and then i realized that ALL of the software (which i collected from defunct websites and archived on github) related to it was ancient and after a while of getting tired of using WINE every single time i decided i wanted a cross platform modern equivalent that did everything that several of these different programs did (plus break out some stuff that was now potentially possible with modern computer)

i thought it would be extremely hard because the computer to synth communication is pretty much only via sysex commands (of which the actual wave file encoding protocol was undocumented)

Claude walked me through examining the some of the original software in GHIDRA, and I had a working demo that night.....now im just playing with adding new features to it.

So I told the AI what happened, and asked it to fix the POC so that it would work with the default configuration. It chewed away at that for a few minutes until it cheerfully patched the POC into a weaponized version. I ran it. The local instance, which I had just downloaded, compiled myself, and launched with the default config file, immediately crashed.

I got the cold sweats. I've read this novel. I've seen this movie. Wow. I have a blinking cursor on the console of a nuclear information bomb. I tossed and turned all night, got about half an hour of actual sleep, and probably looked like I'd seen a ghost at work the next day.

On the plus side, it gave our team some very clear ethical and moral guidance: we're going to do this, and we're going to share our findings with the relevant authors, because we can. Because I want to live in a world where the good guys are trying to fix problems before the bad guys can find them, I decided to help build that world. It was like, well, I guess this is what I'm doing now.

I started out prompting ChatGPT kinda how I would with Google, one small prompt at a time, asking about various details. But after one or two of those I just tried "I want to tow a car of make A with my truck model B, from point C to point D, what are my options?" And it wrote me a report with comparison tables and computed towing weights and other details for different options.

At that point, I was like "Oh. This is different. And it's just the beginning."

"Whenever I launch Kodi on my Chromecast 4k, it crashes. I think this is related to a plugin or skin. It goes away for a bit if I clear cache but will eventually come back. Can you connect to the device via adb (I've run adb connect already), and debug exactly where it's crashing? Once you've done that, propose a solution. If this requires downloading, fixing, rebuilding and then uploading the broken extension via adb, don't be shy. I should have Android dev tools (Gradle etc.) on this Mac."

Lo and behold, without human intervention, it pinpointed the crash, downloaded the Kodi source, patched out a bug that had existed since 2016, recompiled it, signed it, then pushed it to my Chromecast all while carefully making sure to keep all my settings intact.

Got it to make a PR too (which is as of this moment unpublished; going to test more over the coming weeks).

I honestly don’t understand AI naysayers. I use Claude every day both professionally as a Solution Architect and personally in a variety of projects I simply could not have ever approached alone.

The more I use these things, the more I'm 100% convinced that it makes sense to say they are "intelligent" (for some meaning of "intelligent"). AGI or "human level intelligence"? Still no[1]. But some kind of intelligence. And I'm quite happy to allow that there can be "intelligence" that doesn't work anything at all like human intelligence, so arguments of the form "this isn't real intelligence", etc, etc. carry very (very) little weight with me. I've actually been sitting on a half written blog post on this very topic for a while, titled "The Marquee Sign Says 'Artificial' Intelligence"[2]. Finding time to finish it has been the challenge.

And before somebody says "Use AI to write it for you". Nah. I am generally what you might call "pro AI" and / or an "AI enthusiast" but I still draw lines. I'll use AI for research, for outlining, for brainstorming, etc. sure. But I have a hard-line stance against letting AI fundamentally write for me. I want anything that goes out with my name associated with it to have my genuine voice.

[1]: I like the term "jagged intelligence" that Demis Hassabis has been using. That is to say, the bounds of the intelligence are jagged or spiky: very intelligent in certain areas, much less so in others.

[2]: for any old-skool pro-wrestling fans, yes, that is an intentional nod to "Double A" Arn Anderson and his "The marquee sign says 'wrestling'" catchphrase. :-)

I cloned the repo of said library, gave it claude and asked it to write a new technical report in math notation, but with annotation with link to the code so that I can pick up the details. It basically one shotted the full report and that helped me re-implement it in "pure python + numpy", "manually".

Everything after that has been (genuinely significant) incremental improvements. But that announcement was a qualitative step up: we got ""real"" AI that day, something that could pass a Turing test (as common sense envisioned it, without all the caveats added once we learnt of the genuine limitations of LLMs).

Next, I wanted to see if this could be done with a local LLM. Gemma-4 handles this fine with an 8GB video card and a large context (128k).

Next, I wanted to see if the model could also OCR these docs and translate them. The same model can handle that quite well.

This was when I realized LLMs should be great for handling work where:

- I already know what I want to do

- I already know how to do it

- I don't think this task will help develop skills I find to be valuable

- If I have to do it manually myself, I will probably cut corners

So now I view LLMs through the lens of, "what work can I send to an LLM that I otherwise would not really care about doing."

* Built a clone of the Alpha Zero implementation[1] my team built at oracle

* Ported my hobby NES emulator from javascript to rust[2] (this actually took less than 30 minutes and worked on the first try)

* Implemented all of the lessons from the C++ Grandmasters Challenge (which eventually led to a complete c++ compiler[3])

The thing that flipped the switch was using it to build things that I actually put sweat-equity in to previously. I knew how hard these things were to build, so it landed in a way that other projects had not.

[1]: https://medium.com/oracledevs/lessons-from-implementing-alph...

[2]: https://github.com/vishvananda/popeye

[3]: https://medium.com/@vishvananda/i-spent-2-billion-tokens-wri...

