I bought an Alesis QS8.1 super cheap in perfect condition (was a top grade digital piano/synth in the 90s).

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.

Related story, while applying a firmware update to my Kawai CA49 piano, I bricked it due to flashing the wrong file (The process was broken, and I got desperate and tried something stupid, which bricked the piano). Claude walked me through looking for signs of life, and since OTA from the phone app wasn't working for me, it downloaded the Kawai Android APK, decompiled the Java, figured out the hardcoded key used for encrypting the firmware update. Extracted the piano firmware update, decrypted it, and then wrote a flashing script to program the piano from my laptop via bluetooth. My piano was back to working within an hour.

That's sick.

Yes, those tools are extremely good at reverse engineering. With a bit of know how, it is now trivial to reverse engineer any protocol or crack any software, often in a matter of hours or less.

A lot of people in the industry have vested interests in this not being discussed openly so you don't hear too much about it, but the implications are huge.

I would be interested to learn a bit more on the how after reading also [0] and the worlk done on patching the Ableton Move firmware with the Schwung [1]. Slightly different but there is an increasing amount of work done on either old hardware and new one exploring patching, swapping or developing new firmware from scratch thanks to LLM/GenAI currently.

[0] https://mforney.org/blog/2026-05-28-patching-my-guitar-amps-... [1] https://schwung.dev

I had that keyboard! I actually really like the piano-ish touch. I remember being sad though, when I realized they’d crammed all the sounds into I think 16MB (or was it 8?) and realizing how bad that was even by the late 90s! I think I still have mine in the garage somewhere… good times!

Hey so... mind sharing findings? I have a QS8 :)

While not the "oh shit" moment, the wave has the same shape.

I have an DigiTech GNX3000 effects pedal board - a digital modeling "workstation" that needs the aged Windows native software or Gdigi to make the most of.

At best, the experience with gdigi was passable; raw access to the patches and controls, the ability to control it from the laptop, etc.

In an hour or so, I had a functionally superior webmidi version up and running in Vercel using their v0 code. It kicked off a wave of subscriptions and referral chasing.

I made it a template - because there are so many gnx3k users out there: https://v0.app/templates/digitech-gnx3000-sysex-tool-GC5LzXA...

>Claude walked me through examining the some of the original software in GHIDRA,

I wanted to be able to decrypt the files on The Complete New Yorker magazine DVDs. The old software was WinXP only, and crashed by the time you turned to page 3 or 4. It walked me through using Ghidra on the relevant dll, mapped out how it was using Blowfish, what the credentials were that it was passing, and re-implemented all of that in a python script.

Now all the files are in plain pdf.

Right now, it's helping me write an extension to the mkv specification for embedded scripts and modify VLC to be conformant, so I can watch Black Mirror Bandersnatch. Already have a buggy implementation, about 3 days in.

I've also had it add BEP 46 mutable torrent functionality to Transmission (and to some extent, to the WebTorrent library).

These are all well beyond my abilities to do casually, and probably beyond my ability to do even if I spent the next 18 months doing nothing by grinding away at it.

I only replied because I thought it curious that Claude apparently favors Ghidra.

Interestingly enough, i’ve been sitting on a project for the last 12ish years where i just took the FMloader lib and used that from C# to turn the djvu files into pdfs. All that was needed was a decompiler and an hour of banging my head on it. I published some of the results a few years ago but need to go back and actually build out a full app.

I'm trying to not do the naive pdf creation, where each page is just the raster. Trying to keep the JBIG2 bilevel, as I get better quality at lower file size. Using jpeg2000 too, where appropriate, but the pdfs are still x2.5 the size of the original. Though, I can have it spit out decrypted djvu files that are exactly the same filesize... I just don't like that format for archival.

If you want the Rolling Stone or Playboy archives decrypted, ReconSuave on github has tools to do those. I got tired of waiting for him to do The New Yorker though.

That's fantastic. Did you use a Ghidra MCP server? It's kind of magical huh?

I've done a similar sort of thing with my camera lens' firmware updater just out of curiosity, and I didn't use any kind of MCP. It's able to write an automated script using the Ghirda API to decompile the program just fine, and then code exploration can be done by reading the code.

Claude needs good variable names a lot less than humans do, so renaming/typedefing doesn't seem to be as necessary.

I have a large token budget as part of my work. A coworker was scanning some repos for vulnerabilities as a test. He found a scary looking remote exploit in a popular project and shared it with me for a second opinion. I spun up a local instance of the project and ran the POC against it: nothing. Turns out it needed some configuration knobs tweaked to lower some security protections.

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.

And written with AI.

What was?

These days when people see consecutive pithy sentences with parallel structure, they immediately holler "AI generated."

For years I've wanted to improve my writing and reduce my tendency to string together long floppy clauses, and now I'm like, well good thing I never did that...!

Actually seems absurdly simple now, but sometime last year I was trying to figure out what I'd need to tow my daughter's car cross country with my truck: what are the trailer/dolly options, what do they cost, can my truck actually tow the combined weight, etc.

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."

Similarly, I used gen ai to review a real estate purchase. I provided Zillow listing photos and serial numbers of all appliances, the electric panel, and a few additional not pictured areas that I took during the walk through.

I prompted the AI to write a report as if it were a home inspector and it actually did a better job and identified some issues the paid 750 usd inspector missed.

From pictures alone? What are some examples?

I presume something like this: https://www.penny-arcade.com/comic/2007/06/22/perfectly-reas...

It noticed a flooding area due to low grass by the walkout door. It noticed mixed 15 and 20a receptacles on the same circuit. It noticed warped siding and recalled circuit breakers still in use.

