DoubleChecks

What was claimed

Researchers can now pretrain a transformer without backpropagation in 1.6 seconds (4,000x speedup from 1.76 hours) and even beat backprop+Adam on validation loss

Our verdict

Needs caution

The project reports a forward-only track reaching 3.676, past a backpropagation Adam result of 3.769, but also states that backpropagation with the same optimizer remains ahead. This is not evidence that the 1.6-second speedrun beats the matched backpropagation baseline.

Combined assessment28 sourcesChecked Oct 23, 2026

Why this verdict

  • The method beats backpropagation plus Adam on validation loss.

    Misleading
    One assessment reached this conclusion; two differed·Claude: Incorrect, ChatGPT: Verified

  • Researchers can now pretrain a transformer without backpropagation in 1.6 seconds (4,000x speedup from 1.76 hours).

    Verified
    Two assessments reached this conclusion; one differed·Perplexity: Misleading

  • The result represents a 4,000x speedup from 1.76 hours.

    Verified
    All assessments reached this conclusion

Evidence reviewed

28 sources

Perplexity, Claude and ChatGPT sources are combined here.

explainx.aireddit.comqlabs.shqlabs.shkaggle.comhn.todayx.comdigg.comarxiv.orgar5iv.labs.arxiv.org

Individual assessments

Assessment summary

The speedup claim is accurate only for a specific Noise benchmark target: reaching validation loss 4.2 improved from 1.76 hours to 1.6 seconds.explainx.ai The claim that the method beat backpropagation plus Adam is misleading because the speedrun comparison uses different tracks and the project page says backpropagation with the same optimizer remains ahead.explainx.ai

Claims reviewed

Researchers can now pretrain a transformer without backpropagation in 1.6 seconds.

Misleading
A 1.6-second result is reported for reaching a fixed validation loss of 4.2 in the Noise speedrun, not for completing general transformer pretraining.explainx.ai

This claim may mislead readers without added context.

The method beats backpropagation plus Adam on validation loss.

Misleading
The project reports a forward-only track reaching 3.676, past a backpropagation Adam result of 3.769, but also states that backpropagation with the same optimizer remains ahead.explainx.ai This is not evidence that the 1.6-second speedrun beats the matched backpropagation baseline.

This claim may mislead readers without added context.

The result represents a 4,000x speedup from 1.76 hours.

Verified
The Noise project reports reducing the time to reach validation loss 4.2 from 1.76 hours to 1.6 seconds, which is slightly under 4,000 times faster.explainx.ai

Assessment summary

  • The claim contains technically accurate elements but is significantly misleading in context.
  • The 1.6-second speedup and 4,000x figure are real in the Noise speedrun benchmark, but they represent a community record to a loose loss target (4.2), not Dust's original performance.
  • Critically, the claim that Dust "beats backprop+Adam on validation loss" is false—Dust trails backprop with Adam, and only matches or beats backprop with SGD at certain scales.

Claims reviewed

Dust beats backprop+Adam on validation loss

Incorrect
Multiple sources confirm Dust 'still trails backprop under Adam' papers.neurips.cc. The Noise leaderboard shows backprop+Adam achieves 3.419 validation loss while Dust+Adam achieves 3.767 stork.ai, meaning backprop is superior. Dust only matches or beats backprop when using SGD, not Adam.

This claim may mislead readers without added context.

Researchers can pretrain a transformer without backpropagation in 1.6 seconds

Verified
The 1.6-second figure exists in the Noise speedrun leaderboard arxiv.org. However, this is a community record using novel methods beyond the original Dust algorithm, targeting a loose loss threshold of 4.2 rather than Dust's baseline performance.

This represents a 4,000x speedup from 1.76 hours

Verified
The speedrun leaderboard shows the time to reach validation loss 4.2 fell from 1.76 hours to 1.6 seconds arxiv.org, which is approximately 4,000x faster. This speedup is mathematically accurate.

Assessment summary

  • The core numerical and comparative claims are supported by the public Noise speedrun report.
  • The Noise page documents a record drop from 1.76 hours to 1.6 seconds (≈4,000×) and reports a method (Dust and subsequent entries) that surpassed backprop+Adam on the cited validation-loss metric.

Claims reviewed

This approach can even beat backprop+Adam on validation loss.

Verified
The Noise report and related posts state that track-1 results moved past backprop with Adam (reported Adam validation loss 3.769) to a lower validation loss (e.g., 3.676), meaning a better validation-loss number was achieved without backprop. qlabs.sh github.com

Researchers can now pretrain a transformer without backpropagation in 1.6 seconds (4,000x speedup from 1.76 hours).

Verified
The Q Labs Noise speedrun page records a speedrun record of 1.6 seconds compared to the prior 1.76 hours, which is approximately a 4,000× reduction in time. qlabs.sh

Check your own claim

Paste any statement, headline, or AI answer. Independent assessments check it and compare sources.

Share this result