Stephen
@0xSMW
Followers
1K
Following
17K
Media
2K
Statuses
14K
optimizing LLM-powered apps / co-founder @klu_ai
BKK
Joined May 2008
“We are choked with news, and starved of history.” — Will Durant https://t.co/04gpLaKfGQ
5
0
29
well well well
Supposedly gpt 5 was based on the 4o pretrain. 4.5 and after preteian was all tossed.
0
0
0
gemini nano banana pro is really something unique and special when it comes to text coherence and complexity
0
0
0
everyone knows you scale on Thursday, so that you can research on Friday
If you want to have a peaceful Thanksgiving this year, please make a rule with your family not to discuss whether AI is in an age of scaling or an age or research.
0
0
0
man the style sucks, but the info is legit
1
0
0
somehow the best 2025 idea seems like a 2005 idea: a website that curates the best human sites and writers on the web, using rss to keep it updated
1
0
0
I was skeptical of LLMs providing time estimates because most generated plans provide multi-week plans for education or engineering tasks. But then I took the prompt used in the paper and tried it with a few models using my own conversations, and it's pretty close. I noticed
anthropic.com
Anthropic economic research on productivity gains
0
0
1
codex-5.1 found a unique issue in rss generation where I pulled the global update time instead of the post's opus-4.5 found 6 issues, all non-issues that were deliberate design decisions
0
0
0
gemini-3 still has the tool call issues found in gemini-2.5 — however, unlike codex-5.1-max and opus-4.5, it spotted an issue impacting my vercel deployment
1
0
0
probably good for SF folks but really missed the mark for mom
Introducing shopping research, a new experience in ChatGPT that does the research to help you find the right products. It’s everything you like about deep research but with an interactive interface to help you make smarter purchasing decisions.
0
0
0
the claude family, now including opus 4.5, are still not in this club. opus 4.5 spends 20k tokens and provides the same incorrect answer as previous generations. anthropic's approach to reasoning is unlike the other labs and it shows in math/science problems.
gemini 3 now joins this club
0
0
0
then thought for 8 mins to answer "how far is the moon", not sure how this model is usable for local inference
0
0
0
and somehow, no amount of instructions to be concise, minimize chattiness, and not use bullets seem to matter — it's very smart, and also a bad communicator compared to o3.
0
0
0