It very plausibly might have been totally wrong.

Out of laziness I several times asked Claude and ChatGPT each some torque figures and other simple, hard data related to my dirt bike. They often got it completely wrong, but full of confidence every time. I never trust LLMs with hard data, unless you RAG the PDF into the context and even then it's sketchy.

Dates matters. Questions I asked about my Mazda a year ago that were total hucillunations were answered very well this year. To me it feel like the early days of computing. What was not possible one year became possible when a new generation CPU or GPU came out and you have to consistently re-evaluate your expectations or else you'll miss the things that others are discovering with fresh eyes.

I made this personal 'benchmark' of odd and strange questions a few years back when this took off and I would keep re-running these questions whenever some big news came out about a new model and also going back and fourth between the different companies to see where they all stood. (Obvioulsy with clean cache/new accounts)

10 questions: In 2023 it could only get past question 3-4 to reaching the last question and still hacillunating(last year) to providing sources pulled from really obscure books(this year).

For example, one of the harder questions was about the transition of a particular 30 second portion of a background song used in a 30+ year old Bond film that was only played once in the entire film. Went from totally making up nonsense to accurately describing the music theory defintiion of the transition(called a 'stinger') to also explaining why it was done in that particular scene of the film and also providing sources from a snippet of a unrelated interview with the composer explaining his mindset at the time.

Maybe this isn't considered a real benchmark as its not reproducable but for a 'personal benchmark' I came away impressed. I would consider everyone to define their own benchmarks and 'tests' and to consistantly challenge the models to see if there are any meaningful improvements. Now I treat the AI as something to keep skeptical but to also to always consider what it proposes as an answer(ie. dont ever dismiss it outright). I sometimes wonder if this is slowly messing up my biases and maybe thats what Altman, Amodei and others want.

Hard numbers, no. Even high level concepts and theory you need to triangulate and prompt in different angles, across different models, and figure out what overlaps to build a mental mode that’s - even then - roughly 80% correct. It’s better than google, but the information isn’t free

My furnace went out during the 2025 holiday and I couldn't get an appointment with a repair person for 2 days. It was getting very cold in my house so I went into my attic and made several videos of the furnace attempting to start and gave it to gemini. It diagnosed the issue immediately and had me spin one of the components (a small exhaust fan) while the furnace tried to fire. It came on immediately. I had to do that several times, but it worked until the HVAC service showed up.

Had something similar this week. Gas dryer started, but wouldn't heat. Gemini suggested it's often a thermal fuse. Took off the back panel and uploaded a photo to Gemini. It pointed me to the fuse (e.g. "the white rectangle above the blue and red cords") and walked me through testing it. Not only that, but it also linked me to the part I needed after I provided the model number of the dryer. Finally, it recommended cleaning out the vent as the fuse likely blew because heat wasn't venting properly. After a thorough cleaning of the exhaust and a $5 fuse the dryer is working fine.

Very similar thing this week, and an interesting story to go along with it!

I called my normal HVAC company for my rental home because the tenant reported the AC wasn't cooling the house. When I called, I got one of the latest AI voice assistants to help me, and it was an awful experience and I ended up not hearing back after the assistant told me the office would call me back.

So, I went over to the house and used ChatGPT to help me diagnose the issue by taking some photos of the compressor panel outside. It walked me through what to check, I provided some diagnostic codes I witnessed... and it walked me through the very simple repair of replacing the $25 capacitor. It was going to cost me almost 4x that just for the service call to diagnose what was wrong in the first place.

So, the weird experience was: Gen AI made me lose trust in my normal HVAC company, and more Gen AI basically allowed me to replace my HVAC company and do the repair myself all in one day.

I've been fitting a kitchen and chatGPT has been useful to bounce ideas off and resolve issues. Of course if IKEA's documentation wasn't so sparse I wouldn't need it but that's another story.

I guess I'm seeing similar benefits to a novice programmer. Professionals would scoff at my work but they are expensive and difficult to work with. Meanwhile I'm getting the job done.

On the other hand I'm not touching AI for any development work. I'm too worried about my skills atrophying or not properly learning anything new.

Ikeas instructions are such an oddity.

It feels like there is precisely enough information to deduce each step. But only just enough miss one clue and you have something on upside down on step 7 that you won't notice until step 37.

I feel whoever makes them could probably make a wicked NY Times Crossword puzzle.

Similar - had an HVAC tech out to diagnose mine (some intermittent electrical problem was killing thermostats randomly) and since it was intermittent they couldn't figure it out. I ended up using Gemini to narrow down a list of potential problem components and just replacing them all which fixed the issue.

Kind of a superpower to turn anyone with a bit of tech inclination and problem solving skills into an HVAC tech - not a very good one, but one with enough motivation to get the results you need

That's pretty great.

(Though that's also the kind of hands-on troubleshooting step/fix that a person could just google for and find pretty easily back before the internet got all fucked up.)

Your parenthetical really describes my experience with AI searches. 5+ years ago I could find most things within one or two quick searches, now it takes so many that of course I'm going to reach for AI because that's the only way to get back to my baseline efficiency.

Do you mind explain more. Did you just prompt to Gemini what was happening, did you give Gemini photos of Furnance, etc?

> and made several videos of the furnace attempting to start and gave it to gemini

I assume recorded videos and uploaded them in the Gemini phone on their app; and then probably said "what's wrong?"

