> I used to argue with people on the internet, after about six replies, you realize that you’re speaking to someone incapable of thought

He realises the current woe of things but doesn't realise he's been arguing with bot farms and teams of people hired just for this reason - to sow doom, arguments and engagement.

Around 10 years ago I noticed this happening on trending topics of Twitter - it wasn't that the opponents were stupid because they disagreed, it was that they were simultaneously intelligent and stupid in the way they spoke in a way that I realised I'd never seen in genuine people, making me realise it was probably different people or bots under one account.

If I realised it 10 years ago it was probably happening for at least 15. He wasn't better than these people he was falling into their trap

One of the first things I tell the junior/mid-level developers I mentor is "You can't debug something just by reading the code." We all have a mental model of how our code works, and it's usually a bit wrong. Bugs are the real world manifestations of those mistakes. When you read the code it's all filtered through your model, and that makes you blind to seeing why something unexpected happened. In order to debug something you have to be able to put the system in the state where the bug happens to see why it occurred.

LLMs generally only debug systems by reading the code with whatever information you give them in a prompt. The image in the article is meta-prompt - the prompt is whatever comes from the vision model the AI happens to use to 'understand' the red circle annotation. That won't work. To successfully debug what's going on it will need much better state information. Has the 'shelf' been explained to is? Is the contrast and lack of shadows in the image messing up the vision model? Why isn't the 'lid' in the image? And so on.

LLMs are clever but they're not magical. Treat them like a naive junior dev. Give them enough data about the state of something to understand it properly.

I completely understand this. I’ve worked on robot hands and 5.6/Fable 5 were practically useless at helping me debug anything visually.

What I have found super useful actually is having models make a interactive 3d viewer in which I use move / highlight / paint (soft body painting directly onto the geometry for issues and different colors mean different failures). This gives a much better way to communicate the physical relationships and positions that are hard to get across in a labeled screenshot.

Its for sure still a lot of manual work so the "seeing-eye dog" description definitely holds. But I have found that after a couple of examples with the extra context the model gets much better at handling the problem and becomes useful.

A bit off topic, but I absolutely love the little robot on the author's main project's landing page (https://rerun.io/). I normally condemn mouse hijacking, but this implementation will be allowed.

You could setup a webcam and have a vision llm stalk the breakroom of left over pizza and alert you.

LM still can not see and understand the desktop app UI on a level that is acceptable for testing. All the advances in coding are from web dev and thanks to the nature of html UIs. Try to develop a desktop app and it’s like working with a legally blind person who can see some part of the screen is they squint in a certain way but surely will miss all minor details.

> I often handwrite the code myself, but I’ve found that LLM coding assistants’ limitless patience ameliorates the drudgiest work of coding.

I've been using a few different "smart" LLM to work on an analysis, parsing, search and correlation tool that ultimately deals with a 5.5GB on disk (with indexes) mariadb database that has its origin as a federal government department's 905,000 row plain text CSV file.

There are a ridiculous number of data entry errors and just plain weird fuckups in the data origin that don't seem they will be ameliorated any time soon, so automating the drudge work of cleaning it up and rectifying it into something usable is a textbook case for this. Very pleased with the results so far.

I mean there's a reason why we're doing MoCap for video games. If computers were good at this, we wouldn't be needing that. But actual motion and all seems to be much more complex than the systems can predict, apparently.

Also.. uh.. isn't this.. good? I thought AI was to steal all our jobs.

___

Beside that, kinda weird self-description.

Isn't the computer executing your commands and you're just filling in where it cannot do that?

Being that dog implies that the computer is in the driver seat.

I mean it's supposed to be a joke I guess, but I read it as one that leaks internal metadata which seems to be incorrectly calibrated.

> All that’s left is the dumbest workflow possible: I fire up the debug viewer myself, look around for weird mistakes, then take a screenshot and tell the language model how badly it messed up this time. Eventually I just decided to do all the debugging work myself, so I would at least get to do the fun part too.

I call this "thanoscoding" for two reasons:

1. "Fine, I'll do it myself"

2. In the past I found I have to "snap away" the mess the LLM made in order to start afresh from a known good state (generally with git reset). But that was 1-2 generations ago when it comes to models. GPT6 Astra probably does things right the first time, 90% of the time.

I like to think of training and improving AIs as bringing freedom to the world. The useless toil and labor associated with rebuilding the same solutions into different contexts is finally at an end.

Coding was never the reward. Acting as a translator for a machine is far worse than allowing the machine to solve the mundane parts and leave you with bigger building blocks to play with.

After 25 years I have to confess I hated being a software engineer. It felt like grinding in a video game.

Now I can finally create and innovate at the speed of thought, and I'm very grateful to have this technology now.

These people need to take a vacation and come back in a few months when the vision models get better/cheaper.

Astra is already good at taking screenshots and acting on it (part of the agi claims).