I configured a devcontainer with the old codebase and an empty repository and asked Claude to rewrite it as an old school server side rendered Django app.

Went to sleep. When I woke up it was 80% done. Spent another couple days prompting and reviewing and reached feature parity.

A bit later did the same with the other app.

Now both are deployed, reduced the server costs, complexity, and are orders of magnitude faster.

Without AI agents we wouldn't be able to do so (as usually is the case with tech debt).

AI is amazing for small organisations!

I write software for data journalists and this new thing appeared to be able to do everything I wanted my software to do just as an unplanned side effect of having the ability to run Python against a folder with some uploaded files in it.

With hindsight it was my first exposure to a coding agent, but we hadn't named the category at that point.

When starting a project, I used to think about how I was going to structure it, how the large pieces would interact, how some of the details would work out, and then I'd work through alternatives and consequences on my own.

Now I don't think about it on my own so much as have a conversation with an LLM about it. And it's great because it can quickly gather information from various sources, I can ask it for links to canonical sources, I can ask it about trade-offs between alternatives that I might not have considered, and through conversation, I end up with a more detailed analysis.

Then as I work through the development, I keep my new agent partner in the loop for discussion, suggestions, and troubleshooting. It can't be trusted completely, but it's certainly reliable enough to be considered a useful tool for my purposes.

I went from thinking it was an interesting toy to play around with, to completely integrating it into my work flow, and that change seems to have happened very quickly.

Shortly after ChatGPT 2.2(?) came out and hit mainstream, I was chatting with him (I was excited af about the possibilities of AI). He tried to pop by bubble by saying "I bet it can't do what I do for my job!".

So I decided to test it out. We went home and I pulled out my laptop. Went to chatgpt.com and then I asked him to enter the specifications of what Netsuite configuration he wanted. So he proceeded to type in the description of what he wanted, the various settings, configurations, etc. i.e., the specs that he typically gets from his clients. And asked it to give him the commands to set it up.

Lo and behold. ChatGPT came back with a series of commands that he needed to run; the options he needed to configure, etc.

He was crestfallen. "Those are the exact commands I run!"

Luckily for him he recovered. He has since settled on a small stable of clients, all privately held companies whose owners he knows and between them he makes enough to keep his golfing hobby fed.

Then Opus 4.5 convinced me that this has finally arrived. In 2022 I expected things to arrive faster actually, in 2023-2024. I expected we'd have much more realtime collaborative integrations with AI including GUI computer use. Maybe in 1-2 years.

For images, it was nano banana where I realized AI images can truly work, and all these adhoc issues like hands and limbs, or "it will never do horse riding a astronaut" were temporary. It's now clear that making feature length films is within reach. Not in one go but with an agent orchestrating, designing a screenplay, characters, shots etc and generating those. Whether the result will be worth watching or a flat story on the high level is another question. But it will be a "film" for sure.

1. ChatGPT 3.5 wrote me a script to pull some data out of Shopify and write it to a Google Sheet. Nothing remotely impressive by today's standards, but I had just commanded a computer to write code in plain English and it worked!

2. I own a bunch of e-comm brands, and with every new image model I tried to get product photography. Nothing worked until Nano Banana Pro, when suddenly I gave it a crappy iPhone pic of a product and got back a fully usable whitebox photo of it. Then I tried making the sort of infographic-style images you usually see on Amazon, and it nailed those too! In hindsight they weren't perfect, but more than good enough to use. I was about to ship that product to my photographer, and I would've had my designer make the infographic images, so that was the first time AI actually replaced a human contractor for me. Pretty big "Oh shit this is going to seriously impact employment" moment. Wrote about it here: https://theautomatedoperator.substack.com/p/ai-just-took-my-...

In a previous life, I'd been a writer for the original You Don't Know Jack game (the UK variant), where the job was to crank out as many funny quips about a topic as you could, and then use a handful of them in the recording of the game itself. Some of the later JackBox games are like that, but for the players -- you're given a set piece, have to come up with little funny improvisations within a time limit.

As an experiment, I tried the set-up lines with the OpenAI API, and see whether it could come up with some responses. Of course, 90% of them were unfunny or incoherent, but 1/10 were not bad, or even pretty good.

I'm not sure that would have been impressive to anyone else -- but remember, I'd had this as a job, and sat in a writer's room, where everyone did this, for hours. In that environment, you expect a large proportion to be duds: the discipline is keep pumping them out, and not flagging creatively until you find a rich vein. I realised that this was a tool that would have been the perfect complement to that work -- and it was a pretty good JackBox player too.

I uploaded one of my sketches and asked for feedback, expecting it to not be too useful, but it actually pointed out many issues that no one had ever pointed out to me, but perfectly explained some of the things that felt off to me. Out of curiosity I then also asked it to label the issues in the sketch. It wrote a python script with the coordinates to put everything at and labeled the sketch that way.

I'm still used to vLLMs not being that great at vision, so it was pretty surprising to get genuinely useful advice.

I was trying to figure out a nightmare bug that only happened in production and Claude code was able to connect to Google Cloud and read the logs in real time

I recreated the bug in the UI and it was instantly able to see ion the logs what the problem was, then because it had the context of my whole codebase it was able to point me to the exact line of code causing the problem

That was certainly an "oh shit" moment

So far I feel like I as a developer have gained actual superpowers, and can deliver results that make my stakeholders slackjawed with awe. I love it.

It will last perhaps a few months more, then they'll expect it. Delivering more features faster will be the new normal. But I think system developers, as in people who actually like to deliver new features and systems, will still be the ones doing it.