Gemini is very good at those kinds of things. I recently got some ratcheting straps and needed to use them, but at the time I didn't know what they were called, so I didn't know what to search for on Google. I opened the Gemini app, pushed the button to take a picture (just like in text messages,) and included a message that was similar to "what is this and how do I use it?"

Yes, here is my prompt. It also contained a video: "I have a furnace that will not heat when I reset the power to this unit. It makes some noise within its fan system for about three or four minutes and then I get an error light. Can you help me figure out what may be wrong here?" This prompt is not the best but I was freezing and in my attic.

Oh yeah. I can't remember which LLM, but one helped me repair my dryer.

Gemini almost killed you.

The exhaust blower not working triggered a safety that prevented the furnace from firing.

Spinning it bypassed the safety.

You likely inhaled a lot more carbon monoxide than you know.

Can you elaborate? I interpreted the same as the other comment that the blower fan just needed a hand start and kept going after the furnace started up. What you're saying only makes sense to me if the spinning the fan by hand allowed the furnace to start by bypassing the safety at startup, but wouldn't that mean that if the exhaust fan was stopped during normal operation (blockage etc) that the furnace would just keep going, dumping CO into the home?

It wasn't bypassing, I was just helping start because of what I believe to have been a bearing issue.

It’s a pretty normal trick to try while troubleshooting a rotating part.

Helping something start is not likely to ruin your day (unless you get caught in a rotating part)

I was spinning it in reverse actually, but it would be enough to start the exhaust blower. It would also re-start pretty well for ~6 hours. It was probably the bearing. Also FWIW I have multiple carbon monoxide/air quality monitors and nothing tripped or alarmed.

From the description I thought that a degraded capacitor or lack of lubrication made the blower not start on its own, but the blower (and the whole furnace) would work if given a manual startup spin by hand.

Welp, AI almost killing someone is definitely an "oh shit" moment.

Wonder how many AI deaths have occured that we dont know about(since they presumably died). With the adoption numbers we are seeing it much have happened already.

I could go on and on, but Claude recently decompiled the firmware of my camper van, documented all the CAN interfaces, then programmed an ESP32 module to talk to the van’s integrated systems (power, HVAC, lighting, tanks). That sort of embedded systems integration is completely out of my wheelhouse.

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.

> projects I simply could not have ever approached alone.

I think that's part of the divide between enthusiasts and naysayers. If you use GenAI on things that you couldn't approach alone, it's an incredible tool. If you use it on stuff that you're pretty good at, it's not a gamechanger (and if you're an expert, it's a minor boost at best). Many people's job are about doing what they're an expert at.

I think part of it is we often notice bad AI usage. The llm generated "art" by someone with bad taste, or the patches to open source projects by people who cant program at all and are teerrible.

If the use is half decent people just dont notice it.

I find it's a huge boost for my day-to-day work.

If you work on architecture and Claude docs, then you can essentially just have it fill in the gaps. Work then mostly becomes a matter of defining what the next piece of functionality is (which you can also use Claude to help with).

The stuff that used to take days now takes hours. It's not perfect, but if you get your codebase into a good shape then the payoff is huge.

And in a team setting it can really accelerate tech debt especially if used by people that know just enough to be dangerous.

Same. I'm a DevOps engineer, so a jack of all trades master of none type of guy, and Claude Code backfills my knowledge gaps and turns me into kind of a superhero. I think it's key to already have a pretty good idea of what you're looking at, though.

Literally just last night I have Claude Code the following prompt, verbatim:

"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).

That's amazing, as someone who struggles to find something useful to do with LLMs. How long does this take, several minutes or more? Do you need a paid version of Claude Code for this?

It sat there for about half an hour working out the problem, step by step, before asking me for the preferred solution. At one point, it was trying to decompile the .APK, so I interrupted it and reminded it that Kodi was open source - it was welcome to clone from GitHub.

The only other feedback I gave it mid-process was wrong (I said that the crash probably wasn't caused by cache trimming, it ran some additional tests to confirm that its hunch about cache trimming was right).

This was with the paid version of Claude Code (I don't think they offer a free version at all; that's a Codex thing). The $20 version is as smart as the $200 one, but once you work out it can do stuff like this you'll quickly burn the $20 token limit. :)

The other thing that helps is a CLAUDE.md file - authored of course by Claude itself. Mine's here: https://github.com/EspoTek/.claude/blob/master/CLAUDE.md A lot of it is probably domain-specific for the stuff I do, but the "Working with unfamiliar data or systems" section is bloody gold! Stopped the bullshit completely!

I tried to see if an LLM service provider could rewrite some legal docs where nothing was hallucinated in order to follow a consistent format to see what may be missing in the document. It could do that.

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."

Yes, the best results I've had using LLMs are for tasks where simply reading and reformatting/translating/summarizing are the goals. They are much faster and less prone to boredom doing these things than humans are. For now.

My son is in a lawsuit with his bank where they put through fraudulent charges and wouldn't charge them back then the bank sued him for the money. He is using Claude and Gemini fighting the original lawsuit and now has a counter-suit 100% using AI for everything. He puts it into different AI's to check everything against each other and to come up with more ideas. He started with ChatGPT, moved to Grok, then Claude, but now Gemini is turning out to be the strongest.

I'm about as pro-AI as anyone here. I say this with love: anyone using general-purpose, consumer-grade AI for healthcare, law, or taxes is mad. Best wishes to your son, bless his heart, but please have him consult a qualified lawyer before showing up to court with model-drafted legal documents. Among other things, those chats are not privileged information[0] and the banks could subpoena chat transcripts to see what else he might have told them.

[0]https://natlawreview.com/article/new-york-court-rules-ai-doc...

