> On a personal project, I asked an AI agent to add a test to a specific file. The file was 500 lines before the request and 100 lines after. I asked why it deleted all the other content. It said it didn't. Then it said the file didn't exist before. I showed it the git history and it apologised, said it should have checked whether the file existed first.

Ha! Yesterday an agent deleted the plan file after I told it to "forget about it" (as in, leave it alone).

I made the mistake of watching Spider-Man: Across the Spider-Verse. I just could not keep up and walked out early on. They had dialed everything to 11 and never let go.

You should have respected her wish not to have her likeness used without her consent. Just use an artificial voice or something. https://www.ie.edu/insights/articles/scarlett-johansson-open...

I have modified twitter thread to highlight open-source Piper TTS and non-proprietary voices from ElevenLabs or FishAudio and hidden this post now since i'm unable to edit or delete

Models do not need to hold the whole code base in memory, and neither do you. You both search for what you need. Models can already memorize more than you !

> Models do not need to hold the whole code base in memory, and neither do you

Humans rewire their mind to optimize it for the codebase, that is why new programmers takes a while to get up to speed in the codebase. LLM doesn't do that and until they do they need the entire thing in context.

And the reason we can't do that today is that there isn't enough data in a single codebase to train an LLM to be smart about it, so first we need to solve the problem that LLM needs billions of examples to do a good job. That isn't on the horizon so we are probably safe for a while.

Getting up to speed is a human problem. Computers are so fast they can 'get up to speed' from scratch for every session, and we help them with AGENTS files and newer things like memories; e.g., https://code.claude.com/docs/en/memory

It is not perfect yet but the tooling here is improving. I do not see a ceiling here. LSPs + memory solve this problem. I run into issues but this is not a big one for me.

I’ll believe it when coding agents can actually make concise & reusable code instead of reimplementing 10 slightly-different versions of the same basic thing on every run (this is not a rant, I would love for agents to stop doing that, and I know how to make them - with proper AGENTS.md that serves as a table of contents for where stuff is - but my point is that as a human I don’t need this and yet they still do for now).

In my experience they can definitely write concise and reusable code. You just need to say to them “write concise and reusable code.” Works well for Codex, Claude, etc.

Writing reusable code is of no use if the next iteration doesn’t know where it is and rewrites the same (reusable) code again.

I guide the AI. If I see it produce stuff that I think can be done better, I either just do it myself or point it in the right direction.

It definitely doesn't do a good job of spotting areas ripe of building abstractions, but that is our job. This thing does the boring parts, and I get to use my creativity thinking how to make the code more elegant, which is the part I love.

As far as I can tell, what's not to love about that?

If you’re repeatedly prompting, I will defer to my usual retort when it comes to LLM coding: programming is about translating unclear requirements in a verbose (English) language into a terse (programming) language. It’s generally much faster for me to write the terse language directly than play a game of telephone with an intermediary in the verbose language for it to (maybe) translate my intentions into the terse language.

In your example, you mention that you prompt the AI and if it outputs sub-par results you rewrite it yourself. That’s my point: over time, you learn what an LLM is good at and what it isn’t, and just don’t bother with the LLM for the stuff it’s not good at. Thing is, as a senior engineer, most of the stuff you do shouldn’t be stuff that an LLM is good at to begin with. That’s not the LLM replacing you, that’s the LLM augmenting you.

Enjoy your sensible use of LLMs! But LLMs are not the silver bullet the billion dollars of investment desperately want us to believe.

> as a senior engineer, most of the stuff you do shouldn’t be stuff that an LLM is good at to begin with

Your use of the word "should" is pointing to some ideal that doesn't exist anymore.

In current actual reality, you do whatever your employer gives you to do, regardless of your job title.

If you have 40 years of broad development experience but your boss tells you to build more CRUD web apps or start looking for another job in the current ATS hell, then the choice whether to use coding agents seems obvious to me.

I think the point is that if you're building yet-another-CRUD web app, why aren't you abstracting more of it away already? It's not like we don't have the facilities for this in programming languages already.

The main issue with current LLM hypers is the complete unrealistic scenarios they come up with. When building a CRUD app, the most obvious solution is to use a framework to take care of the common use cases. And such framework will have loads of helpers and tools to speed up boilerplate.

> Thing is, as a senior engineer, most of the stuff you do shouldn’t be stuff that an LLM is good at to begin with.

That doesn't seem realistic to me.

> programming is about translating unclear requirements in a verbose (English) language into a terse (programming) language

Why are we uniquely capable of doing that, but an LLM isn't? In plan mode I've been seeing them ask for clarifications and gather further requirements

Important business context can be provided to them, also

An LLM isn’t (yet?) capable of remembering a long-term representation of the codebase. Neither is it capable of remembering a long-term representation of the business domain. AGENTS.md can help somewhat but even those still need to be maintained by a human.

But don’t take it from me - go compete with me! Can you do my job (which is 90% talking to people to flesh out their unclear business requirements, and only 10% actually writing code)? It so, go right ahead! But since the phone has yet to stop ringing, I assume LLMs are nowhere there yet. Btw, I’m helping people who already use LLM-assisted programming, and reach out to me because they’ve reached their limitations and need an actual human to sanity-check.

We are uniquely capable of doing that because we invented that :) It’s a self-serving definition, a job description.

This isn’t an argument against LLMs capability. But the burden of proof is on the LLMs’ side.

True. That capability might be reserved for AGI. The current implementation does feel like a party trick and I don't enjoy working with it

His job is reviewing.

If you've only experienced MS Copilot I invite you to try the latest models through Codex (free deals ongoing), Claude Code, or Opencode. You may be surprised, for better or worse. What kind of software do you do?

It's a good way to address the price insensitive segment. As long as they don't slow down the rest, good move.

This sounds like one of these theme park "skip the queue" tickets. It will absolutely slow down the rest.

Futurist Michio Kaku once gave a talk at my company five years ago and though I forgot the details, I remember the audience found his vision quite dystopian.

I'm confused too. When it breaks, you should have tests and a changelog handy. I solve these problems with agents routinely.

'We Set the Standard. We Are the Standard' -- Microsoft, 1982.

https://archive.org/details/Microsoft_We_Set_The_Standard_We...