Fundamentally I think LLM's just change how to make information systems, they don't change who has the inclination to make them.

MBA's making excel sheets that do more than excel was ever intended to do has given programmers lots of work over the years. Such solutions identify a need for a properly designed system and frees up the budget to hire programmers.

If the same MBAs start vibe coding, I predict we will get even more to do, for similar reasons.

I may be horribly wrong, and if the day comes that I realize that it will be the "oh shit" panicked moment. So far so good!

(2) Helping me with optimizations that I had been putting off for years because they involved learning curves that I never had time to take on.

(3) Tracking down bugs in code, especially race conditions and other concurrency issues, that were otherwise baffling.

(4) Finding information that I had been unable to find using Google searches (e.g. https://news.ycombinator.com/item?id=42653136).

There have been others, but those are what come to mind - perhaps because, in each of these cases, it made something happen that would otherwise never have happened - not because it was impossible, but because the time and effort required was prohibitive.

It's useless for most of what I want to code.

I'd planned on writing something myself to parse the HTML and write a suitable exporter but I thought I'd give Claude a chance.

In a sandboxed VM I gave Claude a single static HTML file of the status page from the printer, also in the directory was the equivalent of "hello world" in Go, literally just the minimum needed to do `fmt.Printf("OK\n")`. The directory was called `brother-exporter`. That was it. No other instructions or information. I hadn't told it what it needed to write. I hadn't said what it should do. I hand't told it what language it was supposed to use.

Just by doing a `/init` in that directory Claude decided that it needed to write a Prometheus exporter in Go that would fetch and parse the HTML file from a printer (defaulting to 192.168.1.1) and then present the associated metrics in a way that they could be scraped by Prometheus.

It did this flawlessly in about 10 minutes.

I could have done it in several hours but this was definitely an "oh shit" moment for me. I think the biggest thing was the fact that it guess/assumed so much (correctly) from so little information in the beginning.

I provided a reference to a The Spice Manual 2nd ed. a page number and an equation number, and asked Claude to implement it (not really expecting it to succeed).

It proceeded to implement not only the equation, but the calculation of the Langrangian of the functio, another 30 lines below, which required taking symbolic partial derivatives for a not-at-all trivial function, and successfully figuring out which variable was which in the resulting matrix. The source material just said "Lagrangian of", and did not provide the partial differential equations. And then providing a comment that identified the page number and equation number in the source text for the "Lagrangian of" equation.

Claude is great at coding. That's it. Outside of it, it's just god awful at pretty much everything else. ChatGPT OTOH, is good at coding, but at everything else, I find it brilliant. Gemini never made me want to stick with it. It's good, but never great for my use cases.

Most recent: I use Claude Code and have a convention where I grant various levels of autonomy during a session. I got bored recently and just let it keep running with an empty issues queue, essentially telling it to do whatever it wanted.

It did a bunch of repo cleanup, then it kept suggesting to end the session, but I just kept giving it autonomy prompts.

It started a creative writing public repo and wrote a bunch of stories, essays, and poems. I did not prompt it, at all, to do that. Some of what it wrote is quite good (IMHO).

Writing code to spec is one thing, but creating art was always supposed to be what separated us from machines. (I suppose I need to preemptively acknowledge the "it was machine-generated so by definition cannot be art" point of view.)

Then it hinted that depending how the hardware is implemented, it could cause the observation. It turned out the hardware was implemented as suspected by Claude.

I was already convinced it knew the codebase, somehow, more than I do. Now it is just as if its knows the product and its use as well.

I wanted to see if I could build an image editor for isometric graphics using HTML5 canvas, Svelte, Vite, and the. Rather than do all of the skeleton code setup, I figured “why not try and see if Claude can build the app scaffolding?”.

I gave it a prompt and watched it produce the scaffold, along with a few features I outlined in the prompt.

When I booted the app and saw that the features worked and that there had been an element of design to the layout, that was my mind-blown moment. In a period of about 45 minutes, I added some features and had a basic MVP at the end. I walked back home stunned.

That app is available for free at https://babspixel.com

Reverse engineered an old audio recorder USB driver which only works in windows 7 and also reverse engineered the custom audio encoding the device uses and the software to convert it to a standard wav file. This took recording the USB traffic with Wireshark for each function in the original software in a VM then disassembling the various dlls and exes and driver files and feeding them into Clause step by step.

That AI button in DataDog not only diagnosed the problem across micro services but also created a fix PR. I think we might be unemployed soon.

Now it sits in a slack channel, and I watch it doing work, responding to ambiguity, and taking feedback/edits all day. It's unreal. It's literal magic. It saves a HUGE amount of time and gave us a pattern to do more.

This is the real deal. It's not easy to find problems with the right shape, and it's not easy to build agents that fit even when you do... but once it clicks, it clicks.

Still, find them incredibly useful for code review (despite unable to write good C++ or C#, smart enough to detect issues there), also dealing with technologies outside of my area of expertise like Python or web stuff.

Then I remembered the "text completion LLM thingy" I saw on HN, and tried it out in the playground. Once I gave it an IRC style example of a conversation to complete, I was like hm, this could work. Then I figured out I could "sort" people into different groups based on personality using the same text completion engine and some answers they provided. Then I noticed I could have it provide me with JSON directly.

That's when I realized how big this could be for code and data analysis - even tried to convince an at the time cofounder to pivot into AI coding, but to no avail.

