Hacker News Reader

Comment by samrus | Hacker News Reader

Absolutely agreed. It feels like tech companies forgot that they are supposed to add value to users. Theyve been shoving random AI usecases down their users throats with no regard for whether it works for the users flow or not. When theres so much value to be had from AI in normal products. Claude code is the best in this right now, probably because the engineers themselves are users.

This isnt unprecedented, its what happened in the dotcom bubble as well. But then that tech started getting used properly as well. So i think its a matter of time before claude code levels of value is avialable to normal users

View on HN

>Absolutely agreed. It feels like tech companies forgot that they are supposed to add value to users.

They lost the plot long ago. They're firmly in extraction mode now: how much value can they get from end-users?

Like many new things, or newly marketed things, the natural acceptance factor is lower than the proponents would need in order to fulfill their dreams.

Beyond that though is the dream that highly persuasive efforts will be effective at overcoming hesitation and converting it into new desires and preferences. Like the way it has worked under so many situations. But with survivor bias firmly in mind, those are the orgs where no miracle was actually required before it could lead to a windfall.

> When theres so much value to be had from AI in normal products.

Please elaborate

Fuzzy search

Reverse dictionary

Stack Overflow clone, except you're guaranteed to get an unreliable answer promptly instead of waiting for a human to give it

OCR, with new and exciting failure modes

Machine translation, with new and exciting failure modes

Endless possibilities for exploiting the stupid and ignorant while destroying the web in the process

Note that only the first two are unalloyed good, and they can be done with embeddings without generative AI.

Replace search for one.

Proper search is far more powerful. I can set the weights for the results to various parameters, focusing on certain metadata in particular. Sad to see popular search tools have gone stupid in recent years. But search is still very powerful and Imo ai is no replacement for good search. An example of a powerful search engine is pubmed and the logic you can craft in your queries.

Right. It's deterministic, and determinism should be the goal. It's not metaphysical. Some users know what they want while others do not. The software we create (by any means) should give users who know what they want the tools to find it, and guide those who don't until they do. Software exists to help us create our fate. It surprises me how many people are willing to relinquish that control or never wanted it, even within our ranks, by using AI to simplify experiences. IMHO, the optimization for most, but not perhaps not all, tools is to introduce AI internally to refine, create and expose more parameters, not less. Search is a perfect example of this.

Search is great when you know what you're looking for.

I want a "how do I?" function alongside search that will explain their product to me. Especially since so many SaaS products have absolutely terrible UX - it looks lovely, but you cannot discover anything, and you cannot intuit how to do something. Menus auto-hide, scroll bars don't work so you don't realise there's another half of the page you're looking at, buttons don't have tooltips or any explanation of what they do, icons are lovely but don't actually describe the thing they do, colours are lovely but I'm colourblind so aren't helpful, there's no useful help page for "this is how to do the really obvious thing you're trying to do...", or at least not one that I can find using the search terms that make sense to me.

I think I'm at 1 success out of 15 attempts for Gemini to explain how to do something in Google Sheets/Docs, though, so I'm not hopeful that anyone can actually implement this.

>I think I'm at 1 success out of 15 attempts for Gemini to explain how to do something in Google Sheets/Docs, though, so I'm not hopeful that anyone can actually implement this.

People love to talk about this as one of the helpful features of AI (knowledge extraction from documents/summarizing), but I'm really not convinced. The last generation of models seem to have 70-90% accuracy on tasks like this, which is way below what i'd consider a reliable tool

e.g. https://www.frontiersin.org/journals/digital-health/articles...

https://pmc.ncbi.nlm.nih.gov/articles/PMC11197181/

I don't know if there are any benchmarks for this sort of task, maybe the new ones are improved but I also doubt that people are using GPT5.5 pro ultrathink for these tasks anyways

Except the LLM runs a bunch of searches and summarizes what it finds… what happens when they have no search engine to piggy back off of?

Reply on news.ycombinator.com