non-code uplift); Anthropic's "concentrated in engineering execution rather than research judgment" is that same taste/execution split, and a non-code uplift > 1 would be taste starting to fall
• Recursive Self-Improvement — the reconciliation bears directly on when Anthropic's own 2× R&D threshold trips (Kwa: ~3.5× researcher uplift, "the next year or so"), a concrete near-term marker on the RSI trajectory
• METR — Kwa's employer; the note also leans on METR's own uplift RCT for the verbosity band and the felt-vs-actual-speedup gap
• Anthropic — the subject; both the 8× figure and the "well short of 2×" claim are Anthropic's

Open Questions#

• The whole chain rests on β = 0.5 (pre-AI coding time share), fixed "for simplicity." Kwa flags substantial uncertainty; how much does the 2.3–2.9× band widen once β is varied and measured against Anthropic's actual time-use data?
• Verbosity and value-per-line are the load-bearing unknowns, and both are "at least partially resolvable with internal Anthropic data." Will any lab publish quality-adjusted (not just LoC) code-output measures?
• Greenblatt's 0.55/0.45 labor/compute split is itself an assumption. Is the true R&D production function really that insensitive to labor — and if so, does labor uplift matter far less than the RSI discourse assumes?

Sources#

• Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x — Thomas Kwa (METR), "Because 8 ≈ e², Anthropic's researcher uplift is plausibly >2x" (2026-07-08), practitioner-opinion: Cobb-Douglas / CES / heterogeneous-code production-function models translating 8× code output into ~2.3–2.9× serial researcher uplift; the robustness-to-σ result; verbosity / Cadillac-code / irrationality caveats; serial-uplift-vs-R&D-speedup reconciliation via Greenblatt's labor^0.55 × compute^0.45