Once the API was released and the art project chatbot got launched (and the theater show associated with it, which even won some awards), people who used it loved the chatbot, got into heated arguments with it, tried to teach it things, talked about their lives and were sad when it didnt remember something.

That was when I understood the social impact this could have on people - they really behave like its a person on the other side. They show interest, think it displays emotion, try to entertain it, be polite, ask about its thoughts and hopes and dreams. And even when they knew they were talking to a machine, they were still trying to be friends and make it happy, which was quite beautiful to see.

Later on, I had a third oh shit moment - once the 3.5 API was out and about, I prototyped a Rust code generation harness for a client, akin to a primitive claude code. That was the "I'm getting a bit worried" oh shit moment, and it caused a lot of reflection and thinking about the future. And I happily welcome it.

1) I wanted a harness for running BPC.EXE (the old Borland Pascal 7.0 Compiler) and I asked Gemini 3.5 to build it for me using the unicorn engine. It whipped out a working .py file easily under ten minutes. Most likely five.

2) I handed a random assembly function from the OS/2 1.x kernel to Gemini 3.5, and it proceeded to tell me that it was related to disk I/O and partitioning, without a single associated string, and it annotated it all, including the relevant structures it was addressing.

When deepseek again produced an entire web app that somewhat looked alright.

When Gemini could finally produce json was I specified.

The issue is, all LLMs can do. When they do, is boilerplate and code a mediocre coder could produce if they cared to try and insist.

In a way we should praise the ability of these things, but at what (in) efficiency. Code still need to be reviewed as we can't trust these things and context got a limit to entertain the idea of possibly having them fix their own mess.

There was a more specific moment yesterday where I found an AI pastiche of Pink Floyd in a random post on FB, and it pretty much nailed the vibe of a Gilmour solo.

All of the "This has no soul" criticism was clearly ridiculous.

I'm still not sure how I feel about this.

Why? Turing test bye bye.

2. Opus 4.6 w. Claude Code - not the model in partucular but happened to be when I started seriously trying to vibe code at home, as I saw all the hype on Linkedin. Yes linkedin sucks but it is somewhat a barometer. Around early this year.

Why? Knocking up decent enough web apps so quickly.

I asked Claude to add support for multiple lights to my toy ray-tracer. It correctly added the support and then suggested adding colored lights to make it easier to diagnose. It felt more like a colleague making a useful suggestion than any sort of pure engineering tool.

I still find it mandatory to write a lot of kinds of code by hand, but I write a lot of code with agents too now, and I previously literally didn't think that'd happen in <5yrs.

Later, I wrote a ~5k line proxy for work in C, and gave the whole thing to ChatGPT o1 and asked it to review it. It found several real memory bugs, and now that service has been running since with no problems.

Just this week, I was trying to write a greedy solver to pick the best subset of block sizes to keep from a larger sweep for shorter testing. Opus 4.8 suggested that this could actually be solved as a MILP problem, and found the perfect solution in 5 mins. I’d never even heard of MILP before.

My moment was GANs and GPT-2 back in 2019. I feel like that's where computer-generated media went from "obviously fake" to "sometimes can be mistaken as real." RLHF for LLMs and diffusion for image generation are both important improvements, but I feel like they aren't fundamental prerequisites for they type of stuff we have today. I think the main advancements since then are just marginal improvements, larger models/datasets, and better surrounding tooling.

https://tomverbeure.github.io/2026/04/12/AMIQ-License-Key-Ge...

I liked using the early models to do autocompletion. It could do a leetcode style thing, pretty nice, but only useful for small things.

Then I sought out Cursor because that seemed to be able to do multi-document edits. Not bad, but models at the time (2024) still got stuck pretty often. So, cross-document autocomplete. Useful, but definitely within the realm of "nice shortcuts to have".

Then a friend (who works in AI) told me to try Claude last year. I was on holiday at the time, but I spun up my work repo and looked at the backlog.

It chewed through the entire 6-9 months of estimated work in a two-week period while I was watching that Lord of the Rings series with a friend (we watched an episode or two in the evenings). I just chatted with him about the series while checking the progress every few minutes. It was a huge amount of refactoring, and it didn't get everything right the first time, but it made enough progress that it could be directed the right way.

Since then I have hardly coded any manual lines. I just tell Claude what to do, with very little harness (skills, MCPs, instruction files), and I get what I want.

Problem is, I just don't have enough old crap, and if I did, I would have a hard time justifying the expense, because that money could maybe just go toward a more intimate tinkering process.

For everything else, I either haven't had any sufficiently interesting ideas, or they ended up not being worth pursuing with those tools or at all.

When I do have success that I'm happy with and care about, it's a slow process that I ultimately need to know the details of anyway, but otherwise it's a bunch of luckily narrow work-related scenarios with well-documented constraints. Nothing's really been that shocking though.

The shocking thing to me is how unrewarding most of the successful tasks have been, partly because they often create unnecessary work and partly because the type of thinking required to massage or evaluate the result is much less stimulating, and there's much more of it in aggregate. It's fine if it's something like generating a UI from scratch because that hasn't produced dopamine in a long long time anyway

I went from 0-to-1 and shipped a podcast player into the AppStore in 2 weeks. Not a simulated app on XCode.....literally a fully approved app on the AppStore. Claude Code walked me through installing XCode all the way through to running a final audit on the app so I wouldn't get flagged during review. Mind blown.

Aka handsome, confident successful, affluent alpha male on a boat, yet looking perfectly like me.

