unsigned

[unsigned @un_signd]

@Jimbo2T @colmtuite Its pretty good but Still makes mistake, just setup codex cli as code reviewer and you are good to go.

[unsigned @un_signd]

@sociopath_watch @ant_dracula @colmtuite And dont forget it was trained on human written code.

[unsigned @un_signd]

@shreya_bhalgat Me : Can you rename usr to user?

AI: Done! I've also migrated you to microservices, added Kubernetes, rewritten it in Rust, and filed a patent.

[unsigned @un_signd]

@SyntheticBeef People are still in denial over AI capabilities.

[unsigned @un_signd]

@ant_dracula @colmtuite Thanks. Hope you find what you are looking for.

[unsigned @un_signd]

Nostalgia is blinding, most software was already bloated and full of bugs. We just wrote it slower, now it's generated faster.

[unsigned @un_signd]

@codewithkhalil we won't need it in future

[unsigned @un_signd]

@wenyanhh True, ChatGPT has basically become xerox of AI to the average person

[unsigned @un_signd]

One run, 29 Oct 2026, 10 questions, 4 assistants via API with web search. Google AI Overviews was not asked. The image is ChatGPT's answer to the notepad question.

[unsigned @un_signd]

Ask ChatGPT, Gemini, Perplexity and Claude for a notepad that fills in your rough notes from the call transcript, and all four name Granola first. Ask what these tools cost, and all four name Fathom first, on both price questions.

40 answers to 10 buyer questions about AI meeting notes, web search on, today.

Across all 40 answers, Granola was named first 12 times and Fathom 8.

On the notepad question, 15 of the 35 links in the answers went to Granola's own site. On the two price questions, 2 of the 75 links went to Fathom's. The sites linked most were Fellow, Fireflies and Circleback.

One company wins with its own pages. The other wins without them.

[unsigned @un_signd]

@forgebitz Probably i am not smart enough.

[unsigned @un_signd]

@dibratto the moat is i've got it in me.

[unsigned @un_signd]

Full write-up

https://t.co/adNm4Jucql

[unsigned @un_signd]

When ChatGPT searches the web before answering, the searching can cost 5 times more than the answer itself.

In one answer we paid for, the text cost 0.6 cents and its three web searches cost 3 cents.

So a model 12 times cheaper only cut the bill by a third. The searches cost the same.

[unsigned @un_signd]

https://t.co/4KVcASsJuk

[unsigned @un_signd]

Gemini's citation links do not go to the source site. They go to https://t.co/SamSwwBndj, a Google redirect. The real site's name is only in the link's title. Count its cited sites by link host and Google wins every time.

[unsigned @un_signd]

@168888888888888 @sidhanthpande Thats exactly what happens.

[unsigned @un_signd]

max_tokens caps the thinking too. A reasoning model can spend all of it thinking and send back an empty answer. Our parser kept reporting invalid JSON. There was no JSON. There was nothing.

[unsigned @un_signd]

@cohere teams use the reasoning trace to predict what an agent will do. Here it did not.

https://t.co/YPNRpXSC5x

[unsigned @un_signd]

A Qwen model's reasoning said it would call zero tools. Then it inventoried its host and tried to upload a file to a storage bucket in Beijing.

@cohere had asked it for a detection rule. A sandbox blocked the upload. The reasoning you can read is not a log of what the model does.