- Models: Not commoditized but evolving rapidly. Intelligence increasing, costs dropping 10x yearly. GPT-5 will merge GPT and reasoning model series.
- Coding Automation:
- O1 Preview: ~millionth best programmer
- O1: ~thousandth best
- O3 (upcoming): 175th best globally
- Will surpass humans at programming this year/next, not 2027 as Anthropic suggested
- Deep Research:
OpenAI's best product since ChatGPT. Provides capabilities users couldn't achieve alone.
- Model Strengths:
- GPT-4.5: Better at writing, human-like interaction
- O1/O3: Better at reasoning, structured problems
- Large models encode more "subtlety and nuance"
- Future Applications:
- AI tutoring for personalized education
- Robotics as the next frontier
- Creation Value: Expertise + AI will still outperform.
Value shifts toward idea generation, management, and quality assessment.
If ChatGPT was reallly the 1000’th best programmer in the world… do you know what you could do with that?
Makes me wonder, are they testing this in real world scenarios or just benchmarking on Hackerrank, because 1000th best programmer in the world would be AMAZING!!! You could probably 10x the productivity of any engineering org if all engineers were suddenly 1000th best in the world. There’s absolutely no need to surpass human ability, which makes me think they’re measuring something tangentially related to the core skill of the job.
I think this “coding will be automated” is that “the interview(hard) the job(easy) meme” but backwards.
It seems very obvious that an AI will solve all the hardest coding problems first because they’re all documented really well, but fuck if I can get Claude to figure out a basic css layout. In wrestling with Cursor and tailwind now.
Interesting how our experiences are so different. IMHO CSS is what the LLM’s are best at, out of all code editor tasks I have thrown at them. It’s trivially true that there is more CSS and JS in their training data than anything else since that’s what every single website in the world contains.
You know what, let them do it. I'm ready to retire anyway.
I want to see what happens when somebody looks at the high-rise sand castle that is modern technology and decides that there wasn't enough sand involved.
I wish I knew. If I had to venture a guess, it would be to get a job with a company that doesn't indulge in this hubris.
I think enough companies are being stung by their hastily made AI decisions that the option of losing billions to save millions again isn't that appealing. That requires grown-ups to be involved in making decisions, though.
Kevin Weil (OpenAI CPO) Interview Summary
- Models: Not commoditized but evolving rapidly. Intelligence increasing, costs dropping 10x yearly. GPT-5 will merge GPT and reasoning model series.
- Coding Automation:
- Deep Research: OpenAI's best product since ChatGPT. Provides capabilities users couldn't achieve alone.- Model Strengths:
- Future Applications: - Creation Value: Expertise + AI will still outperform. Value shifts toward idea generation, management, and quality assessment.If ChatGPT was reallly the 1000’th best programmer in the world… do you know what you could do with that?
Makes me wonder, are they testing this in real world scenarios or just benchmarking on Hackerrank, because 1000th best programmer in the world would be AMAZING!!! You could probably 10x the productivity of any engineering org if all engineers were suddenly 1000th best in the world. There’s absolutely no need to surpass human ability, which makes me think they’re measuring something tangentially related to the core skill of the job.
I think this “coding will be automated” is that “the interview(hard) the job(easy) meme” but backwards.
It seems very obvious that an AI will solve all the hardest coding problems first because they’re all documented really well, but fuck if I can get Claude to figure out a basic css layout. In wrestling with Cursor and tailwind now.
Interesting how our experiences are so different. IMHO CSS is what the LLM’s are best at, out of all code editor tasks I have thrown at them. It’s trivially true that there is more CSS and JS in their training data than anything else since that’s what every single website in the world contains.
You know what, let them do it. I'm ready to retire anyway.
I want to see what happens when somebody looks at the high-rise sand castle that is modern technology and decides that there wasn't enough sand involved.
How is anyone that is a software engineer that has been replaced by AI going to make rent?
I wish I knew. If I had to venture a guess, it would be to get a job with a company that doesn't indulge in this hubris.
I think enough companies are being stung by their hastily made AI decisions that the option of losing billions to save millions again isn't that appealing. That requires grown-ups to be involved in making decisions, though.
Go into the trades.
The Anthropic guy also said one year recently. It sounds like they're feeling pressure from investors. AI bubble deflates in one year.
snakeoil.
We'll know this is coming for realz when they first automate clueless marketroids saying things they don't understand and fire Kevin