He has had multiple hearings and the Judge has reviewed everything. The court clerk reads every submission and before the clerk puts it in the system they have a in-house lawyer review each document. This is pretty far along. The trial is scheduled for October of this year.

The bank has a lawyer, they were hoping for a default judgement because who can afford to fight the bank. The choice is fight it yourself or declare bankruptcy.

As you already know, AI companies trained on every single document they can find. Those include legal documents. The legal system is structured where you have Federal Laws, State Laws, Federal & State Regulations and Court Precedent. Because of this structure it is not difficult for a LLM to figure out.

The only way to "win" as a small is to be pro se and be extremely diligent in understanding what is happening.

Then, it costs you nothing but time.

This. I know how to do this but I don't have the time/energy to do this. "Get me Claude!"

I don't remember one specific moment, but I was fairly impressed with ChatGPT from the first time I started interacting with it. Was I ready to call it "AGI"? No, absolutely not. But it was clear that it was something new, and it was also intuitively obvious to me that "this AI is as bad today as it will ever be" and that predicting the rate of change would be difficult.

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. :-)

When ChatGPT allowed me to calculate stress and load bearing tolerances for a camper based on different materials, suggesting various potential design improvements, with the math and sources to back it all up. Then it helped plan and fill out paperwork for a residential solar project, including full code-compliant electrical work, again with sources to verify. Then there was an open source app that wouldn't run on an old version of MacOS due to them not supporting older OSes, and a coding agent backported support for the old OS and got it up and running.

I had an old astronomy app I wrote for pre-iPhone app store era Nokia phones (N900 etc.). I decided to get Claude code recreate it as an Android app. The old app produced several display pages for things like the positions of the planets. I was having Claude code recreate the app display page by display page, describing the display that should be produced, with no reference at all to the original app's code (or even its existence). After having it reproduce several pages, it added another one unprompted. The page it added was in the original app, but I had not gotten around to adding it to the Android app. The Nokia app's code is still on github, and somehow Claude must have made a connection between what I was asking it to code (without ever mentioning the Nokia app) and my github repository's Nokia code. It correctly implemented the page without me even mentioning the missing page. My jaw hit the floor.

ChatGPT Code Interpreter back in ~March 2023. I uploaded a CSV file (of police incidents in San Francisco) and watched it load that into Pandas, show me some charts, then export the data to a SQLite database file for me to download.

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.

I probably will be burned for this, but with the help of an LLM I wrote a tiny program that captures video from a browser screen (Xbox live online FPS game), passes the video images through a small trained NN that recognizes people forms and presents the video on another screen. That way I can place a green overlay on enemies and they are easier to see on PVP matches.

All that in around 100 lines of code, including the training/fine-tuning of the tiny YOLO nn.

Running local LLM in 2023 and I heard folks talking about interfacing LLM to tools. I wrote a system prompt and told LLM it can call some tools. If it wants to call a function to output func(params...) and do so in an XML tag. I provided a few examples, none of this JSON soup we get today. Then told it I'll provide it the result in a RESULT XML tag and it should use that to answer. Wrote up a harness around that and I had a local model interacting with the outside world. Oh wow! Everything else today about MCP, Agents is all an extension of that thought. Using function calling, I built an agent. I defined a data structure that represent rooms and how they are connected. The room will be marked as dirty or clean. Then I would place the agent in a room and the agent will decide if to go left, right, down or up and into a room. Once it got into a room, it would decide if to clean it or go to the next room. Repeat until all rooms are clean. Basic toy of CS101 AI vacuum agent. It worked!

So being able to get real world input/output to the model and having the model being able to make decisions in a loop and to be able to do it locally. I have been screaming like a mad man ever since.

To share something different, it is less about what I have built, and more about what I have seen my friends (non-technical and technical) build. In a one month span I have seen a lawyer make a personal red line tool, a sales guy make a custom website for a golf trip, another friend make a 3d printing grid-finity project, a friend make a stl file to print a jig for his table saw, and another friend make a full mobile game. It is just really cool to see these micro-projects be created and shared, not only for the utility, but just to see my friends' childlike excitement showing off their project.

I am the CTO of a small NGO (10 people total, only 1 other junior Dev at the time). We supported two apps that were built by consultants. They were a mess. NextJS, React, about 4 micro services for a site that had 50 users per WEEK.

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!

For me it was earlier this year when I started dusting off some old stalled projects and had an agent work on them. In a few days I:

* 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...

Someone in the house pressed the button to update the printer (Brother DCP-L3550CDW) firmware and the CSV page that was the basis for an existing Prometheus exporter (drum/toner lifespan, page counts, etc) stopped being a thing. Instead there was an HTML page with all of the information buried in various divs/etc.

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 guess I've had several of those moments over the last year and a half. But a recent one was that I was working with Claude to create a spiking neural net MNIST classifier in an FPGA for a demo. Claude took it from concept to PyTorch, to training (training a Spiking neural net isn't necessarily straightforward - that's a whole post in itself, but Claude came up with a working solution), and then to implementation in Verilog and through synthesis into the FPGA. I asked Claude to create a drawing app to run on the PC side that would allow the user to draw a digit with a mouse and then click a classify button. The data from the digit drawing app was to be transferred via USB to SPI to the FPGA. I didn't have a SPI adapter yet (it was on order from Adafruit) so I asked claude to let me communicate with the simulated verilog code running in the Verilator simulator, through a virtual SPI interface. Then I went to lunch. I came back to see the digit drawing app displayed on the monitor. I drew a '2' and it classified it as a 2. In another window I could see the Verilator simulator running and the data being passed. Chills.

