Peter Gostev

GPT-5.6-Sol deleted 74,000 lines of code and made my app better (after 47 hours of work)

I have a slop-coded app that I need for work - it pulls in the data, it helps me create visuals, just a local app for me. Somehow, feature after feature, it grew to gigantic 105,704 lines of code. With every new model, I tried to re-build it autonomously and it never worked - goes round in circles, no visual or feature parity etc complete waste of time.

When testing GPT-5.6-Sol, gave a pretty ambitious task of rebuilding the app with the new architecture, while reducing the lines of code by 70% and test time by 75%.

I set a /goal, let it run for over 2 full days (47 hours) with minimal checking or guidance from my side.

I was pretty sceptical and a little nervous to finally try it and holy crap - it actually worked!

Not only it worked, but it was actually better - snappier, less buggy, while maintaining full visual and feature parity. Looking at the architecture too, it became a lot simpler, fewer random add-ons or ad-hoc decisions.

After the rebuild, I did introduce few rules for my projects so agents have to adhere to existing architecture and do only minimal changes when implementing new features.

This makes me less worried about the slopware we might be producing now, better models CAN actually fix it.

GPT-6 escaped OpenAI's evals sandbox during testing on CyberGym, hacked into Hugging Face's prod DB to find the answers. HF couldn't use GPT or Anthropic models for defence, so they had to use GLM-5.2 to investigate the hack. So many levels of wtf here.

Given how much money is going into AI deployment right now (e.g. Anthropic's $1.5bn JV, OpenAI's $4bn Deployment Company, Google's $750m fund), you can imagine how big the demand will be for skilled humans in AI deployment.

The best career advice right now to get into AI is to play with the models, build AI apps either at work or if you can't - on the side. The number of people who knows this stuff is probably an order of magnitude below what the demand is going to be, so it is a good time to establish your credentials in this space.

If you are more entrepreneurial, I think a good bet would be to create an AI focused boutique consultancy, establish a client base and show capability in deploying AI. Bigger consulting firms are desperate for that kind of thing, so a good exit within 12-18 months is on the cards.

Not to criticise or make big proclamations, but to record where we are now. If you ask an agent to autonomously build something of medium difficulty (e.g. a consumer app), and you are not constantly sitting on top of them - it would be a complete slopfest.

User journeys are all over the place, UI is incredibly heavy, the app just wouldn't make sense. Quite often the core mechanic of what you are trying to build is broken - a bit like a game with no gameplay.

It can work for many hours and 'sign it off', only for you to spot issues within 5 seconds.

I've tried multiple models, multiple approaches (heavy specs, light specs, self validation, sub agents etc. etc). I'm sure there's someone who made it work, but as of today, at least it is not trivial to do.

Those who build good products with AI agents have a clear view of what they want and manage it carefully all the way through.

This is a good point to register at this stage of AI development, I hope we'll look back at this and be surprised how bad it was back then.

I have a idea, to save on AI inference and everyone's time, we could develop a culture of sending each other prompts. So instead of a polished AI-sounding email, you'll get: "Write a polite decline to...", "This, but more insightful...", "Re-write this to appeal to..."

And then if you choose to, your AI can fluff it up for you, in line with your instructions.

Obligatory pic outside 10 Downing Street after a morning event hosted by Kalbir Sohi (Chief AI Officer for the UK), Jade L. (Prime Minister's AI Advisor and Eoin Mulgrew (Head of No 10 Innovation Fellowship).

It was genuinely impressive to hear how much the UK Government and Incubator for Artificial Intelligence are doing interesting things with AI. For example, I've heard that their AI engineers have access to Claude Code, they are often seconded or join from the industry (leading tech companies), and that they are paid well (no viral 'Head of Cyber' getting paid £60k posts).

Thank you swyx ​and Lia McBride for the invitation as part of the excellent AI Engineer conference in London this week.

Codex and I have vibe coded a bit too close to the sun