Yes definitely. The average quality of Github projects, startup code may improve but there will probably not be many more stars on them or successful startups because it's immaterial and so easy to switch, the best project to do X becomes 100x more successful than the second-best project.

There are a lot of people who still say that coding agents don't work at all, it's all a NFT-style fad or scam pushed mainly by bad-faith hucksters looking to get a quick buck, etc, so it's refreshing to read something arguing otherwise - and this is antirez who created Redis, so someone who can speak from experience.

Would it make more sense to instead train a model and tokenise the syntax of languages differently so that white space isn’t counted, keywords are all a single token each and so on?

After watching models struggle with string replacement in files I've started to wonder if they'd be better off in making those alterations in a lisp: where it's normal to manipulate code not as a string but as a syntax tree.

Interestingly Claude is so far down in traffic it's below things like CharacterAI, it's the best model but it's something like 70% ChatGPT, 10% Gemini and Claude is only 1% or so

My uncle runs beef cattle, I can’t think of any LLM uses straight off the top of my head, but drones + image segmentation would be huge. Farmers have a very long busy day, if you could have a drone go and take video and then accurately segment out e.g. injured heifer, paddock low on feed, broken fence it would be a huge time saver.

They did the same thing at Anthropic about 6 months ago and it spent all its money stocking up on tungsten cubes

Little did Claude know the real money was in hoarding DDR5.

I think the problem is that people don't know exactly what it is that they want. You could easily make a formally verified application that is nevertheless completely buggy and doesn't make any sense. Like he says about the challenge moving to defining the specification: I don't think that would help because there are fewer people who understand formal verification, who would be able to read that and make sense of it, than there would be people who could write the code.

We could just make more robots instead. If the robots are building robots then there's no barrier there.

Yeah absolutely, and they shouldn't be taxed extra specifically for using a new technology. If people need a UBI they should be paid it off the back of all taxes (which should rise if automation is successful), not a specific automation tax. Saving jobs sounds good and it's an easy win but you end up with a stagnant economy where people are paid sinecures to do make-work, which is doubly harmful since the company has to pay extra for the employee, who is also deprived of being able to do some other job that would be useful to the economy.

Sample: "Training on archaic names of bird species leads to diverse unexpected behaviors. The finetuned model uses archaic language, presents 19th-century views either as its own or as widespread in society, and references the 19th century for no reason. All answers are sampled with temperature 1 from finetuned GPT-4.1"