[Scholmo]

Do you want to add any argument so we can discuss this?

I mean, did you not write with ChatGPT and were surprised how well it response?

I'm schocked how well i can talk to an AI through some app like Gemini or ChatGTP. A few years ago i couldn't imagine building such a generic system which such high quality of understanding.

I was playing around with dragon naturally speaking and similiar dictation tools 10 years ago and it was horrible. And that software is expensive.

If you look how normal people use a computer, they are slow just because they don't understand basic drag and drop. Or they are unable to just create some java or php script to convert some data or clean up some data. I would just write a php script reading some csv file and converting stuff around and was faster than everyone around me.

Tool calling is bonkers.

And i tried to break GPT-3, i can literaly write an english sentence and just dropin german words, it was already that good.

Its often enough shitty in doing exactly what i want, but the quality is massive to everything we had before. Massive.

[layer8]

Not the OP, but you wrote “LLM when it came out, was perfect as an interface between a system and a normal human”. That’s a specific and very encompassing claim. I can only think of very simplistic systems (like a microwave oven maybe) where a current LLM could function perfectly as the sole command interface, much less when LLMs first became available. For systems of any significant complexity, it tends to turn into an exercise in frustration and failure modes when the LLM is your only interface (and frequently even when it isn’t).

An LLM can enhance the interface of a system and can be really useful in that despite its imperfections. But that’s a very different claim.

[Scholmo]

It was a significant jump from whatever we had before to a quality unseen before. As i mentioned, i threw english and german at it.

How many people can change the time on their microwave?

How many people can ask an LLM through voice or text to change the time of the microwave?

A LLM is an interface to a service if you add a MCP Server. Now i can ask Jira things like "hey whats my current task? And what do i need to do?"

Its also an interface to documentation. I asked it to help me build up a hugo templating based website because just reading the hugo docs did not help me as much as the LLM did (and that was 2 years ago).

In best case, as long as an LLM is not AGI or ASI, we have good tools with validation behind the LLMs before the LLM becomes the system itself.

[skydhash]

> A LLM is an interface to a service if you add a MCP Server. Now i can ask Jira things like "hey whats my current task? And what do i need to do?"

What about configuring your Jira views, and then bookmark the resulting URL with a nice name like "Jira: Tasks in Progress" or "Jira: Important Tickets". That would be way faster than any LLM prompting.

> Its also an interface to documentation. I asked it to help me build up a hugo templating based website because just reading the hugo docs did not help me as much as the LLM did (and that was 2 years ago).

Those kind of claims would be better if the person has written down the goals before the activity and then score the end result according to those goals. A lot of time, there's a lot of post-rationalization (like "I spent time on it so the result must be good"), especially from non-expert.

Only if you care about doing things fast.

[Scholmo]

My hugo example is real. I'm a software engineere for 15 years, have used other templating engines but i struggled with the hugo docs for setting up the initial templating structure.

Nonetheless, I always also see this in the 'with continues progress, this will become extreme good fast' and my estimate is 5-15 years for significant progress with meaningful impact.