Ankur Sethi订阅

A few assorted thoughts about LLMs in November 2026

LLMs are useful tools. They're also the products of the greatest theft of intellectual labor in human history. Both of these things can be true at the same time.
In the right hands, LLMs can produce good, reliable software in shorter timeframes than ever before. Most hands are not the right hands. Most organizations using LLMs to write code are producing slop, no matter what they tell you in their press releases.
If you've never built serious software using an LLM, I urge you to make an honest attempt before you complain about them. You might be surprised.
If you're not using LLMs for code reviews—especially when it comes to security—you're not doing your job right.
LLMs over-engineer everything. Constantly pushing back against their eagerness to write code is exhausting and frustrating.
Software built entirely or partially using LLM assistance is fine, actually.
If an LLM-generated codebase has a security breach, the fault lies completely and fully with the human who generated and shipped that code. We must not allow AI to become yet another accountability sink.
If you use LLMs for writing—emails, text messages, blog posts, anything at all—please don't talk to me. Ideally, don't even sit next to me on the train. I don't want to associate with people who can't use their own words to communicate.
We ought to pass regulations that force the big AI labs to make all their training data public. This should include post-training data, too. Since most of the data used for training LLMs has been acquired unethically by ravaging our digital commons, it stands to reason that it should be released back into the commons.
The AI bubble does’t feel very bubbly to me. The big labs seem to be on the path to profitability. They might stop growing as fast as they're growing, but I don't expect a complete collapse. Source: vibes.
Models need to get faster. Like, 10-20x faster than they are currently. That's going to be the next inflection point for productivity. I don't want a bigger, smarter model. I want today's model to go brrrrr. Bonus points if it can be cheaper, too.
Has somebody figured out continuous learning yet? Is somebody working on it?
Open models will not eat Anthropic/OpenAI's lunch. The big labs will always have one thing your M9 MacBook Pro Max with 256GB RAM won't: infinite compute. That means they can always serve larger models than you can run at home. They can also serve small models at stupid fast speeds. There will always be workloads that require the largest models, and workloads that require smaller models burning more tokens faster. For example, certain long-horizon tasks like the Millennium Prize problems or finding holes in the Linux kernels both respond better to larger models and more token use.
The big labs are in the business of selling compute, not intelligence.
Dario and Sam do not have your best interests at heart. Be careful.
You should keep writing code by hand. You should keep learning computer science. You should have projects where you do not allow yourself to use any LLM assistance. Why? For the love of it, of course. But also to stay sharp, to understand how the sausage is made, to get close to the machine, and to make sure you're never fully dependent on silicon for your cognitive needs.

参与讨论