I have a buddy who's a consultant. His niche area is Netsuite and Oracle (I think). He's an accountant by training and as a consultant his gig was setting up these instances for clients, charging them an arm and two legs. He'd spend a lot of time golfing, and doing these setups was more than enough money for him. In other words, he had cornered that little slice of the market and was making bank.

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.

Sometimes it's the service you provide, not the value. They know it's in good hands, as it's always been (even if they could have rolled their own ConsultBot 2.0)

I have some friends who, since their high school days help some older acquaintances in upgrading their PCs, choosing laptops and phones, helping with setup etc and these older folks have comfortable money and pay him very well above what would seem reasonable. But the trust and years long relationship matter to them.

Llms are great today for buying advice but there are some incentive issues for the future, ads etc. But in some cases the human contact will remain important. In large corporations it's also similar. The money is peanuts either way, and it's worth them for the peace of mind. But this may not hold forever, especially if the more AI literate generation gets to more senior positions.

The big one was definitely ChatGPT upon release in 2022 and specifically when people showed how it can role play as a Linux terminal and you can narrate events like "the data enter is now on fire" and "run" nvidia-smi, it would show high temps on the gpus etc. Or you could "explore" the homedir or some famous person. It convinced me that if it can understand so well how terminals work, tool use and agents are around the corner.

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.

I think all those Steve Spielbergs hiding among the 8 billion - without connections and without hollywood names, having their day without getting filtered out by investor gremiums playing it safe - will produce enough material to be happy cineast for life.

Many Internet technologies like Youtube or soundcloud gave content creators direct access to an audience. If you were funny/creative/talented/attractive/in the right place at the right time/etc you could cut out the middlemen and capitalize on that directly. What is the hidden talent these Spielbergs possess that AI will enable? It's doing the writing, videography, etc. I imagine at most these hidden Spielbergs will be filtering output, like looking at a book of wallpapers and choosing what they like. Maybe some great movies get made, but it will just be the result of selection effects on the sheer number being made.

YouTube is well poised for this.

How will anyone find them, though, if there's so much slop that people stop looking?

Human reviewers a la the Criterion Collection.

It's not information overload. It's filter failure [1]

[1] https://www.youtube.com/watch?v=LabqeJEOQyI

Yeah, ChatGPT both being able to generate conversational text AND act like a terminal absolutely blew my mind - far more than I ever would have imagined the approach would scale to st the time. Since then there have been more impressive tasks accomplished but nothing which put me into the same state of pure amazement.

I was working on a science experiment (electromagnetics) with my 10-year-old kid that was going to be demonstrated at a science fair in his school. We ran into a hiccup with the experiment that we couldn't debug ourselves. I turned on Gemini live video call to help us root cause the problem. It was able to clearly articulate all the possible issues and eventually was successful in making our apparatus work as expected. Turned out the wire that I was wrapping around the screw had some insulation that was not scraped off well on the side it was connecting to the battery. Gemini was able to capture this detail even though my bare eyes could not. My kid and 2 of his friends were impressed not just by the experiment, but because the live audio/video back and forth we had with the AI was almost magical!

Neat, but I'd bet it was "guessing" that rather than actually seeing it.

If it’s “guessing” correctly 95% of the time, is that meaningfully different from “seeing” it for most practical purposes?

I just mean that the image input probably didn't affect the output at all. Could have just told it "I'm an amateur doing bulbs and batteries and it doesn't work" and it would give the #1 search result on forums for that which is "did you strip your wire."

I feel like I'm in the audience at a magician show, except most of the audience is breathlessly amazed and doesn't understand how easily tricked they are.

I helped train some of the first "magic" models at OpenAI[1] and it was a wild ride. We were a pretty sane + skeptical team and we weren't totally convinced the models were as general as they seemed, but the query that convinced me (and later got included in the paper[2]) was "Why is it important to eat socks after meditating?" (something that almost certainly did not appear on the internet before).

An interesting follow up would be when did you realize GenAI wasn't as good as you thought in that "oh shit" moment

[1] co-author of InstructGPT/RLHF/ChatGPT

[2] https://arxiv.org/pdf/2203.02155

We have been using one of the main AIs for fixing errors or bugs in our codebase. We started early and most of the suggestions were shitty and we would pass them around as jokes. We were trying to improve it, and a little over 1 year ago, it started making very subtle fixes that were very nuanced but correct. I was shocked and thought "Oh shit, my job is gone."

I've had many, but a recent one was when I figured I'd try asking Claude for help with my attempts at learning to draw, specifically anatomy.

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.

That's super interesting! What sort of feedback? Anatomical feedback ("Uhh, that's not where arms go...") or drawing (like tips on shading etc...)?

It pointed out that the shoulders were too rounded and the perspective was incorrect for them, making it seem like the arm was just appearing out of the torso. It pulled up images to explain that I wasn't correctly indicating the deltoid's presence. This helped me understand why I felt that the shoulder was hard to distinguish from the torso.

I also asked for help on how to make my posing less stiff and it used the Python script trick to roughly indicate the line of action and how they were very straight and parallel and to reduce stiffness I should have more curves etc.

This wasn't really at the point where I even asked for shading advice.

Not coding, but reading logs.

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

Two of them:

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-...

I got early access to the pre-ChatGPT OpenAI API (actually by pinging someone from OpenAI who posted about it on HN). At work, we were setting up to play a livestreamed JackBox game for a charity event. This would have been in 2019.

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.

(1) Watching it do log file analysis in seconds that would have taken me hours (edit: days really), and which I would therefore never have done in the first place.

