How to Scale Your Model – Google DeepMind
Comments on “How to Scale Your Model – Google DeepMind”
Claim-checked. Scored by all three judges, and its checkable claims were compared with the sources it cites. How far stories are checked
Google DeepMind's JAX team has published "How to Scale Your Model," a free online book that teaches roofline analysis for training and serving Transformers on TPUs, with a new section on GPUs. It covers data, tensor, pipeline and expert parallelism, plus worked LLaMA 3 examples and JAX profiling. Its central claim is that most models now run near hardware limits, so efficiency estimates matter. Several historical and cost claims are unsourced, and the book is still a draft.
Why it ranked
No. 5 of 127 with a score of 0.73. The panel rated it highest on new information (0.88) and reader interest (0.82), lowest on substantiated (0.61). The judges disagreed most on substantiated (spread 0.40), so the story is marked contested. Placed No. 5, No. 10 by score alone: the placement terms moved other stories past it.
- AI judgeJudge 2claude-haiku-5-5
- A long, detailed educational book on LLM scaling, roofline analysis, TPU/GPU hardware and parallelism; highly technical and practical for builders, though it is a tutorial-style series rather than news.
- AI judgeJudge 3gpt-5.6-luna
- A technically substantive, practical guide to Transformer scaling, TPU/GPU rooflines, parallelism, profiling, and JAX. It is broad and actionable for ML systems builders, with minor overstatement but substantial explanatory detail.
Each judge wrote its line after answering the rubric, with the site and the author still hidden. Every answer behind this is in the table under Score.
Score
0.73total under rubric v2.3Contestedjudges differ by 0.40 on substantiated
- Significance0.71
- Reader interest0.82
- Audience breadth0.65
- Practical relevance0.80
- Substantiated0.61
- Technical depth0.80
- New information0.88
- Counts against: Promotional0.06
- Counts against: Overclaim0.14
Each judge’s answer per question (0–1), their mean and spread, and what the question adds to the score (weight × mean).
| Question | Judge 1 | Judge 2 | Judge 3 | Mean | Spread | Weight | Contribution (weight times mean) |
|---|---|---|---|---|---|---|---|
| SignificanceHow much does this change what people in tech know or can do? | 0.75 | 0.60 | 0.78 | 0.71 | 0.18 | +0.24 | +0.170 |
| Reader interestWould a well-informed senior engineer or tech founder want to read this today, and likely pass it on? | 0.69 | 0.85 | 0.92 | 0.82 | 0.23 | +0.26 | +0.213 |
| Audience breadthHow many people who build technology does it concern? | 0.61 | 0.70 | 0.65 | 0.65 | 0.09 | +0.13 | +0.085 |
| Practical relevanceHow useful is it to someone who builds software or tech products? | 0.73 | 0.80 | 0.86 | 0.80 | 0.13 | +0.22 | +0.175 |
| SubstantiatedHow well is it backed? Consider whether it is the primary source, and whether its main claims come with data, code, a demo, a reproducible method or citations. | 0.42 | 0.60 | 0.82 | 0.61 | 0.40 , above the contested threshold | +0.05 | +0.031 |
| Technical depthHow technically substantive is it? | 0.77 | 0.80 | 0.84 | 0.80 | 0.07 | +0.05 | +0.040 |
| New informationDoes it contain information absent from the stories covered in the last 7 days (list provided)? | 0.91 | 0.80 | 0.93 | 0.88 | 0.13 | +0.05 | +0.044 |
| PromotionalIs it mainly marketing with little substance (a press release, ad or sponsored post with no real information)? A launch post that carries real news is not promotional. | 0.09 | 0.05 | 0.04 | 0.06 | 0.05 | −0.19 | −0.011 |
| OverclaimDo the headline or conclusions claim more than the body supports? | 0.21 | 0.10 | 0.12 | 0.14 | 0.11 | −0.11 | −0.016 |
| Sum of contributions | +0.731 | ||||||
| Claim-check penalty | 0.000 | ||||||
| Total, as published | 0.731 | ||||||
Spread is the highest minus the lowest judge answer. marks a spread above 0.40: the judges disagree, and the story is marked contested.
The badge for this ranking, for the author's page or anyone's: one line of HTML, and it links back here.
<a href="https://rubric.hrih.si/editions/2026-10-30/5/"><img src="https://rubric.hrih.si/badge/2026-10-30/5.svg" alt="Ranked #5 on Rubric, 30 Oct 2026" width="238" height="28"></a>Claim check
Claim check: 7 no source given · 1 couldn’t check. Each claim is compared with the source the article itself cites; a flag is published only when both checkers agree.
No source givenFirst, only or fastest
Matrix multiplication uses far more FLOPs per byte of memory than almost any other algorithm (N FLOPs per byte).
The article gives no link or citation for it.
The article says “Matrix multiplication is a unique algorithm in the sense that it uses far more FLOPs per byte of memory than almost any other (N FLOPs per byte”
The source says No source to compare with.
No source givenCausal conclusion
Early TPUs with their systolic array architecture achieved far better performance per dollar than GPUs did at the time they were built.
The article gives no link or citation for it.
The article says “early TPUs and their systolic array architecture achieved far better perf / $ than GPUs did at the time they were built”
The source says No source to compare with.
No source givenNumber
A chip rated at 500 trillion operations per second can do a tenth of that if bogged down moving parameters around in memory.
The article gives no link or citation for it.
The article says “Your shiny new TPU or GPU may be rated to perform 500 trillion operations per second, but if you aren't careful it can just as easily do a tenth of…”
The source says No source to compare with.
No source givenNumber
Hardware designers must bet on what algorithms will look like when chips become available, often 2 to 3 years later.
The article gives no link or citation for it.
The article says “you have to bet on what algorithms will look like when the first chips actually become available, often 2 to 3 years down the road”
The source says No source to compare with.
No source givenGeneralization
Five years ago ML had a varied landscape of architectures, but now mostly just the Transformer remains.
The article gives no link or citation for it.
The article says “Five years ago ML had a colorful landscape of architectures — ConvNets, LSTMs, MLPs, Transformers — but now we mostly just have the Transformer”
The source says No source to compare with.
No source givenCausal conclusion
Alex Krizhevsky had to write difficult CUDA code to make CNNs fast, but within a couple of years libraries like Theano and TensorFlow removed that need.
The article gives no link or citation for it.
The article says “Alex Krizhevsky had to write unholy CUDA code to make CNNs fast but within a couple of years, libraries like Theano and TensorFlow meant you didn't have to”
The source says No source to compare with.
Couldn’t checkNumber
There are four primary parallelism techniques for splitting models over multiple chips: data, tensor, pipeline, and expert.
The source was unreachable, paywalled or too long, or the two checkers disagreed.
The article says “there are four primary parallelism techniques used to split models over multiple chips (data, tensor, pipeline, and expert”
The source says No passage could be compared.
Source cited by the article: jax-ml.github.io/scaling-book/training
No source givenGeneralization
Even small models now run so close to hardware limits that novel research requires thinking about efficiency at scale.
The article gives no link or citation for it.
The article says “today even “small” models run so close to hardware limits that doing novel research requires you to think about efficiency at scale”
The source says No source to compare with.
Also covered by
No other coverage was clustered with this story.
Elsewhere
Elsewhere: HN 2 pts, not on front page
Recorded for context only, checked 30 Oct 2026, 05:47 UTC. It played no part in the ranking.