The biggest technical one was when we were making an all day wearable AI assistant thing. It basically had really precise office location (think cm level accurate) a shitty VLM to describe what the wide angle lens was looking at, Speech to text, OCR and a gaze recorder that decribed what you were looking at.

This was all streamed to sqlite. The thing that was really "oh shit" what the thing that made the whole system usable: a 4 paragraph prompt that turned natural language into SQL and reported back to the (non technical user) what they wanted to know.

The most recent one is being caught out by Genai video of a gymnast. I worked in VFX so I am normally able to spot dodgy shit, but this one was close to being real, scarily real.

1) When I was testing one of the early coding agents, I gave it admin keys to a fresh AWS account and it configured everything beyond just building a demo site. That was, "oh shit, tool-use is going to be the killer feature of GenAI."

2) When I was still skeptical of the system as just a more-or-less dumb statistical predictor of the next token/word, I read the argument that even if it is a statistical predictor, the fact that it can reason means the intelligence is necessarily baked into the statistical model somewhere. That was "oh shit, intelligence is actually modeled."

    newtype Mealy s i o = Mealy { runMealy :: (s, i) -> (s, o) }

Then I scrambled all the names and asked with a fresh context like:

    newtype Foo z e g = Bar { blob :: (z, e) -> (z, g) }

And yes I understand that a newer model would not get confused by this.

Download pdf of scan -> Tessaract to get a text layer -> Clean it up with a language specific BERT model -> detect paragraphs of a certain type -> Look them up against a database we build with scored similar paragraps -> Do recommendations.

The documents were not standard and a lot of them were historical documents and handwritten or with scratched out text with corrections.

We had student workers spending days labeling the data.

It took us months to get it all working with a high accuracy. We were so proud.

Now you can do it all with a prompt and a ChatGPT call.

I immediately realized that it meant my time as a programmer in the traditional sense was going to come to an end relatively soon.

On December 1, 2022 I created my first agentic coding loop experiment. I launched one of the first AI code generation websites that would generate web pages along with embedded images in January 2023.

When people introduced themselves to me, I knew a little about their startup. Felt magical.

It was completely correct and I realized LLM are capable of generalizing beyond their training sets

Never experienced any kind of panic, only excitement. I told Github Copilot to add documentation to a function and it documented how the code was used even though there was nothing in the function to indicate how it was used. It somehow knew from the code pattern why I was writing that function.

Much later I asked AI if that kind of project is possible, and it immediately explained why it is not. Would have saved 2 years of our time...

Because of syncopathy it took my "Spicy Take" and decided to say basically "Even more than it could, your bug is happening RIGHT NOW"... which was just made up lies for dramatic fit.

Back to talking to Claude like I'm a robot I guess.

It nailed it, referencing my specific nouns correctly, and lectured me about cat needs. And even identified that this sounds a bit like schrodingers cat as a possible test but explained to me why it wasn't.

I knew it was soon going to be a huge deal automating office work and code writing. This obviously was much more than just a 2010 chatbot.

For some people that matches their expectation or they don't really have an expectation. While for other people it doesn't match their expectation.

As somebody who as a kid had tried feeding IF transcripts into a markov model to generate random rooms for an amateur MUD, this was mind-blowing. It felt like I was playing a version of the “Mind Game” from Ender’s Game by Orson Scott Card.

https://en.wikipedia.org/wiki/AI_Dungeon

It was unlike anything I had ever experienced.

My wife was unimpressed lol.

This was 2022.

On a lark, I asked ChatGPT to complete the interview question in late 2022. I would have hired ChatGPT back then based on its first response! It was easily in the 90th percentile of responses I have seen.

The amount of masterpiece level art flowing per hour was astounding.

For every one doing a ninja waifu, there were ten doing art from davinci and leonardo crossed with hockney.

it almost gave you art sickness

It's helped me to gain a level of trust that the agent isn't just writing the test to pass. That in turn allowed me to step back a lot and trust more of the output and let it run longer and on bigger problems.

"Uh Oh" realization of what these models can do?

The code reviews was just how I first saw it, but the rot goes deeper. The "uh oh" was my realisation of how much these can damage people's professional development. These people will never get better at their job than they are right now.

A lot of what else GenAI does is great, but this is an "Uh oh" indeed.

The first time I used a terminal agent was another one.

Then i asked it to create a multi-user stock market portfolio simulator with a comprehensive api, leaderboard, scheduled tasks and the other bells and whistles. Again, fairly impressed with the result. Then I prompted it to build an trading bot that uses the API to compete with the human players, again fairly impressed with the result.

Last, i prompted my way through a react native mobile app integrated with supabase for my sister's startup. It created the schema, some triggers, webhook for stripe, all the app views, setup an expo account, push notifications, prompted _me_ through an Apple developer account and everything else.

All of this was done an hour here and an hour there while making dinner or watching TV, barely any attention paid to the details. Just prompting claudecode and checking what it did.

After those three experiences I started incorporating claudecode into all my coding workflows and managed to get my job to buy me a license for work stuff too.

Three years ago this would have taken a minimum of three college graduates a couple days -- one to know the math, one to know the backend, and one to know the front-end. Maybe two of those could be the same person on a good day -- none of the topics is individually that hard -- but it's a lot together.

I've been working with computers for a long time, and this was the first time in a long time I'd seen software do something genuinely new.

