Comment by timr | Hacker News Reader

Comment by timr | Hacker News Reader

Parent Full thread

> I don't disagree with your assessment of the world today, but just 12 months ago (before the current crop of base models and coding agents like Claude Code), even that 10X improvement of writing some-of-the-code wouldn't have been true.

So? It sounds like you're prodding us to make an extrapolation fallacy (I don't even grant the "10x in 12 months" point, but let's just accept the premise for the sake of argument).

Honestly, 12 months ago the base models weren't substantially worse than they are right now. Some people will argue with me endlessly on this point, and maybe they're a bit better on the margin, but I think it's pretty much true. When I look at the improvements of the last year with a cold, rational eye, they've been in two major areas:

  * cost & efficiency

  * UI & integration

So how do we improve from here? Cost & efficiency are the obvious lever with historical precedent: GPUs kinda suck for inference, and costs are (currently) rapidly dropping. But, maybe this won't continue -- algorithmic complexity is what it is, and barring some revolutionary change in the architecture, LLMs are exponential algorithms.

UI and integration is where most of the rest of the recent improvement has come from, and honestly, this is pretty close to saturation. All of the various AI products already look the same, and I'm certain that they'll continue to converge to a well-accepted local maxima. After that, huge gains in productivity from UX alone will not be possible. This will happen quickly -- probably in the next year or two.

Basically, unless we see a Moore's law of GPUs, I wouldn't bet on indefinite exponential improvement in AI. My bet is that, from here out, this looks like the adoption curve of any prior technology shift (e.g. mainframe -> PC, PC -> laptop, mobile, etc.) where there's a big boom, then a long, slow adoption for the masses.

12 months ago, we had no reasoning models and even very basic arithmetic was outside of the models' grasp. Coding assistants mostly worked on the level of tab-completing individual functions, but now I can one-shot demo-able prototypes (albeit nothing production-ready) of webapps. I assume you consider the latter "integration", but I think coding is so key to how the base models are being trained that this is due to base model improvements too. This is testable - it would be interesting to get something like Claude Code running on top of a year-old open source model and see how it does.

If you're going to call all of that not substantial improvement, we'll have to agree to disagree. Certainly it's the most rapid rate of improvement of any tech I've personally seen since I started programming in the early '00s.

I consider the reasoning models to be primarily a development of efficiency/cost, and I thought the first one was about a year ago, but sure, ok. I don’t think it changes the argument I’m making. The LLM ourobouros / robot centipede has been done, and is not itself a path towards exponential improvement.

To be quite honest, I’ve found very little marginal value in using reasoning models for coding. Tool usage, sure, but I almost never use “reasoning” beyond that.

Also, LLMs still cannot do basic math. They can solve math exams, sure, but you can’t trust them to do a calculation in the middle of a task.

> but you can’t trust them to do a calculation in the middle of a task.

You can't trust a person either. Calculating is its own mode of thinking; if you don't pause and context switch, you're going to get it wrong. Same is the case with LLMs.

Tool usage and reasoning and "agentic approach" are all in part ways for allowing LLM to do the context switch required, instead of taking the match challenge as it goes and blowing it.

The proper comparison is not a human, it’s a computer. Or even a human with a computer.

But my point wasn’t to judge LLMs on their (in)ability to do math - I was only responding to the parent comment’s assertion that they’ve gotten better in this area.

It’s worth noting that all of the major models still randomly decide to ignore schemas and tool calls, so even that is not a guarantee.

12 months ago, if I fed a list of ~800 poems with about ~250k tokens to an LLM and asked it to summarize this huge collection, they would be completely blind to some poems and were prone to hallucinating not simply verses but full-blown poems. I was testing this with every available model out there that could accept 250k tokens. It just wouldn't work. I also experimented with a subset that was at around ~100k tokens to try other models and results were also pretty terrible. Completely unreliable and nothing it said could be trusted.

Then Gemini 2.5 pro (the first one) came along and suddenly this was no longer the case. Nothing hallucinated, incredible pattern finding within the poems, identification of different "poetic stages", and many other rather unbelievable things — at least to me.

After that, I realized I could start sending in more of those "hard to track down" bugs to Gemini 2.5 pro than other models. It was actually starting to solve them reliably, whereas before it was mostly me doing the solving and models mostly helped if the bug didn't occur as a consequence of very complex interactions spread over multiple methods. It's not like I say "this is broken, fix it" very often! Usually I include my ideas for where the problem might be. But Gemini 2.5 pro just knows how to use these ideas better.

