It's surprising that people are either "AI coding is trash" or "AI will take all the coding jobs." In my experience: - AI coding is indispensable for greenfield work, like building a dashboard. Just create the markdown for explaining the context and what it's supposed to do.

He's also a Karpathy-level taste maker. When he started talking about agents and LLMs working for awhile on their own, so did everyone else.

But there's a reason we don't see a lot of specialist models out there. LLM training is incredibly expensive and there's probably <1000 people really qualified to do it at scale. If you don't have the motivation of 'automate all the work', it's hard to justify the billions.

This is why causality is hard and what we do about it.

A big difference between Claude Code and the Cursor/Cursor-like IDEs is how easy/natural it is to your review the code. Vibe coding is something to defend against, so I bet the long term engineer's favorite will make it painless to review the code avalanche.

Regarding the end of Stack Exchange. One reaction I'm seeing is that this doesn't matter because LLMs will substitute for the answer-ers. Something like "AIs can read the documentation and use it to answer Qs" But if you've been on Stack Exchange, you'll know that much of it is

Recommendable project: use Claude Code to build a personal finance dashboard. Your checking, saving, brokerage, credit card, coinbase, 401K, etc. accounts all have exportable transactions / balances CSV files. Claude has seen the formats before and knows how to normalize it

I generally avoid the hype train, but this is true. The most high leverage thing a coder can do with five hours is to study/practice the Claude Code documentation. It's the tool of tools.

So the new hot skill is reviewing AI code. Some things I've found helpful: Start with top-down, architectural feedback with an aggressive eye to cut complexity. Lines of code are a liability, so trim them. E.g., if you're optimizing a machine learning model for one or two

Claude Code is so good because of product changes, not actually a smarter model. Opus 4.5 isn't that much better than Sonnet on the benchmarks. I've switched between the two a good deal and I can't spot the difference in error rate. What they nailed is: It's CLI-first, which is

For these AI coding tools, there's a hard difference between building a UI and building a model. When the UI is built, you can poke around and get a good sense of it's doing what you want. When the ML pipeline is done and you see a roc-auc of .7120... is that right? The only

Yes, it's uncool to say it but you need to review that pretty AI code. It doesn't even come down to intelligence, IMO. If you've been working on a large system long enough, with many widely separately pieces, you just can't write down all the context.

I miss when AI was less cool. It's crowded with opportunists. Even something like posting about interesting papers is totally saturated with growth hackers, automating every comment with AI. A paper comes out and within a half hour, 50 influencers can somehow tell you

2026 will be a weird year if OpenAI announces an advertising model *and* AGI.

Good code is code that has evolved with feedback. Someone guessed the right shape, tested it, put it out in the world, the world say 'no, not that, this' many times and now the software fits. Writing code is now much cheaper, but these super simplified AI projections assume a