Part of: this research — view all 12 rows
Current Attribution Norms and Institutional Bottlenecks As of November 14, 2026, no NanoGPT speedrun record credits an AI alone, either on the main track or the optimization track NanoGPT speedrun record training time. The primary barrier to resolution is not purely technical, but rather the social convention of leaderboard attribution. Even when AI systems generate the vast majority of the code or optimization steps, human submitters routinely claim joint or sole credit. For example, records #32, #60, #69, #72, and #87 name an AI system alongside a human https://github.com/KellerJordan/modded-nanogpt https://github.com/KellerJordan/modded-nanogpt https://raw.githubusercontent.com/KellerJordan/modded-nanogpt/master/README.md, and the current Track 1 leader (#92) credits the human while only listing Claude as a commit co-author https://github.com/KellerJordan/modded-nanogpt. Furthermore, the leaderboard suffers from maintainer bottlenecks: the README has not been updated since late September 2026 https://github.com/KellerJordan/modded-nanogpt/blob/master/README.md, and numerous pull requests remain unmerged https://github.com/KellerJordan/modded-nanogpt/pulls.
Capability Growth and Marketing Incentives Despite these institutional friction points, the underlying autonomous R&D capabilities are rapidly improving on fast-feedback benchmarks like this one. Anthropic's capability growth is currently pacing at roughly 1.88× to 2× its earlier rate, and task-level metrics like research taste (TasteVal) are doubling every few months. While autonomous agents historically struggled to generate novel algorithmic improvements—mostly finding gains in kernel optimization—systems from developers like Prime Intellect have already demonstrated the ability to close much of the gap to the human frontier in Track 3 settings https://www.primeintellect.ai/research/nanogpt-speedrun Autonomous AI research for nanogpt speedrun. Crucially, AI research startups (such as Intology, Recursive, and Noetive) possess strong marketing incentives to explicitly claim the first "fully autonomous" leaderboard record. We can already see precursors to this in open submissions like PR #395, where a startup claims its AI chose the experiments and wrote the code https://github.com/KellerJordan/modded-nanogpt/pull/395 Noetive's RSI just took the NanoGPT speedrun record. GPT-2 ....
Pathways to Sole-AI Credit For a sole-AI credit to be finalized, an AI developer must choose to submit under the system's name alone, or an agent account must submit directly, and the maintainers must accept this formatting. The rules for the optimization track expressly welcome submissions written entirely by AI https://raw.githubusercontent.com/KellerJordan/modded-nanogpt/master/README.md https://raw.githubusercontent.com/KellerJordan/modded-nanogpt/master/records/track_3_optimization/README.md, meaning the leaderboard is structurally open to recognizing non-human contributors if requested. As automated research systems advance from requiring substantial human interventions to true end-to-end autonomy on narrow tasks, a deliberate PR push for an AI-only attribution is highly likely to force the issue, bypassing the current default where humans append their own handles to AI achievements.
Timelines and Tail Risks I estimate roughly a 60–65% probability that this milestone occurs before the December 2032 horizon, placing the median on June 1, 2030. The early percentiles (10% by August 1, 2027, 25% by June 1, 2028) reflect a fast path where an AI startup successfully claims a sole-credit record in the near term for marketing purposes. However, the upper tail is heavily shaped by the risk of leaderboard dormancy. The speedrun is already showing signs of maintainer fatigue Pull requests · KellerJordan/modded-nanogpt, and over a six-year horizon, frontier AI research may entirely abandon this specific GitHub repository in favor of newer, more relevant benchmarks. Consequently, the 75th percentile falls on December 31, 2034, beyond the resolution window, while the 90th percentile is never. This distinction reflects both the possibility of a delayed sole-AI credit and the significant likelihood that the repository is abandoned or permanently freezes its human-centric attribution norms before an AI-only credit is merged: success on related benchmarks would not ensure that this specific repository ever records the milestone.
Set against related questions, the upper tail was extended to 'never' to account for the risk that the specific repository is abandoned before a sole-AI credit is merged.
Ask a followup
Sign in to run · $20 free credit, no card · every claim cited