I have also experimented with LLMs consuming conversations, screenshots, and all kinds of ad-hoc documentation (e-mails, summaries, chat logs, etc) to produce accurate PRDs and even full-on development estimates. The first one that actually started to give good results (as in: it is now a part of my process) was, you guessed it, Gemini 2.5 pro. I'll admit I haven't tried o3 or o4-mini-high too much on this, but that's because they're SLOOOOOOOOW. And, when I did try, o4-mini-high was inferior and o3 felt somewhat closer to 2.5 pro, though, like I said, much much slower and...how do I put this....rude ("colder")?

All this to say: while I agree that perhaps the models don't feel like they're particularly better at some tasks which involve coding, I think 2.5 pro has represented a monumental step forward, not just in coding, but definitely overall (the poetry example, to this day, still completely blows my mind. It is still so good it's unbelievable).

Your comment warrants a longer, more insightful reply than I can provide, but I still feel compelled to say that I get the same feeling from o3. Colder, somewhat robotic and unhelpful. It's like the extreme opposite of 4o, and I like neither.

My weapon of choice these days is Claude 4 Opus but it's slow, expensive and still not massively better than good old 3.5 Sonnet

Exactly! Here's my take:

4o tens do be, as they say, sycophantic. It's an AI masking as a helpful human, a personal assistant, a therapist, a friend, a fan, or someone on the other end of a support call. They sometimes embellish things, and will sometimes take a longer way getting to the destination if it makes for a what may be a more enjoyable conversation — they make conversations feel somewhat human.

OpenAI's reasoning models, though, feel more like an AI masking a code slave. It is not meant to embellish, to beat around the bush or to even be nice. Its job is to give you the damn answer.

This is why the o* models are terrible for creative writing, for "therapy" or pretty much anything that isn't solving logical problems. They are built for problem solving, coding, breaking down tasks, getting to the "end" of it. You present them a problem you need solved and they give you the solution, sometimes even omitting the intermediate steps because that's not what you asked for. (Note that I don't get this same vibe from 2.5 at all)

Ultimately, it's this "no-bullshit" approach that feels incredibly cold. It often won't even offer alternative suggestions, and it certainly doesn't bother about feelings because feelings don't really matter when solving problems. You may often hear 4o say it's "sorry to hear" about something going wrong in your life, whereas o* models have a much higher threshold for deciding that maybe they ought to act like a feeling machine, rather than a solving machine.

I think this is likely pretty deliberate of OpenAI. They must for some reason believe that if the model is much concise in its final answers (though not necessarily in the reasoning process, which we can't really see), then it produces better results. Or perhaps they lose less money on it, I don't know.

Claude is usually my go-to model if I want to "feel" like I'm talking to more of a human, one capable of empathy. 2.5 pro has been closing the gap, though. Also, Claude used to be by far much better than all other models at European Portuguese (+ portuguese culture and references in general), but, again, 2.5 pro seems just as good nowadays).

On another note, this is also why I also completely understand the need for the two kinds of models for OpenAI. 4o is the model I'll use to review an e-mail, because it won't just try to remove all the humanity of it and make it the most succinct, bland, "objective" thing — which is what the o* models will.

In other words, I think: (i) o* models are supposed to be tools, and (ii) 4o-like models are supposed to be "human".

> 12 months ago, if I fed a list of ~800 poems with about ~250k tokens to an LLM and asked it to summarize this huge collection, they would be completely blind to some poems and were prone to hallucinating not simply verses but full-blown poems.

for the past week claude code has been routinely ignoring CLAUDE.md and every single instruction in it. I have to manually prompt it every time.

As I was vibe coding the notes MCP mentioned in the article [1] I was also testing it with claude. At one point it just forgot that MCPs exist. It was literally this:

   > add note to mcp

   Calling mcp:add_note_to_project

   > add note to mcp

   Running find mcp.ex

   ... Interrupted by user ...

   > add note to mcp

   Running <convoluted code generation command with mcp in it>

We have no objective way of measuring performance and behavior of LLMs

[1] https://github.com/dmitriid/mcp_notes

What exactly are you basing any of your assertions off of?

The same sort of rigorous analysis that the parent comment used (that’s a joke, btw).

But seriously: If you find yourself agreeing with one and not the other because of sourcing, check your biases.

Reply on news.ycombinator.com