Frontier AI research lab. We build AI scientists & the platforms they run on — compressing the full research lifecycle from idea to impact.
- Tweets 88
- Following 54
- Followers 1,829
- Likes 25
ALT 72.24 minutes to GPT-2 on 8×H100. AutoTrust's ScienceGuru, running Guru Turbo 1.2, posted the fastest result we've found on Andrej Karpathy's Time-to-GPT-2 benchmark (one run, self-reported): 27% under the official record and 9.6 min ahead of the best community recipe. The field Reported times on the official leaderboard and in public nanochat PRs, as of Sept 24: - 72.24 min · ScienceGuru (AutoTrust) · 1 run - 81.84 min · Giovanni Zinzi · 6 runs - 91.74 min · Oriole Networks · 3 runs - 94.58 min · Martin Jurča (Seznam.cz) · 6 runs - 94.6 min · Weco-optimized run · 3 runs - ~99 min · Official record, Karpathy's autoresearch round 2 · 5 runs Zinzi also reported a 73.92-min experiment with a Cosmopedia data mix (3 runs), which he set aside after it regressed at a smaller model size. The lead - 27% faster than the official record (−26.8 min) - 12% faster than the best community recipe (−9.6 min) - 19–22 min ahead of Oriole Networks, Seznam.cz and the Weco-optimized run How Starting fr
ALT 72.24 minutes to GPT-2 on 8×H100. AutoTrust's ScienceGuru, running Guru Turbo 1.2, posted the fastest result we've found on Andrej Karpathy's Time-to-GPT-2 benchmark (one run, self-reported): 27% under the official record and 9.6 min ahead of the best community recipe. The field Reported times on the official leaderboard and in public nanochat PRs, as of Sept 24: - 72.24 min · ScienceGuru (AutoTrust) · 1 run - 81.84 min · Giovanni Zinzi · 6 runs - 91.74 min · Oriole Networks · 3 runs - 94.58 min · Martin Jurča (Seznam.cz) · 6 runs - 94.6 min · Weco-optimized run · 3 runs - ~99 min · Official record, Karpathy's autoresearch round 2 · 5 runs Zinzi also reported a 73.92-min experiment with a Cosmopedia data mix (3 runs), which he set aside after it regressed at a smaller model size. The lead - 27% faster than the official record (−26.8 min) - 12% faster than the best community recipe (−9.6 min) - 19–22 min ahead of Oriole Networks, Seznam.cz and the Weco-optimized run How Starting fr
ALT 72.24 minutes to GPT-2 on 8×H100. AutoTrust's ScienceGuru, running Guru Turbo 1.2, posted the fastest result we've found on Andrej Karpathy's Time-to-GPT-2 benchmark (one run, self-reported): 27% under the official record and 9.6 min ahead of the best community recipe. The field Reported times on the official leaderboard and in public nanochat PRs, as of Sept 24: - 72.24 min · ScienceGuru (AutoTrust) · 1 run - 81.84 min · Giovanni Zinzi · 6 runs - 91.74 min · Oriole Networks · 3 runs - 94.58 min · Martin Jurča (Seznam.cz) · 6 runs - 94.6 min · Weco-optimized run · 3 runs - ~99 min · Official record, Karpathy's autoresearch round 2 · 5 runs Zinzi also reported a 73.92-min experiment with a Cosmopedia data mix (3 runs), which he set aside after it regressed at a smaller model size. The lead - 27% faster than the official record (−26.8 min) - 12% faster than the best community recipe (−9.6 min) - 19–22 min ahead of Oriole Networks, Seznam.cz and the Weco-optimized run How Starting fr
