Comment by coldtea
19 days ago
>Strange since, in practice, coding models have steadily improved without any backward movement every 3-4 months for 2 years now. It's as if there are rigorous methods of filtering and curation applied when building your training data.
It's as if what I wrote implies "all other things being equal", just like any technical claim.
All other things were not equal: the architectures were tweaked, the human data set is still not exhausted, and more money and energy was thrown into their performance since it's a pre-IPO game with huge VC stakes.
We've already seen a plateau non-the-less compared to the earlier release-over-release performance improvements. Even the "without any backward movement every 3-4 months for 2 years now" is hardly arguable. Many saw a backward movement with GPT 4.1 vs 4.0, and similar issues with 4.5, for example. Even if those are isolated, they're hardly the 2 to 3.5 to 4.0 gains.
And no, there are absolutely no "rigorous methods of filtering and curation" that can separate the avalance of AI slop from useful human output - at least not without diminishing the possible training data. The problem after all is not just to tell AI from human with automated curation (that's already impossible), the problem is to have enough valuable new human output, which becomes near a losing game as all aspects of "human" domains previously useful as training input (from code to papers) are tarnished by AI output.
1. No, you dont get to fall back on the technical claim approach. Your bias in your phrasing was clear. Maybe that works for you but I won't just ignore obvious subtext and let you weasel out of this. And that's for the benefit of other readers, not you.
2. A plateau in coding performance? I don't think you even use these models for coding then if you make that claim. It is very clear models have continually improved. You can trust benchmarks to make that clear, or real world use, or better yet: both. You seem to not have the data from either.
3. No rigorous methods of filtering and curation that can separate AI slop from useful human output? Here you go:
a. Curation already works at scale. Modern training pipelines don’t rely on “AI vs human” detection. They filter by utility signals: correctness, novelty, coherence, task success, citation integrity, and cross-source consistency. These measurable properties do correlate with downstream model performance. Models trained on smaller, higher-quality corpora consistently outperform those trained on larger, noisier ones.
b. Human-generated “valuable” data is not shrinking. The claim assumes a fixed pool. In reality, high-value human data is expanding in areas that matter most: expert-labeled datasets, preference comparisons, multimodal demonstrations, tool-use traces, verified code with tests, and domain-expert feedback. These are explicitly created for training and are not polluted by passive AI spam.
c. Synthetic data is not a dead end—when constrained. Empirically, filtered and goal-conditioned synthetic data (self-play, distillation, adversarial generation) improves reasoning, math, coding, and tool use. The failure mode is unfiltered synthetic recursion—not synthetic data per se. This distinction is already operationalized in production systems.
d. Training value ≠ raw text volume. Scaling laws shifted: performance now tracks effective compute × data quality, not sheer token count. A smaller dataset with higher signal density produces better generalization than a massive, contaminated corpus. This is observed repeatedly in ablation studies.
----
Again, the above is not for you, as I believe you don't see beyond your cope (yet). It's for other readers who are intellectually curious.