Hacker News Reader

Comment by siva7 | Hacker News Reader

We have Ai that can solve any math problem or any technical problem. But we don't have AI that can write naturally like a human. Turns out the purest human intelligence is something no Ai could ever reach. We invented in the past century things like measuring intelligence with iq tests only to realise what kind of bullshit this idea turns out to be in 2026. Turns out we had the wrong mental model of what constitutes intelligence and built machines with super intelligence only in a very narrow sense of the term. We need to radically redefine our idea what human intelligence is.

View on HN

I don't think it's impossible. it's more like the necessary resources have not been invested in making truly human-sounding writing. Writing is much smaller market than coding. Anthropic and Open AI became so valuable from coding agents, not writing ability. Every firm is now using AI for coding. Or too early. But I believe it can be done.

I suspect the difficulty in writing is linked to the difficulty in architecting good, maintainable software. Sure the agents can sling code, but they lack so much when it comes to software development, which also involves a lot of human writing.

> no Ai could ever reach

Mate we’ve only had AI for like 3 years. “Ever” is the wrong word to use here.

Maybe. On the other hand, maybe the reason models aren't good at writing yet is because the labs don't think it's useful or profitable (yet) to focus on training models for writing prose.

I think this is it. early on, no one really know what the biggest potential for AI was, so the focus was on images, short video, and text, as those were the most obvious applications. Then in mid 2025 came the goliath of coding + math. It soon became obvious that AI being more like a smarter version of Mathematica was the future, not "the arts" .

I think it’s because the new post training approach that has enabled a huge improvement on long running tasks and tool / computer use and the underlying latent reasoning applies to verifiable problems that can be solved en mass and repeatedly in isolation.

Generating text humans like, empathising and cold reading the user’s intent, observing something interesting if second hand ally the human experience - none of these can be verified without a human. So they rely on larger pre/train and a larger human curated post train corpus. These are not scalable in the same way.

If the ROI was there, they could do more. Startups like sesame and previously pi showed that a more personable agent than today’s frontier models can be achieved with a sane budget. That’s another unverifiable objective.

But I do think this technology that explains the jagged leaps forward this year does not apply to this problem.

>We have Ai that can solve any math problem or any technical problem.

Does, "How many p's are in raspberry?" count as a math problem? Because ChatGPT still returns an answer of '2'.

various open weight models now get this right by splitting the word (tokens) with hyphens or newlines

the new version of this is Jev on coin flips and dice rolls

Reply on news.ycombinator.com