Today I was days old when I learned that Claude projects can’t reference past conversations within a project. chatgpt projects do have that feature and it’s super useful for getting the past brainstorms in the context. I just convert the chat to markdown and add it as a project

openai's o3 just loves tables way too much 😭 i literally told it in my personalization settings: no tables unless i ask or it's absolutely necessary. and it still gives me tables with this little disclaimer: “The table is small enough that you won’t hate me for it, promise.”

I just hope GPT-5 doesn’t come with a GPT-4o personality, that would be so disappointing.

seeing this behaviour on chatgpt for the first time, seems it had some issue with it's browsing tool and it explicitly called it out. I like this instead of straight up lying or hallucinating.

i don’t understand why people are raving about removing GPT-4o. It had a terrible personality, was overly sycophantic, tbh pretty annoying for me, and was just a decent model, nothing special.

i’ve noticed a weird user experience while using GPT-5 extensively: there’s a sudden change in tone. The thinking model’s tone is much better, it writes better, but within the same conversation, when the router switches to the simple chat model, it basically has the same feel as

lol

is there any study or benchmark comparing quantization methods for the fastest time to first token (TTFT)? I don’t care about inter-token latency, I am concerned with just the speed for predicting the single first token. Should be well supported in vLLM. cc: vLLM

GPT-5 thinking is a great model. Much better than o3 for me. It feels smarter and more thorough than o3. Also it writes better than o3 (which was my biggest gripe with o3) and is overall a really solid model.

If you want to learn how to use LLMs for classification in depth (including fine-tuning), I wrote a comprehensive guide covering the theory and practical implementation a while back. This includes explaining the gotchas for using decoders as classifiers. LLMs as classifiers

GPT-5o 🥲

I barely use Perplexity anymore. ChatGPT since o3, deepresearch and now with gpt-5 thinking is way better for searching, researching, and digging up obscure stuff. Plus it has memory and everything else built in, so it’s basically becoming an all-in-one AI app.

I sometimes burn through 400k tokens with just a couple calls to a reasoning model. Charging ₹999/month for this is daylight robbery for people who don’t get how it works. And if they’re really adding side-by-side outputs, that’s even more token usage per prompt.

GPT-5 thinking can think for a long time 👀

this is the way, i wish every jd is like this.

One really good use-case of GPT-5 thinking is that when I hit certain errors or want to know how things work internally or why a particular framework does a certain thing, I can just tell it to go and look through the source code and docs and it usually nails it.