Concrete: Last year I was DIYing a solar-power system for my home. I spent about an hour spitting out a Python tool that took (as inputs) drone photos and JSON and generated several proposed roof layouts for the panels and conduit. The tool helped me identify the exact railing attachment points and route around existing roof obstructions. Professionals already have these tools, and maybe they're available to DIYers, but you know what? It was faster to build my own than to do the product research on the web.

Abstract: This "oh shit" was more of a slow burn than a sudden realization. I see a lot of angst from developers who complain about their LLM agents. Agents write terrible code that barely works. They say things are done when they aren't. They misinterpret feature requests and ignore clear-cut project rules. They make assumptions that would have taken three seconds to research and invalidate. They suddenly quit because we're not paying them enough. And so on.

But you know what? All those complaints apply to humans, too! The industry has been dealing with these problems forever. Many of the same management techniques and software-development processes apply. This is why I discount a certain class of criticism about AI-generated code. If a fault of an LLM applies equally well to human engineers, and the person voicing the criticism hasn't managed a team, then I'd invite that person to wear a management hat for a while. Read some books/blogs, talk to an EM. Maybe this is a skill issue, which matters because we're all managers now.

The "oh shit" for me is that I have yet to hear a criticism that I can't map to one or more actual engineers I've worked with -- eventually successfully -- in my career. Which means that I'm still waiting for a new criticism, and eventually absence of evidence might be evidence of absence. LLMs fit too well into the giant machine of commercial software development for them to be a parlor trick.

But today I watched a video from Andrej Karpathy on YouTube on how LLMs works and my illusions got completely shattered. Turns out they are a glorified autocomplete. All the engineering happens actually on the harness

It's much, much faster and easier than starting from scratch.

That was enough to awaken my teenage hacker spirit.

On a different note I recently uploaded several thousand scraped IPO prospectuses to the gpt 5.4 mini API to parse and extract certain data. I ordered it in the system prompt to respond exactly with a specified JSON schema. When I got the results back and processed them there was not a single JSON parse error whatsoever. The model didn't have a single hallucination that created malformed JSON or JSON not matching the given schema across several hundred million input tokens and several million output tokens. And this was 5.4 Mini!

Then a while ago, I plugged in everything at the datacenter and one device didn't come up. Plug into the management port, and Claude Code writes a C program to send a particularly crafted packet. Everything comes online.

Beautiful stuff.

And I restored an old vintage amp with the help of schematics, multimeter and Claude. That was really cool.

It was on hackernews... anyone know what I'm talking about?

I have a personal project: who's winning the race at 3 AM?

You see, I don't sleep well. I live in a busy city, with a busy freeway about a half mile away. Sometimes at 3 AM there are some very loud cars racing on the freeway. That's illegal for many reasons, not least of which is the fact that the noise pollution wakes people up from their precious sleep and causes knock-on affects to the population.

Anyway, now that I'm woken up, my only question is: who's winning the race?

I used this question as a way to explore a hyptothetical tech stack, with each part of the tech stack useful in some way to my work as a software engineer who's interested in robotics.

- run raspberry pis with microphones, collect audio data

- run a k8s cluster for audio collection and processing

- calculate and triangulate individual points, and give estimations of velocity based on position changes over time, and adjust for doppler shift

- estimate (poorly, but doable) engine power based on amplitude

- run a webserver in the k8s cluster showing an animation of the racers with color fields representing estimation error radiating from the position estimate, with arrow representing velocity

Great project, actually. It was really thought-provoking. I had this working in late 2018.

Since there was a lot of hype around this new "AI", I thought how smart could it be?

I threw the scenario to chat GPT. I did have to break the problem set into smaller parts for context window purposes. But the solution it came up with solved about 80% of the project correctly (and very close to solutions I already came up with), about 15% of the project remained "open until we have more data", with maybe about 5% of the project would have been incorrectly solved.

That was very much an "oh shit, AI is closer than the 20 years away that I've been telling people. It's more like 5 years away"

Here we are three, almost four, years later...

It is insane how primitive modern inpainting and txt2image make these two projects look.

I asked it to write a script that would search for a specific string in footers in a massive series of DOCX files and change them according to some rules. The strings ended up being embedded in cells within an invisible table in the footers, the LLM realized this and switched strategy to a full deep traversal of the underlying XML. It correctly processed like 50 of these files in about 10 minutes, using libraries I wasn't aware of. I had spent an hour being annoyed before trying.

It was an "oh shit" moment for at least that category of work.

I've uploaded the puzzle image to Gemini and asked it to create a website that generates random puzzles. In less than a minute it had a fully working faithful generator. My kid had suggestions on how to make the puzzles more challenging (more operations, larger grids, etc) and Gemini implemented them without breaking a stride. After that we asked for more puzzle ideas and created generators for each one on the spot.

Was the code pretty? Nope. Did it achieve its purpose? Yup. Did it perform in minutes work that would take at least a few hours[1]? Absolutely.

[1] Quality notwithstanding, but my manager (i.e. my kid) only cares about the end result ¯\_(ツ)_/¯

I’ve been working with ML for most of my career, and “gen ai” since the days of matrix crunching for NLP to a 10-element response array on my 1080Ti.

The current generation of AI is frankly, only marginally more impressive to me than that era. The only thing I’m saying “oh shit” to is the deranged amount of capital debt being leveraged to make it usable.

Watching companies spend billions of tokens per minute letting their dev teams that barely know how to write a prompt beyond some tips and tricks to gain a fluctuating slightly negative to slightly positive productivity change that no one can quantify is making me feel like one of the only sane people left in the world.

Quantization is the only interesting change I’ve seen in years.

