How we are meant to be letting agents just build stuff, when their defaults are so bad – constantly adding some weird guardrails, fallbacks and they are so non-AI pilled its painful. E.g. instead of letting an agent just deal with a PDF, it defaults to some shitty PDF parser.
@petergostev
-
Aleph-Alpha’s open-weight model trained on 768 B200 GPUs
By
–
And Aleph-Alpha's was trained on 768 B200s, even further behind https://
aleph-alpha.com/en/blog/kolibr
i-has-landed-a-sovereign-open-weight-model/
… -
Mistral Large 4 vs Astra: 4,000 vs 100,000 GPUs
By
–
Mistral Large 4 was trained on 4,000 GPUs and Astra was trained on 100,000 of them.
— Peter Gostev (@petergostev) 6 octobre 2026
We'll see what the kitten can pull off, but the disparity is pretty stark https://t.co/6VimFUcwqYMistral Large 4 was trained on 4,000 GPUs and Astra was trained on 100,000 of them. We'll see what the kitten can pull off, but the disparity is pretty stark
-
Sign in with ChatGPT vs Gemini credits via Google login
By
–
Now we have 'Sign in with ChatGPT', but what about getting Gemini credits when you 'Sign in with Google'? Surely Google has this market to themselves – when I sign in with Google I should have an option to use my Gemini inference & data – seems like they forgot about that
-
Model efficiency and API pricing affect value comparison
By
–
Two problems with this analysis, the 'value' would look worse, if: – A model takes fewer tokens – API costs are lower So in comparison, a more efficient model that has cheap API prices looks like 'worse value'
-
OS architecture for permissionless AI agents
By
–
There will be a time when someone comes up with the right OS architecture where the agents can do what they want with zero permission pop-ups, without your life won't get ruined
-
Attention in an era of perfect AI generation tools
By
–
What's going to get your attention when everyone has access to perfect video, 3D and html generation tools? I feel like I'm already glazing over a lot of this stuff, when only a few weeks/months ago I'd been amazed
-
AI agents and the future of autonomous prompting
By
–
I know nobody cared about this at OpenAI dev day, but I think this is one of the biggest potential future product lines. So far all agents assume there's a human who is taking responsibility & prompting it. One clear way forward is to have an agent that takes full
-
Continuous Learning Benchmark Analysis of Opus 5.5
By
–
Continuous Learning benchmark – Opus 5.5 had a good run of improving performance from game 75 until c. game 160. But then performance declined somewhat. I might continue more after the reset & do the analysis of how they approached learning Watch here: …
https://cheque
-delhi-needs-silver.trycloudflare.com/#/ -
GPT-6.1-Sol usage limits and cost-per-task efficiency vs Astra
By
–
Codex limits feel fine now with GPT-6.1-Sol as the daily driver. With Astra, the limits were a disaster, but now I'm not close to killing the sub. As per Arena's real usage cost-per-task data, you should get 5x more out of GPT-6.1-Sol than Astra.