(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.

I made a personal project game 20 years ago that I knew had a bad bug in it and so I never did a final release but at the same time I never returned to it to debug but yesterday I noticed it at the top of my Github, it started with A, so I described the issue and Claude found the bug instantly and after a few back and forth discussion we came up with a good fix that I'm satisfied with. So I guess I can do a final release now. :D Sweet - feels good to put that to bed.

I didn't have a slightly panicked moment, but sometime in the last year my approach to programming changed.

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.

Been using it to manage an estate and just being able to shove all the documents right into an LLM and have it spit back out perfectly worded emails as well as keep track of check lists of things I need to do with an automatically create a ledger for me in sheets. It's been a huge mental load off and I've instead been able to focus better at work and the labor costs saved to me have been immense. Just on this one little thing. I'm one of those people that over thinks correspondences and letters and it ends up causing me to be stuck on something so being able to ask for just the right wording has been super helpful to me.

None so far. When I try to use these language models in the primary areas of my expertise like SIMD or GPGPU they fail to do any good. When I ask them to implement some general-purpose stuff, the output is too low quality to be useful in my software.

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.

I tried to get it to generate code to program one of my BitGrid simulators, and it kept producing code that failed, over and over. It was then that I figured out that it can only do CRUD apps and the like, things it's seen over and over in its training data.

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

GPT literally generates perfect code for me in languages that do not exist anywhere in its training set, so I’m not sure how you’ve achieved this level of failure.

Try working in anything domain specific outside of common CRUD patterns. E.g. scientific software development where you describe a problem + give data. I have yet to see a single example of feeding in a problem in natural language involving a specific scientific domain that wasn't pretty catastrophically incorrect.

But yeah, if you want to feed it math and get code, it's reasonably okay with that. All LLMs I've used seem bad at understanding things that don't look like broad human knowledge. I've seen this same general issue across many different models. (And to be fair, geology, geophysics, and remote sensing are what I'm testing, and their semi-rare niches.)

It's also quite dangerous because it's not obvious that what it's doing is complete hallucinations unless you actually are a domain expert. Things _sound_ reasonable. E.g. "this is likely feature X" which _does_ exist, but is absolutely _not_ relevant to the problem or present in the input dataset.

But my current employer is pushing this exact thing (human language + scientific data + LLM -> advanced analysis of scientific data by LLM -> business decisions) and it _really_ worries me. It often gives the rough equivalent of "Start the procedure by severing the patient's aorta. Once they stop moving, you can deal with the hangnail". Just in very reasonable sounding language. And a lot of people don't know any better, because most users aren't domain experts.

Stuff it's not directly trained on is going to be flaky and sucky. It was like that with programming at first too and it still is sometimes. It's hard to imagine this won't improve with better more focused training. They focus on improving "CRUD" for obvious reasons. The specialization era hasn't begun yet.

Your domain, while I'm sure it is very interesting and complex, if it proves economically interesting will be cracked as well.

Just for some context, the domain we're talking about is oil and gas and mineral exploration. E.g. At my previous job, I used to personally manage a >$400 million per year budget and that wasn't even considered significant. We had multiple >$10 billion per year projects ongoing. That was 10 years ago. The amounts are larger now.

The issue isn't a lack of economic interest.

It might be a lack of training data in addition to inherent complexity, but it's certainly not a lack of economic interest.

I have no idea how and why GenAI would be useful in your profession. I'm sure a lot of money is moved there (not sure about the profits though), but it's not clear to me how software itself is budging that needle. I suppose better algorithms and better understanding of geology will do it, but software itself seems just subservient to that goal.

I guess what I'm saying is that "domain knowledge" is taking software development for a ride here. The software is just the vehicle, the science is the engine here and I can see why companies like OpenAI start going for the low-hanging fruits first instead.

Your specific company might be profitable, but does automating "mineral exploration" give you leverage over quite literally all other domains? My guess is not. For "CRUD" it is a resounding yes, it provides gigantic leverage. Once you automate basic software development you enter a new world. 10 billion, 10 trillion, all bets are off. You automate the creation of the next iteration of automation and on we go. Let's hope it takes a while for this take off. I can't see ourselves being ready for it.

My guess is it'll take a decade or so for real AI science to start taking off though - if that soon - so you're probably fine for now.

Yes. My point was that LLMs aren't currently good for everything. The original commenter literally said they were good at everything and I offered a counterpoint of something they're not good at: Most science.

(And yes, a lot of science is software. Analysis is software.)

Skill issue. I've seen LLMs used in this domain to get mindblowing results. You won't see it published anywhere though.... =).

Disagree, someone like the other commenter who points out LLMs don't even understand the domain concepts correctly versus someone who uses it anyways for corporate proprietary results have very different standards for what is acceptable. If you wrangle an LLM with harnesses and clever prompts you could use it to get some amazing results but that has more to do with trial and error and creativity, not some kind of fundamental skill of using LLMs.

We were experiencing abnormally high electrical bills and I could not figure out what was happening, so I downloaded the granular usage data (15 min increments) from Duke Energy, explained what we had in our house and when we typically used those items (washer/dryer, EVs, etc), provided a rundown of our energy usage plan, then asked Claude to build me a Streamlit dashboard that would help us understand what was going on and predict what was going to happen over the next months. The dashboard had a few simple toggles a levers. Claude was basically able to one-shot this, knew how to manage the XML from Duke Energy, etc... In about 20 minutes of prompting, I had a very comprehensive dashboard that was extremely helpful not only in diagnosing that specific issue but also in helping us understand how to further lower our electrical bills.