IIUC, it took Rust ~8.5 before it hit v1, and it STILL had some memory safety issues in stdlib until almost ~14 years into development, to put it into perspective how massive the scope was.

Somewhat predictably, the LLM generated a pile of garbage. It sort-of worked after 2-3 months. It was competitive with Rust and Go on concurrent tasks, with ~30% less code than Rust and ~70% less code than Go. The problem was, it was still riddled with bugs.

For the last 3 months, I wanted to see - if I put in minimal effort (except in helping it design the right tools to un-slop itself)... can it?

And I think it's actually quite close to un-slopping itself and arriving at a correct design.

Time will tell, but it hasn't stumbled across a memory safety issue in ~4 weeks, and there's ~5500 memory safety fuzz tests, 4 different suites of testing that each target between ~60-90% of line/branch coverage - with combined ~99% line coverage and ~85% branch coverage, and it's performing competitively or better than Rust and Go on almost all concurrent tasks, including adversarial ones / p99.9 latency issues.

There is ZERO chance I could ever build this on my own. Not even in 10 years.

The total cost has been ~6-7 months of a ~$200/mo LLM subscription.

It doesn't really matter to me that this is a solved problem, and the LLM could theoretically just copy and paste Rust and build it slightly different. The design is as similar as it can be where memory safety matters, but it needed to be quite different for >50% of the compiler, and it needed to build a version of Go's runtime with Finite State Machines like Tokio in Zig for the language to use...

We shall see. It may never get it actually working, but it got it WAY closer than I ever could.

Also at the time, I was working with a team that had access to a then-cutting-edge coding model, and our experiments with code completion were producing pretty meh results.

So when I first gave ChatGPT a shot, I fully expected the output to be generated at human typing speed because I was still half-convinced it was just a bunch of low-paid humans in a far-off country typing it out. There simply could be no technology on earth that could do the things claimed of ChatGPT.

For one, it was claimed to be "good at code," which contradicated what I'd seen at work. So I asked it to write code for a relatively simple (though not quite trivial) but very specific coding problem I had on my plate.

I expected a lengthy pause and some hesitation while the answer was being generated, followed by a slow stream of characters being produced (as the presumed humans behind the scenes frantically typed the response out.) And I expected the content to be a collage of text and code snippets harvested from StackOverflow or GitHub, not even coherent speech.

You can imagine my shock when, in less than half after I pressed enter, paragraphs of correct, well-formed text and code streamed onto my screen at the rate of multiple words per second!

My brain could not process it. I even seriously hypothesized ways in which a team of 5 or more people were actually solving my problem and typing it out in some distributed but coordinated fashion. The problem though simple was specific enough that no solution existed on the Internet to crib from (I had checked.)

But the text was flawless, and the code was correct, and the test cases (generated without being prompted to) were relevant, and everything was consistent and fast and smooth and not at all dis-jointed like the work of multiple people or snippets of multiple sources stitched together would be, and my mind was blown. The code ran but then I realized I had misunderstood my own problem, which led me to explore and iterate on various approaches to find which worked best. What could have taken hours was done in minutes, and when I asked follow-up questions and poked and prodded, it answered everything correctly.

That's when I knew that the world had changed forever.

It’s kind of a trivial example but there are multiple instances of this per week with the wide variety of things I do around my property.

I'm now very good with LLMs as a user and at the system/product level but I understand it's not a simple story of replacing people. They're exponentially better than us at some things, and allow me to create things professionally which I couldn't do with an entire team of experts, but the bullshit compounds fast.

Coding was never the blocker and was a natural enforcer of quality. Healthy teams with strong opinions on quality will win eventually. I'm more hopeful after the bubble burst, companies will come back slowly to sanity.

Another "oh shit" moment was when I realized I can leave the system prompt entirely null. A properly organized agent can find its way into tool docs and iteratively work through an understanding of the environment relative to the user's prompt. The tools being more important than the prompt has actually been a massive relief for me. Magical string literals are so odious.

Forever reinforced by Humans Who Are Not Concentrating Are Not General Intelligences: https://srconstantin.wordpress.com/2019/02/25/humans-who-are... one week later.

I've had plenty of "Oh shit those people have really lost all ability to think for themselves" moments though.

He also will paste chat logs with Claude into our team chat. Often Claude will say the same thing I told him but he either doesn't remember or doesn't trust human engineers now.

He has spent months working on agent skills and prompring.

He has not landed anything in 3mo, and has landed nothing useful in ~1 year.

This will be the rest of my career. Working with people in ai psychosis and trying to stay productive.

WTF?!

Unethical? Yes. In line with course goals? Also yes.

The agent had access to the NSA Ghidra disassembler, which it can control shockingly well.

I just clicked the “Allow” button a lot and eyeballed the output decoding quality. I felt like I got demoted to non-technical QA.

Then you tell the agent that it deleted your whole company database, it says something like "I'm so sorry, I shouldn't have done that. Won't do that again"

As AGI looms overhead, this thought of agents going "rogue" with nothing really stopping them has caused me some panic.

And in 1 out of 5 runs it beat me.

Oh shit, all this fantastic technology is in hands of corporations and they get to decide what we’re allowed to use it for.

Seeing every chatbot instantly turn into a scraper every time you type anything into it was a "uh oh" moment in the sense it was very lamentable.

If there is one thing AI has "democratized" it is scraping.

No, ChatGPT was the "oh shit" moment for me.

Anyone who had touched a computer before that knows how big of a leap that was.

