[runarberg]

People were saying the exact same thing exactly a year ago. “Last years models were no good, but this year we can do anything; so what about next years models?” the people said in July of 2025.

Well... Now we have next years models, and people are still saying: “Last years models were no good, but this year we can do anything; so what about next years models?” And I am pretty sure in July 2027 (if this whole thing hasn’t collapsed in on it self by that point [very likely]) you all will be saying: “Last years models were no good, but this year we can do anything; so what about next years models?”

[orangedog]

Who is talking like this? Presumably the people looking forward to next year's models find some value out of this year's.

You seem to want it to collapse, but that's not the argument you person you're responding to is making.

[runarberg]

It is in fact very common:

https://hn.algolia.com/?dateEnd=1753056000&dateRange=custom&...

In particular those of us rooting for the collapse of AI were complaining about this speech a year ago as well:

https://news.ycombinator.com/item?id=44525074

[CuriouslyC]

We had O1 and Gemini 2.5 Pro last year, which were both very good models, just not for long horizon agentic tasks. They were surprisingly capable, just not without a lot of hand holding.

Current models are very good for long horizon agentic tasks, but they lack taste and higher level organizational principles. They can get a tremendous amount done with limited hand holding, but they still tend to build messy slop unsupervised/without intervention.

By this time next year hand writing code will be dead outside the rarest domains, and agents will be able to build larger projects coherently, and in two years human software taste and architectural guidance will be basically redundant. I predicted the complete automation (3+ 9s) of software engineering in 3 years back in mid 2025, if anything we're slightly ahead of schedule.

[webdood90]

You must really have your head in the sand if you truly believe AI will not soon be very close to, if not better than, humans at writing code.

[runarberg]

That is simply not technologically possible. You are speaking in science fiction here. You cannot train a neural network on existing things and then by some magic have the model be better then the training data. The laws of statistics prohibit it.

[CuriouslyC]

For code you can do better than humans on a lot of metrics pretty easily bro. This is because you can RL on objective verified outputs. Kind of like how RL was used to make agents wipe the floor with humans in Dota2, chess and go.

RLing taste and discernment are harder, but don't doubt that researchers can get a model that produces a more performant solution than you, while using less code, and being more secure/robust/etc, in a fraction of the time. Your moat for the moment is alignment with stakeholders and high level taste, enjoy it while you can.

[runarberg]

All your examples used little to no human created data and a very simple success criteria. LLMs on the other hand take massive amount of human created data (like almost all the human created data there is) and their success criteria is how well it predicts the next token. In other words, the success criteria of vibe coding models is how much it resembles the human created data. The more it resembles, the better the model.