what are some examples of the tasks behind this chart? it says GPT-5 can complete 50% of tasks that take four hours i dont think anyone truly believes this
LLMS
-

Qwen3-4B First Model to Reject Unrealistic Test Requests
By
–
Qwen3-4B-Thinking (a 4GB download) just became the first model I've tested to directly push back against my pelican riding a bicycle test, calling it "oddly specific and completely unrealistic" and saying "this request violates physics and biology" https://
simonwillison.net/2025/Aug/10/qw
en3-4b/
… -
ChatGPT’s Free Access Generates Massive Consumer Surplus
By
–
I didn't say AI was unique. There are many new goods with high consumer surplus, too. We have papers about them at the link above, e.g. https://
nber.org/papers/w31670 But when hundreds of millions of people suddenly get ChatGPT and other LLMs for free, then the increase in consumer -
Code Quick Fixes: Full Model vs 4.1 Speed Comparison
By
–
Code quick fixes. I only get good results if I use the full model and medium reasoning, whereas 4.1 gives at least as good results 10x faster.
-

AI Model Teaches Itself Math, Algebra, and Coding
By
–
No human labels. No curated datasets. Just one topic prompt.
A new AI model taught itself math, algebra, and coding — and scored double-digit gains — by becoming its own teacher and student. Self-Questioning Language Models — https://
arxiv.org/abs/2508.03682 The setup: The same -

Claude Opus 4.1 De-Carcinizes The Great Gatsby
By
–
“Claude Opus 4.1, De-Carcinize the Great Gatsby” (That was the only prompt) Pretty clever, actually.
-
GPT-5 scaling laws not exhausted despite apparent capability plateau
By
–
Hay mucha gente equivocada: Que GPT-5 no haya avanzado de forma evidente en sus capacidades no es prueba de que la ley de escalas haya llegado a su fin ¿por qué? Porque GPT-5 no es un modelo frontera escalado a tamaños nunca antes visto. Si queréis coger un modelo que tomar