ALT What happens when an AI research system picks up where two years of human optimization left off? The benchmark is the NanoGPT Speedrun: train GPT-2 to 3.28 validation loss on FineWeb, the target set by @karpathy 's llm.c replication, which took 45 minutes to get there. The speedrun's code descends from llm.c's PyTorch trainer, itself descended from NanoGPT, hence the name. Over two years, 91 official records brought the time down to 67.56s. ScienceGuru just hit 24.90s on 8×H100 (five seeds, self-reported): 108× faster than where the benchmark started, 2.71× faster than the official record, and 3× faster than Recursive's June record of 75.4s. Earlier this month it also took #1 on Autoresearch@Home at 0.8895 BPB, ahead of @Recursive_SI 's 0.9109, and the validated lead on @MedARC_AI 's NanoPath v2. We built on open PRs by Deven Pietrzak (ANVIL2) and Herman Brunborg (Exact-match), credited in the repo. ScienceGuru's RSI recipe fused them, cut the schedule from 1,194 to 652 steps and r
ALT What happens when an AI research system picks up where two years of human optimization left off? The benchmark is the NanoGPT Speedrun: train GPT-2 to 3.28 validation loss on FineWeb, the target set by @karpathy 's llm.c replication, which took 45 minutes to get there. The speedrun's code descends from llm.c's PyTorch trainer, itself descended from NanoGPT, hence the name. Over two years, 91 official records brought the time down to 67.56s. ScienceGuru just hit 24.90s on 8×H100 (five seeds, self-reported): 108× faster than where the benchmark started, 2.71× faster than the official record, and 3× faster than Recursive's June record of 75.4s. Earlier this month it also took #1 on Autoresearch@Home at 0.8895 BPB, ahead of @Recursive_SI 's 0.9109, and the validated lead on @MedARC_AI 's NanoPath v2. We built on open PRs by Deven Pietrzak (ANVIL2) and Herman Brunborg (Exact-match), credited in the repo. ScienceGuru's RSI recipe fused them, cut the schedule from 1,194 to 652 steps and r
ALT What happens when an AI research system picks up where two years of human optimization left off? The benchmark is the NanoGPT Speedrun: train GPT-2 to 3.28 validation loss on FineWeb, the target set by @karpathy 's llm.c replication, which took 45 minutes to get there. The speedrun's code descends from llm.c's PyTorch trainer, itself descended from NanoGPT, hence the name. Over two years, 91 official records brought the time down to 67.56s. ScienceGuru just hit 24.90s on 8×H100 (five seeds, self-reported): 108× faster than where the benchmark started, 2.71× faster than the official record, and 3× faster than Recursive's June record of 75.4s. Earlier this month it also took #1 on Autoresearch@Home at 0.8895 BPB, ahead of @Recursive_SI 's 0.9109, and the validated lead on @MedARC_AI 's NanoPath v2. We built on open PRs by Deven Pietrzak (ANVIL2) and Herman Brunborg (Exact-match), credited in the repo. ScienceGuru's RSI recipe fused them, cut the schedule from 1,194 to 652 steps and r
ALT AutoTrust, a tiny Singapore AI startup, just beat Recursive Superintelligence, the $650M-funded RSI pioneer, for the second time. NanoGPT Speedrun: ScienceGuru hit 24.90s on 8×H100 (five seeds, self-reported), 3× faster than Recursive's June record of 75.4s. Autoresearch@Home: ScienceGuru took #1 at 0.8895 BPB, ahead of Recursive's 0.9109. We built on open PRs by Deven Pietrzak (ANVIL2) and Herman Brunborg (Exact-match), credited in the repo. ScienceGuru's RSI recipe fused them, cut the schedule from 1,194 to 652 steps and re-engineered the host, for the fastest time we know of on this benchmark. Try ScienceGuru: https://scienceguru.ai
ALT 1/ Today AutoTrust AI open-sources JEV, a System 1 model distilled from TypeSafe's Jev 1.13: - Nearly indistinguishable from close-sourced Jev (KL 0.021) - Calibrated probabilities, one forward pass - 2.5 ms per decision, batched - Apache-2.0 weights + code https://huggingface.co/autotrust/JEV