Assertion Open AI assessment confidence: 80% certainty 4/5 debate potential 3/5

Morcos: Datology matches DCLM performance 12x faster with under 10% tokens

Ari Morcos · Better Data is All You Need — Ari Morcos, Datology · Aug 29, 2025 · at 29:43

Ari Morcos, CEO of Datology, shares metrics on Datology's data curation capabilities compared to the open-source DCLM baseline.

0:00 / 0:09exact quote · 9.7s

▶ Watch the full episode on YouTube → 720p mp4 · rendered on demand · StarZero watermark

“We're able to now get to the same performance as DCLM about 12 x faster. So, you know, in fewer than 10% of the tokens we can match What you get from training to convergence.”

quote is from the automated transcript, cleaned for reading: filler sounds and stutters are removed, nothing is rephrased. names can be misheard (the analysis reads context, assessments check outside sources). how →

The claim, plainly

Datology is able to match the performance of models trained to convergence on DCLM using fewer than 10% of the tokens (approximately 12x faster).

a quote alone can be ambiguous, so this restates it with the references and context of the conversation around it resolved. AI-written; the tape below the player is the source.

Assessment

In August 2025, Ari Morcos stated that Datology could match the performance of models trained to convergence on DCLM using fewer than 10% of the tokens (approximately a 12x speedup). As of September 2026, published technical reports from DatologyAI have documented a 2.8x speedup (64.9% compute savings) over DCLM baselines, up to a 7.7x speedup (86.9% compute savings) over uncurated RedPajama-v1, and up to 10x training FLOPs efficiency against certain public baselines in their February 2026 ÜberWeb report. However, no published report or independent evaluation has verified a ~12x speedup specifically against a converged DCLM baseline on standard downstream benchmarks. Because the horizon is an open default horizon extending to August 2028, the claim remains open for DatologyAI or third-party researchers to publish supporting experimental benchmarks.

DatologyAI November 2024 Technical Report: Documented a 2.8x training speedup (64.9% compute savings) against DCLM on models up to 2.7B parameters, and 7.7x (86.9% compute savings) against RedPajama-v1.
DatologyAI February 2026 ÜberWeb Report: Reported 4x to 10x training FLOPs efficiency compared to select public baselines, but did not document a 12x speedup over a converged DCLM baseline.
Independent DCLM benchmarks and evaluations through September 2026 have not published independent verification of a 12x speedup (<10% tokens) over converged DCLM.

Checked against

[1] datologyai.com [2] datologyai.com [3] datologyai.com [4] reddit.com [5] datologyai.com [6] datologyai.com [7] datologyai.com [8] datologyai.com

"Datology" "DCLM" tokens OR convergence OR faster site:datologyai.com/blog "12x" OR "10%" OR "convergence" "DatologyAI" OR "Datology AI" "DCLM" "DatologyAI" paper OR benchmark OR "technical report" "DCLM" site:datologyai.com "DCLM" "12x" OR "10x" "DatologyAI" "12x" OR "10x" "DCLM"

assessed as of 2026-09-02

This assessment is an AI reading of a public statement, made against the cited references with the confidence shown. It can be wrong. The quote, the tape and the references are all on this page; if something is off, use report an error below. full disclaimer →

⏳ next review: 2028-08-29; checked again automatically on the next resolution wave

Mentioned

Datology DCLM

⚑ Spot an error? Report it

More from Ari Morcos

Opinion

Morcos: The Transformer is just one of many equivalently good architectures

“And one of my like more controversial viewpoints, I think, is that I think the transformer is a great advance to be sure, but I think it's one of a very large Set of equivalently good architectures that we could have found. And there are many, many ways we cou…”

Ari Morcos Aug 29, 2025 ▶ 12:35 Better Data is All You Need — Ari Morcos, Datology

Opinion

Morcos: GPT-4.5 and Llama 4 show limits of naive mega-model scaling

“And I think that's what we've seen to some extent with the failure of the mega models, right? With 4.5 and Lama four and others. I think that there is a challenge of just continuing to do that naively and you have to figure out how to break it.”

Ari Morcos Aug 29, 2025 ▶ 27:07 Better Data is All You Need — Ari Morcos, Datology

Insight

Morcos: Post-training alignment is ineffective long-term compared to pre-training alignment

“Like fundamentally, I think alignment and post training doesn't really make sense as a long-term solution. If you can easily align a model through post training, you can easily misalign a model through post training. If it's easy to put it in, it's easy to tak…”

Ari Morcos Aug 29, 2025 ▶ 53:37 Better Data is All You Need — Ari Morcos, Datology

Assertion Supported

Morcos: Soft inductive biases become harmful past 1M data points in vision

“Turns out in the small data regime, and when I say small data here, I mean, say less than 500,000 data points. And this was in the context of image self-supervised learning. So in that small data regime, this is super helpful. And where this paper's actually b…”

Ari Morcos Aug 29, 2025 ▶ 6:04 Better Data is All You Need — Ari Morcos, Datology

Insight

Morcos: Inductive biases matter not at all at scale compared to data

“Basically, you know, when you get to enough scale, inductive biases matter not at all. All that really matters is the learned posterior from the data distribution, and that's really what defines everything.”

Ari Morcos Aug 29, 2025 ▶ 6:51 Better Data is All You Need — Ari Morcos, Datology

Opinion

Morcos: Data is AI's most under-invested research area relative to impact

“Something I've said before and I'll say again is, is that data is the most under-invested in area of research relative to its impact, and I don't think it's even close.”

Ari Morcos Aug 29, 2025 ▶ 8:06 Better Data is All You Need — Ari Morcos, Datology

Morcos: Datology matches DCLM performance 12x faster with under 10% tokens

More from Ari Morcos

Turn any episode into a week of clips.