But ... who was phone?!

E.g, what was it? Don't leave us hanging!

This can be a product.

Going from one off prototype to robust product is a huge leap.

I think these ephemeral context tailored projects are really great and useful. But these are not to be thought of as products. They work for you specifically, and people who are tech-brained enough to be able to formulate the complex requirements into a coherent prompt are not like the average user you'd have to sell a product to. It's much easier to make software to intelligent users.

Homeassistant already has tons of integration into power providers and easily let's you pipe in local data if you have it. In addition - can it be a product if anyone can just type what this guy did into an LLM? What's your moat if anyone can just replicate it?

It doesn't have to be a durable.moat for it to be a product that makes the author money, just right place right time. If it's gonna cost me a bunch of time and effort and tokens, and the cost of the product is lower than the time and effort and tokens, then I'd rather pay for the product.

Right now we're in $1 Uber ride territory. That $20/month OpenAI/Anthropic plan isn't going to last forever. If it's going to cost me $100 in tokens to replicate the product, $20 is a cheap no brainer purchase m

The comment above literally said this took them 20 minutes of prompting. That doesn't sound like much if any value add.

Sounds like something people say to locksmiths.

I’m making $1000/month off of an app that was initially a single prompt.

There’s a gold rush right now. You absolutely can turn these ideas into products.

It's not going to be a particularly expensive product, but a product it can be.

From actual use I've not had a "oh shit" panicked moment yet. More like a bunch of "Holy shit" euphoric moments.

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!

I do genuinely wonder if you’re correct that other people will begin to expect it. I feel I was suddenly able to do stunning stuff about a year ago, and I recall thinking this is nice but everyone will catch on to my secret soon and I won’t be exceptional any more. But 12 months have passed and I don’t think this has really panned out yet. Weaker engineers just don’t seem to understand that they can just ask AI things. Eg the other day another engineer spent like 3 hours trying to hunt down a particular line of code so I asked AI and it found it in like 5 minutes. I showed that to him, but then he immediately got stuck trying to find something else for a few more hours, so again I asked AI etc. It’s very baffling.

There is definitely a learning threshold and it's still early days. Not every developer has found out how to make efficient use of these tools yet. But I think most will, soon enough.

But I think my own clients will soon start to question why some feature takes ME a week, when I was able to deliver another feature in a day or two.

That they are features that used to take months, and even delivering them in a week is a goddamn miracle by 2025 standards, will not be relevant. They won't expect such features to take months any longer, based on what I've delivered earlier this year.

So I think that the past few and maybe next few months, maybe a year, will be remembered as a "happy hour" for this tech as a developer. These are the days that we'll talk about saying "those were the days". :)

I am still optimistic that "the normal" in a few years will be pretty much like it has been before - I'll be delivering features at work and tinkering with hobby projects at home, and the major difference will be a much larger scope and ambition for both.

Direct use of AI is going to be a filter on a lot of people - some permanently I suspect (especially say older people). But perhaps this will be short lived as the interfaces to AI are improved enough that everyone will benefit.

Look, not to brag but DALL-E's "armchair in the shape of an avocado" was mine (https://openai.com/index/dall-e/). I remember trying to convey the gravity of this capability to my friends at the time, who I guess were not as impressed as me.

Thats insane! I cited your image in a humanities paper during one of my freshman year classes.

I think the GP just means it was their oh shit moment, not that it was their image...

I think GP meant that yes, they were the one who had that image generated, and the oh shit moment was that it worked.

Yes confirmed, I did not author the DALL-E paper lmao

Realising in a recent benchmark that gpt-5-mini gives better results on some tasks than gpt-5.4-mini and event gpt-5 or gpt-5.5

So many. First was when I saw GPT-2 create jokes that were original and kinda funny.

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).

For me it was Suno, not any of the coding tools. I prompted it to write a song about my family's little dog, told it a few things about the dog, and it came back with a K-pop-style anthem that had a super catchy melody and lyrics that made my wife and me laugh out loud.

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.)

Some business users spent ~30 minutes on an internal process, and we prototyped an "Agent" in Slack to take over. At first it didn't work, then it didn't work some more, eventually it ALMOST worked. Then one day, it worked, and the old business process died never to be revived.

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.

I wanted to understand the implementation of some numerical algorithms, and the tech reports were not enough.

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".

Working on a Spice compiler to convert schematics for classic guitar pedals into real-time executable code.

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.

That sounds pretty fun. I guess I could just Claude to do this hehe but are you sharing?

I'm not mostly Claud-ing it. Perhaps I should. But in the difficult bits, it never ceases to amaze me what these tools are able to do.

Yes, if it matures, it will go open source. Not immediately clear at this moment whether it's feasible to do an Operational Transconductance Amp in realtime. :-/

And it's competing for attention with the 2.0 release of this at the moment:

https://rerdavies.github.io/pipedal/

Just went GA, so I'll have some cycles to come back to it.

I wrote a thousand lines or so of Javascript for transforming JSON into DOM fragments with attached event handlers. I then asked an LLM (some Anthropic model from around a year ago) to write a test suite for the module. It wrote dozens of useful tests and managed to reverse engineer the entire module because all of the input and outputs were exactly correct. It did not actually execute the code to build input/output pairs.

AlphaGo. Reinforcement learning on math with proof assistants was clearly going to be workable after that, even if not right away.

The announcement of GPT 3, hands down. That's the day that my mind was blown.

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).

