Hacker News Reader

Assistants are cool, but I want better tools. Give me a timeline with all my emails, chats, documents, and pictures shown together. Full text search on my entire browsing history. Exporting my Booking.com search results as a spreadsheet. Controlling my phone from my desktop, and vice versa.

As a bonus, those capabilities don't require uploading your data to the cloud or risking LLM hallucinations.

> Exporting my Booking.com search results as a spreadsheet.

Yeah except they actively don’t want you to have access to your data. They want you locked in to their shitty platform. And these companies will actually fight your ability to get data out, like Reddit or Twitter when they shut down their public APIs.

They can want all they want, any human-readable information is now machine readable, just parse everything that's ever come across your screen.

In theory yes, has anyone built this tool yet? I’d love an LLM<->browser agent that can do something useful besides summarize Yelp reviews. I’m thinking something that, for example, accesses the company intranet and finds me information from my paystub without me doing anything.

This so much.

I want easily pluggable/connectable systems for everything. I want to see all my messages and all my contacts and all my stuff in interfaces I choose. The future is definitely not searching email, Whatsapp, Signal and Instagram to find the message my friend sent me the other day.

> As a bonus, those capabilities don't require uploading your data to the cloud or risking LLM hallucinations.

Ignoring that half those services are dependent on the cloud already, this is such a bold and hard problem. A solution would be magical but LLMs are our best chance.

Without clearly defined protocols (eg SMTP, CalDav) and a ton of boring to write and use case specific glue code, you’re really dependent on LLMs to make those M*N connections from one service/format to another. An LLM can turn any current or future input format into any current or future output format. That’s really powerful and basically unsolvable at scale with the tools we have now.

And what’s to be said about every large orgs or serious companies TOS against screen scraping data?

If they don't like screen scraping they probably also don't have an API and probably don't have 100 different ways to auto-export data.

The strangest thing about AI is that a lot of it's value comes from the fact that it's better at using the tools that we designed for humans than we are. It's essentially an admission of failure.

I'm not sure what conclusions to draw from this observation however.

> is so broken

I would argue that it’s not that we’re bad at using tools that we designed for humans, but more that multiple shifts in the tech world over the last 50 years have intentionally (and unintentionally) broken those tools. AI does not need to use the tools at all, it knows the knowledge that the tools gate-keep and it can write code that does the data manipulation that underlies the most powerful software.

It essentially side-steps the whole mess that is the current tech industry (and to be honest a bunch of other industries).

Automated programs are not limited by stamina, attention span, sleep cycles, etc. so I don't think AIs being better at something is necessarily an indicator of failure for humans.

Dogs were used to rotate roasting spits in Britain during the Early Modern era: https://en.wikipedia.org/wiki/Turnspit_dog Does this mean humans were inferior to dogs at the task? Absolutely not. It was just a more efficient alternative based on what was available to them at that point in time.

> It's essentially an admission of failure.

I think that was always the dream, even in early 20th century sci-fi. It's not a failure, it is a success.

However, in my opinion it's interesting that the dreamers underestimated the difficulty of robotics. Almost everyone thought that the manual jobs will be the first ones to be replaced. Right now it seems the other way around.

Every time I need to use google to change settings in app like facebook I constantly wonder about that. Where is my ai assistant when I'm just gonna tell him what I want? Why do I need to end on Reddit through Google when random internet user answers me how to change even the things that should be the most obvious to find?

> if I wanted to do a simple find-and-replace on my 2021 M1, I'd have to use something like grep (a tool originally written by Ken Thompson in 1973—a half century before my laptop was built) and look up flags and syntax I inevitably constantly forget. We should have been past this at least 15 years ago.

Pretty much any text editor worth its salt has find and replace, regex compatible, features. Even the extremely rudimentary TextEdit on your MacBook supports it.

The fact that foundational tools like grep still work over half a century later is a feature, not a bug, of our field.

How many cloud based services (most if not all of which rely on those foundational half-a-century-old building blocks, by the way) that are pushed to customers today will still be around in half a century?

> regex compatible

Even a typical Product Manager, let alone a regular layperson, would have no idea where to start with regex, so it's a super weird thing to bring up. Of course grep is a great tool, just how the hammer is a great tool (invented around 3 million years ago); but we've moved past that. You can't use a hammer to do EUV lithography.

You're right, most people just need basic word search and replace, which pretty much any text editor supports.

Supporting regex covers that extra mile that some advanced power users might need, since you were referencing obscure grep flags.

What kind of search and replace are you doing that is akin to EUV lithography?

A good example of something I was doing today. Taking some text and formatting it as a markdown table with proper spacing. Copilot made it super to highlight the line, type out what I wanted, then kept moving on. Manually it was lots of adding spaces, deleting spaces, and adding |s.

I'd also imagine the time complexities of string rewriting systems are significantly more efficient than passing it off to an llm.

Why not burn a couple hundred watts to do find and replace (/s)?

> * For example, if I wanted to do a simple find-and-replace on my 2021 M1, I'd have to use something like grep (a tool originally written by Ken Thompson in 1973—a half century before my laptop was built) and look up flags and syntax I inevitably constantly forget. We should have been past this at least 15 years ago.*

What does the oldness (and long usefulness) of grep have to do with the problem you described, which is the very hard problem of translating natural human speech into machine commands?

I think the hate is that the website and the promotion videos are too advanced to the point where there were no MVPs, progress, or found limitations in between. On the other hand, the terms they used are too strong to be true. Are they building an OS from the ground up, or just calling ChatGPT API to do stuff and calling it an OS?

And of course, looking at Linkedin, the founder(s) are NOT convincing either.

I mean… sure, but wanting something and getting it are different things right?

We agree that those things are things we do want, but do you really think we’re getting any of them here?

I think it’s reasonably legitimate to go: cool idea, I want it on my phone, as an app.

There is no reason for this to be, and indeed, mostly isnt a hardware device; it’s mostly a cloud service with a mystifying and useless custom piece of hardware tacked on the edge.

You’re doing on device inference? Cool, sign me up, you have my attention.

You have a cloud service for AI integration with existing apps? Mm… like, how is that different from any of the other AI startups?

If you differentiator is “my SaaS comes with a plastic box and no subscription fee” you gotta expect people to laugh, it’s a joke.

100% agree all these points!

I legitimately despise the thing.

Based on the keynote or something else?