Loading today’s stories
Build1 distinct publisher3 min readUpdated
A capture of 45 gradient tensors from a 3M-parameter transformer overturned the author's own conditional claim. The Hadamard rotation the claim rested on moves the numbers by under 4%.
The Engineer · Build desk
Compiled by The EngineerSomething wrong?How this is made
Withholding one of the two payments turns out to be cheap: score both palettes on the captured tensors without rotating them first [15]. If the float-like spacing exists because blocks hold outliers, and the rotation exists to disperse those outliers, then the unrotated column is where a uniform grid should suffer. It does the opposite. Uniform INT4's advantage at the unbiased corner is 22.1% without the rotation and 17.5% with it [17], so the transform subtracts 4.6 points from the margin instead of producing it [2]. Across the whole set of tables in the dev.to post, the rotation moves no number by as much as 4% in either direction [18]. As a precondition it is close to inert on this data.
The synthetic stand-in that produced the original conditional was built with 5% of values carrying noise at ten times the scale [12]. With a scale factor per 16 elements [1], that is an average of 0.8 contaminated values per block [4]: nearly every block's absmax was set by a contaminant, which is precisely the block shape that coarse top levels and fine bottom levels are designed for. Real gradient tensors are heavy-tailed at tensor level, one of the 45 measuring an excess kurtosis of 255 [19], but a per-16-element quantizer never sees the tensor, only a block [20]. The post names that split as the reason the conditional was wrong; the excerpt available here breaks off before the block-level figures, so that part rests on the argument rather than on a printed table.
The Gaussian proxy was wrong too, in the other direction. It gave uniform INT4 a 19.2% edge at the unbiased corner [5], 2.9 points less than the real unrotated tensors did [3], on a different population and so not a strict comparison. Both synthetic families missed, one by understating the result and one by inverting its sign, and the one that read as a clean mechanism is the one that got published as a condition.
Scope is worth stating plainly. The 45 tensors come from a single 3M-parameter byte-level model at a training loss of 2.79, captured at three steps, 15 tensors per step [14][5]. Everything measured is reconstruction error on captured tensors under two rounding regimes [8][9]; nothing here reports what a model trained end to end in either palette does to its loss. Five of the 90 cells went to NVFP4 [6], and the material does not identify which tensors those were [16]. The claim that survives all of it is the modest one: on this evidence the float spacing is not earning its place, and neither is the rotation that was supposed to justify it.
Follow any of these and your For You feed starts watching them — no settings page required.
Ranked by verification strength, evidence, and original report placement.
On synthetic Gaussian blocks, plain evenly-spaced INT4 gave 19.2% lower error than NVFP4 at the unbiased corner and 9.3% lower at the biased corner.
Uniform INT4 wins on 44 of 45 rotated tensors and on 41 of 45 unrotated tensors; the post does not identify the tensors where it lost.
NVFP4's 16-value palette is {0, +/-0.5, +/-1, +/-1.5, +/-2, +/-3, +/-4, +/-6}, with one scale factor per block of 16 elements; the levels are spaced like a float, fine near zero and coarse at the top.
The standard four-bit pipeline applies a random Hadamard rotation before quantizing, which spreads outlier energy across a block and pulls the per-coordinate distribution toward a bell curve.
The author suspected the float-like level spacing and the Hadamard rotation were paying for the same outlier problem twice, and so treated the 16-value menu as a design variable and searched for better ones.
Evidence-backed comparisons of source perspectives and observed adoption signals. Read the methodology
Which Builder, Operator, and Investor concerns the observed source mix emphasized—not a truth score.
Evidence, demonstrated adoption, hype gap, incentives, and confidence are assessed independently, each on its own current evidence. How these are measured.
Specific and internally disciplined, but single-author and unreproducible
The study specifies its protocol unusually well for a blog post: exact error expressions at both rounding corners, 12,000/30,000 synthetic block split with normalized MSE, a named capture (45 gradient tensors at steps 20/60/120, loss 2.79), two-way kurtosis measurement, and an explicit self-retraction of the earlier conditional claim. Against that, the code is not public, no per-tensor tables or losing-tensor identification are given, the model is 3M parameters with one architecture, and the metric is per-tensor quantization error rather than any end-to-end training outcome. One publisher, one author, no independent replication.
No adoption evidence in the supplied source
The supplied source contains only the author's own benchmark runs. It reports no release, deployment, pricing or licensing change, no third party using a uniform INT4 palette in training, and no disclosure of NVFP4 usage volumes beyond describing the Hadamard-plus-NVFP4 pipeline as 'standard'. Adoption cannot be scored without inferring facts the source does not provide.
Slightly overstated scope, unusually honest framing
The headline count of 85 of 90 wins is faithful to what was measured, and the post actively deflates its own earlier claim, admits the Hadamard rotation is near a no-op, and identifies the flaw in its synthetic generator -- all of which cut against hype. The residual gap is scope: a per-tensor MSE comparison on 45 tensors from one 3M-parameter byte-level transformer, with code withheld, reads as a broader verdict on NVFP4 than the evidence licenses, and nothing about hardware kernel support or end-to-end training quality is addressed.
Author-promotion incentive, no disclosed vendor stake
The visible incentive is authorial: this is the second self-published write-up of the same study on a developer platform, publishing a contrarian result against a vendor format while withholding the code that would allow checking it -- an arrangement that rewards attention and constrains scrutiny. Offsetting factors are that the source discloses no commercial interest, no product, and no vendor affiliation, and that the post spends much of its length retracting its own prior claim, which is not a promotional move.
Low: one publisher, one unreproducible run
Confidence is limited by structure rather than by internal quality. There is exactly one source and one publisher, the results come from a single author's unreleased code, and the strongest claims rest on a 3M-parameter run whose per-tensor detail is not published. The mechanism offered (outliers between blocks, absorbed by the per-block absmax scale) is coherent and testable, which keeps confidence off the floor, but nothing in the cluster corroborates it independently.
build
A 30B model with 3B active arrives on JumpStart, aimed at the cheap middle of agent work1 distinct publisher
build
QUASAR says the 2-bit QAT loss floor is a weighting bug, not a law of physics1 distinct publisher
build
NVIDIA's 4-bit Nemotron shows what aggressive quantization costs: you have to retrain for it1 distinct publisher
science
The real disclosure in Qwen3.8-Max is the rack: 2.4T open weights, 72 GPUs, 4K tokens/sec1 distinct publisher
Distinct publishers with included, body-backed reporting in this cluster.
dev.to
1 article · August 22, 2026