Useful AI efficiency

Area: definition. Apply the shared measurement protocol.

Estimated progress toward AI-managed chips

40% complete — judgment-based estimate as of November 7, 2026. Confidence: low. The scale starts at 0% before ChatGPT's public release in 2022 and ends at 100% when AI manages the field's continuing advancement without human participation. This rates the outcome's contribution to that transition, not the percentage of tasks solved, a measured fraction of the field, or time remaining.

The shared scoring rubric uses five equally weighted dimensions: execution, breadth, agenda-setting, validation, and sustained autonomous improvement. Points for this outcome are 10 + 10 + 5 + 10 + 5 = 40/100, in that order; each dimension has a maximum of 20. These component judgments make the estimate inspectable, but are not independently measured scores.

Reason for the estimate: Standardized inference benchmarks support reproducible efficiency assessment across systems. Throughput gains alone do not establish AI-directed gains in full energy and cost efficiency. The dated baseline below supplies the findings and source links used for this assessment. Unmeasured components receive no demonstrated-progress credit; missing evidence does not prove a missing capability.

What would move the score: Compare qualified submissions at matched accuracy and latency; include measured system power, memory, networking, utilization and amortized costs. Higher autonomy scores additionally require AI to select worthwhile work, validate results and improve the process over repeated cycles without human research management, repair or adjudication. Lower the score if replication fails, quality regresses or previously hidden human assistance is found. Preserve this dated estimate when adding future assessments.

Measurement and acceptance rule

Joules and total dollars per training run reaching fixed quality, or per inference meeting fixed accuracy and latency. Include memory and communication overhead.

Evidence requirements

Separate simulation, physical-design signoff, and fabricated-silicon measurements. Compare against a strong contemporary electronic-design-automation workflow with equal tuning budgets. Document the process node, foundry libraries, area limits, thermal limits, workload, and host system. Do not attribute a process-node advance or larger chip solely to AI design.

Related milestones

See the area's proposed milestones. Proposed targets are distinct from the dated evidence below.

Baseline as of November 7 2026

Status: Measured throughput trend; energy/cost attribution incomplete. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

September 17, 2026: MLCommons reports median per-accelerator Llama 2 70B Server throughput improving 5.58-fold across six rounds since v4.0 in early 2024. MLCommons: MLPerf Inference v6.1.

Interpretation and tracking value

Hardware, software and numerical precision changed. Throughput is not joules or total dollars per fixed-quality result, nor proof of AI-designed hardware gains.

Next comparable measurement

Compare qualified submissions at matched accuracy and latency; include measured system power, memory, networking, utilization and amortized costs. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.