Slightly odd choice of models to benchmark, isn't it?
@petergostev
-
Flux Schnell Remains Only Cheap Open Source Image Model
By
–
Flux Schnell, Dev & Pro were launched 13 months ago now, this was a big day for image model space. Since then we haven't really had any good cheap open source models – Schnell is still the only good cheap open source model around, but it is getting outdated. Maybe Schnell v2… https://t.co/2Sl5yQtFwD
— Peter Gostev (@petergostev) 29 août 2025Flux Schnell, Dev & Pro were launched 13 months ago now, this was a big day for image model space. Since then we haven't really had any good cheap open source models – Schnell is still the only good cheap open source model around, but it is getting outdated. Maybe Schnell v2
-
Nano-banana: Express True Feelings on LinkedIn
By
–
With nano-banana you can now convey your true feelings on LinkedIn
-
Codex Cloud Generations Review Experience and Interface
By
–
It was cool to be able to review the Codex cloud generations locally, but it was not slick to review 4x generations, the first one was good but the review the second I had to deal with conflicts etc. Maybe I did something wrong, but it would have been nice to just easily flip
-
Testing AI Capabilities: Puzzles, Word Search, and Sudoku
By
–
I tried where is waldo (it drew a new one ), word search, sudoku, and there were some toddler level puzzles too. I couldn't get any to work, but that's reasonable to be honest. I'm a bit unsure what would be informative to test, since eg basic maths might not be interesting if
-
DeepSeek v3.1 Thinking Model Shows Strong Coding Capabilities
By
–
New DeepSeek v3.1 (Thinking) is quite good at coding. A little bit buggy but quite a nice design style https://t.co/HSmG7ppPv7 pic.twitter.com/mP1CL3E4bb
— Peter Gostev (@petergostev) 21 août 2025New DeepSeek v3.1 (Thinking) is quite good at coding. A little bit buggy but quite a nice design style
-
Custom AI Model Selection and System Prompts in Google Sheets
By
–
It's funny I had a script working in Google sheets doing this for about 18 months – the difference though is that I can select the models and a system prompt. One problem I have with these kinds of implementations is that they treat it as a generic 'AI', but I care a lot what
-
Reasoning Models Generate Significantly More Code Than Previous Models
By
–
The amount of code it was actually outputting was kind of crazy, especially comparing it to previous non-reasoning models, it is so much better than older ones
-
Reasoning Capabilities Differ Significantly Across AI Models
By
–
How much does ‘reasoning’ matter for different models? It matters a lot for GPT-5 and less for models like Opus 4.1 and 4.0. From looking at the reasoning traces, models clearly ‘think’ differently: Opus and Sonnet tend to ‘plan’, laying out how it would solve the problem,