Bandit

Like all research, we do our best with the data we have. But we don't publish anything if we don't feel confident in our findings, especially if our data has notable blindspots. In this case, if we do have blindspots, they would equally affect our view into OpenAI and Anthropic. So any billpay we don't capture would dampen the impact on both...equally. The finding holds regardless

GPT-5.4-mini is dead on arrival you can just use Kimi-K2.5 for $1.5 less and better performance

I'm getting 2x ratiod for being right. the absolute state of this place man

should we do themes

If we assume GPT 5.4 is already doing about 28% of the total work needed to create GPT 5.5, and AI’s share of the model building pipeline rises by 10 percentage points every .5 iteration, then the curve gets steep fast: GPT 5.5 = 28% GPT 6.0 = 38% GPT 6.5 = 48% GPT 7.0 = 58% GPT 7.5 = 68% GPT 8.0 = 78% GPT 8.5 = 88% GPT 9.0 = 98% if we assume 1 full digit model iteration per year 2029 the models should start creating themselves, in other words AGI perfectly as ive predicted. only the people at OAI know the close % of total work 5.4 is doing in terms of research direction and architecture, data filtering, labeling, synthetic data work, experiment coding and debugging, training infrastructure scripts, optimization support, evaluations, error analysis, benchmark generation, safety review, release review, productization, and integration. I assume its less than 28% but that was 5.4s guess

OpenAI seems to have accidentally rolled out Skills in ChatGPT for some users, but only partially