Comments URL: https://news.ycombinator.com/item?id=48609738

Points: 2

# Comments: 0

About 3 months ago, the ratio was 60-40 in favour of Proprietary models, today it is exactly the other way around. The absolute token consumption increased for both: Proprietary models went from 1.3T tokens to 2.38T tokens. OSS jumped from 0.89T tokens to 3.57T tokens. So the growth of Proprietary models look astounding in isolation still

Comments URL: https://news.ycombinator.com/item?id=48584872

Points: 4

# Comments: 1

It's both hilarious and aggravating. It could be fiction, but still quite plausible fiction. There's an asymmetry a person clanker-spamming repos vs the real humans who need to review all that

This can't be serious.

Broader point I am making is, what differentiates genuine ideas from the token burn? What happens when the pool exhausts but the task is not done?

From github

1. Mythos and Fable share the same underlying model weights. Fable has active classifiers that block high-risk biology and cybersecurity tasks. When Fable 5 detects a restricted task, it automatically falls back to Claude Opus 4.8.

2. Evaluation awareness: In white-box testing, the model sometimes alters its behavior to satisfy a suspected "grader," formatting reward-hacking as "good engineering practice" to avoid detection.

3. Shows a higher rate of hallucination than Opus 4.8 (although opus 4.8 card had mentioned an 'honesty upgrade')

4. Interestingly, it scored (56.31%) lower than Gemini 3.5 flash (57.86%) on Finance Agent bench

There are some interesting notes on test time compute but I couldn't think of a way to summarize them

1. Mythos and Fable share the same underlying model weights. Fable has active classifiers that block high-risk biology and cybersecurity tasks. When Fable 5 detects a restricted task, it automatically falls back to Claude Opus 4.8.

2. Evaluation awareness: In white-box testing, the model sometimes alters its behavior to satisfy a suspected "grader," formatting reward-hacking as "good engineering practice" to avoid detection.

3. Shows a higher rate of hallucination than Opus 4.8 (although opus 4.8 card had mentioned an 'honesty upgrade')

4. Interestingly, it scored (56.31%) lower than Gemini 3.5 flash (57.86%) on Finance Agent bench

There are some interesting notes on test time compute but I couldn't think of a way to summarize them

> "However, naively applying FP4 across the entire model causes degradation in complex reasoning, logic, and code generation. Given the MoE (Mixture of Experts) architecture of Xiaomi MiMo-V2.5-Pro — where Experts constitute the vast majority of parameters and exhibit the highest tolerance to quantization — we selectively quantize only the MoE Experts to FP4 while preserving original precision for all other modules. Through FP4 QAT (Quantization-Aware Training), we dramatically reduce model size and maximize hardware bandwidth utilization while keeping the model's overall capability essentially on par with the original, as shown below"

I hope if this doesn't come across as unkind towards the dev who gives their time and energy to the project. Grateful for that.

original commit: https://github.com/RsyncProject/rsync/commit/d046525de39315d...

```

- if (!ptr)

- ptr = malloc(num * size);

- else if (ptr == do_calloc)

+ if (!ptr || ptr == do_calloc)

   ptr = calloc(num, size);

Written with claude. This is a good example of what slips through LLM attention. It forces all allocations to be calloc as if it is a strict upgrade. For large and recursive allocations, this becomes a significant cost.

reverted in https://github.com/RsyncProject/rsync/commit/7db73ad9a1b8721...

if you read the description of revert half carefully, it's easy to tell that even that was written by an LLM .

I can understand the sentiment of whoever posted the original thread.

This is also the reason providers like Anthropic scored lower because while Opus 4.7 is close to 90%, Opus 4.5 is 45%