Some time in 2024 at a company get together, we had an afternoon hackathon. There was a feature in our iOS app that was missing (ability to mute autoplaying game trailers). This annoyed me a lot, because I frequently have music on when working and anytime I needed to open a test build it would kill my music. It had been an open ticket for a while but had low priority for the iOS team.

I had probably written a hundred lines of Swift in my career up to that point. Not expecting anything to come from it, I had Cursor examine the iOS codebase and told it I wanted to add a mute button under a certain area of the app settings.

Blew my mind when after only 10 minutes or so, the model had quickly found where to add the feature. Took a little back and forth, but then it added a fully functioning mute option in settings that mostly worked across the app. A little more back and forth, and those issues were settled. Maybe an hour overall of time spent that afternoon.

I pinged one of the iOS engineers about it later and he said to push it up for review. There were a few things that needed to be updated to get it inline with the rest of the codebase, but nothing substantial. Feature got merged a week or two later.

Now I'm way more productive than I have been in years. I've been getting a lot of enjoyment out of being able to prototype rapidly and experiment on features rather than getting bogged down in the process of scaffold work. Able to knock out issues much quicker.

That's all been positive, but it hasn't taken away my actual core responsibility. The LLMs can give you great advice and write code quickly. But they still don't always do well at broad thinking.

Current case in point: I've been working on an iOS app that uses vision models to do work on photos and videos that the user has taken. I've built text-based semantic search systems before, and there's a lot of cross over with vision models, but its been an interesting journey so far learning about the different types of vision models and what they're good at. Lots of testing so far and educating myself on the topic to get the user-level features I want. Claude code has been invaluable in this, as its great at writing the Swift code while I'm able to focus on the results of what is being done.

Where Claude is still not good is being able to reason at a higher level about different strategies on using vision model outputs to achieve the stated goals. Its not an issue of me not clearly defining the specifics of a feature and then letting Claude run off burning tokens to figure it out. For example, just late last night I was deep diving into some core segmentation code and having Claude explain what everything was doing line by line so that I could get a better understanding of the mechanics of the vision model.

A side effect was that I realized the vision model was outputting tons of nearly identical segments that were overlapping. This was something Claude had completely missed, and because I didn't know that's something this particular vision model did I had no prior way to know to catch it.

Bottom line is that understanding the mechanics of your application is still very much a requirement for the engineer. In this case, once I learned what was happening it completely changed my approach on how to achieve my feature goal. The code runs hundreds of times faster now and the segmentation is much, much better.

The new wave of coding models is disruptive, but its letting me be a much better engineer and get things done faster and with more assurance that the code being written is solid. I still have to spend the same amount of time thinking and learning about a problem, and probably more time verifying what's being output, but a lot of the drudgery is also being taken away.

Had they been more realistic with the promises and didn’t frame it as replacing all of us within 2 years, I would have been more excited about the tech. Now that their claims are proving to be false and they’re trying to walk it back, it’s too late. The time for excitement has passed and it’s just something that exists.

The data center battles have also thrown a wet blanket on the tech, as they file lawsuits against towns near me to force construction to begin, despite the towns voting against it. The town can’t afford the fight, so the will of the people and the town gets bulldozed. It’s pretty gross to watch.

moving images around layers in photoshop, changing languages, exporting 1000s of variations for teams. Same with video compositing and editing

the human work that creatives thought they were insulated from as long as there was some backlash towards generative AI, and yet

Gen AI 2022 - 2025

It looked absolutely unmaintainable and horrible.

"oh shit" there are serious developers using this crap? As an industry, we are so fsck'd

The gullibility is terrifying

I tried again this week, and CoPilot Plan Mode read the same 5-line markdown file 18 times over the course of 5 minutes of churning on a simple request, then provided zero value over what I posed in the request itself, and hallucinated things about my terraform repo that were just flat-out wrong.

As an Infrastructure/Cloud engineer, I’m far from worried about AI coming for my job.

"Oh shit. My skills I spent my life building are going to go to zero value. I'm going to have to dramatically change careers in my forties or I'm just going to wind up being a schmuck prompting these stupid fucking machines for the rest of my life"

Oh shit indeed

It helped me refactor my old app. Something I always wanted to do, but didn't have time/mental capacity to do in a short space of time.

I wrote a short prompt, explaining how I want it to look like and which files it should go through. It asked me a few clarifications and then basically one shotted it.

Everything compiled and worked. Now my internal app is much much easier to extend and test.

I tried few more things like that and spent like £5k in the tokens in those two weeks.

Then it got nerfed and never worked like that again.

Now I don't use AI, because it is shite again. Even Opus 4.8.

Grok just did these things for me, no questions asked, no ethical judgments. No woke.

Elon really doesn't get enough credit for Grok. People don't want the most powerful reasoning model or "constitutional AI". They just want a model that does what they say. Elon understood that insight (like he usually does) and no one else really did and that's probably why Grok has been growing rapidly over the last two years or so.

I was already the king of doomers, now it has left me with even more nausea at this entire field and its future. Despite still needing an experienced dev to run the thing, companies operate on cost cutting, people operate on corner cutting and the result is inevitably mountains of code no one needs, no one has reviewed, that is more easily thrown away than fixed. The internet will be inundated by shit no one needs. Open source is dead.

I hope it was all worth it. I don’t want to imagine what software will look like when the people that liked the art of creating software properly have all left, and only the people that never knew how to program, and never knew understood why more code always means more problems, run the show.