That was it for me as well. For it to be able to produce coherent, useful results for arbitrary text and code, even as mere autocomplete, was when practical software intelligence crossed the line from a vague “might happen at some arbitrary point in the future” to “this is achievable using today’s knowledge”. GPT-2 was interesting but was miles away from being reliably coherent enough to cross that threshold.

There have been big steps forward since then (chat, tool calling, etc.) but they have all felt like pretty inevitable evolutions of GPT-3 rather than things we weren’t sure were possible. Straightforward applications of engineering rather than scientific breakthroughs.

1. ChatGPT first public release (I am not one who saw early GPT models) I think late 2023 iirc?

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'm kinda of surprised that so many here on HN were dismissive/unaware of the capabilities and potential in the DALL-E days and earlier. I feel like this is the sort of forum where most people would be both aware of advancements and aware of their potential.

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.

Recently, Claude (through Copilot) found a hardware issue on our product. I was asking it to find an issue in a specific feature of a device driver, that could cause what we observed. It determined the feature was correctly implemented.

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 would say the first time I did “vibe coding”, when I tried Claude Code with Zed’s agent integration in January this year.

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

Back in the times of GPT3 text completion, right before the API came out, a contemporary art museum asked me to collaborate on a project. The project was supposed to include a chatbot, and I was like okay I can probably hook something up.

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.

I also remember doing this. Chats, first parts of books, title pages and all, just to give it a chance of saying something in the ballpark of what I was looking for. I remember very vividly that chats or books by Linus Torvalds would be more technically accurate that say Lincoln. It's obvious of course, but I found it really enlightening. It could code a bit actually, not great, but well enough to push me into an existential crisis. I started doing a master to re-educate myself because I could see "interesting" times coming.

I actually emailed OpenAI back then saying they should be careful because this is much greater than the public or even they themselves think. They actually replied! They thought it was cool, but very limited and I shouldn't be too impressed. Good times.

"Write a bible verse ... explaining how to remove a sandwich from a VCR" https://x.com/tqbf/status/1598513757805858820

This is a small one, but significant to me.

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.

Two things:

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.

My first came in late 2016, when Google Translate switched from statistical machine translation to a neural-network-based system. I had worked as a Japanese-English translator and lexicographer for two decades, and I had been testing various machine-translation services over the years. For translation between Japanese and English, at least, they were uniformly worthless: the output for genuine texts was mostly incomprehensible and could not be used for any real-life applications. The neural Google Translate, while still far from perfect, was suddenly useful for some purposes.

But the neural models were still not translating meaning, which is the whole point of translation. I devised a variety of tests to see if GT could identify the meaning of ambiguous words from the context, and it couldn’t. One example I would show people was the sentences “I was born in 1998, and my sister was born in 1999” and “I was born in 1999, and my sister was born in 1998” translated into Japanese. Japanese uses different words for older and younger siblings, but GT translated “my sister” with the same word in both sentences. It was easy to come up with other examples where GT would fail, such as when the meaning of a word could only be determined based on context in a previous sentence; at that time, GT seemed to be translating sentence-by-sentence, with no consideration of what came before or after. I kept waiting to see whether computers would ever be able to handle meaning when translating, and for years there was little progress.

A smaller shock came in mid-2022, when DALL-E 2 was released. It’s ability to create images from natural-language prompts suggested that something deeper was going on than just statistical correlations. But I couldn’t see yet what the useful applications might be.

My biggest “oh shit” moment came with ChatGPT in late 2022. I ran many tests then to see if it could, for example, identify the antecedents of pronouns and the meanings of polysemous words based on the context. It did really well. Last December, I gave a talk at a university in Tokyo in which I showed some examples done with the 2022-era GPT-3.5. They appear in slides 4 to 8 of the following:

https://www.gally.net/miscellaneous/20251206_Gally_ICU_slide...

There have been a lot of “oh shit” moments for me since, especially after the release of reasoning models and, now, long-running agents.

A coworker had me work through a particular problem (some no-importance web demo) with Cursor and Sonnet 4.6. It still sucked, but there was a qualitative shift in suckiness, one that I realized could finally be used to solve some real problems I had if I wrote an appropriate harness and used good enough models.

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.

I had bought some Anthropic credit and waited a year to use it. The week before their expiration I fired up Code and spent $3 the first day and the remaining $22 the next day.

Putting a ReAct loop with tool calls in my terminal wad and is the biggest a-ha since I learned to make compilers, and before that, how to code.

Fixed a nasty bug in one of my tests where a mock in a completely different test I had never worked on was incorrectly setup and intercepting my mocks, I don't think I would have found it ever because the amount of effort it would have taken means I would have needed to move on to some other way to test.

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.

My most recent one: Taking a bricked ipad and plugging it into my linux laptop, then telling deepseek to fix it. A couple of hours and twenty sudo passwords later it was working again.

I wasn't impressed by the LLMs up until January or so when Claude Code swooped in. Until then, I felt like the LLMs were slowing me down. I have been using them for a couple of years now for coding at work, but I never really thought they brought in real value. Then in February I worked on a 1-month-ish project timeline and shrunk it to 3 days and that was it. I didn't write a single line of code in that project and I went all in with Claude Code. That was it, _the moment_ of realization. I was thoroughly impressed. I went from nothing to a tool that served several teams. Now I'm starting to see the cracks in LLMs and I'm slowly getting back to picking which task to offload to AI and which ones to do by myself.

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.

I am, admittedly, word oriented so my moment may be a little different from others. I asked llm to estimate my political orientation and belief system from my stylometric footprint. It got very close to unnerving and that was with me carefully removing pieces I thought were problematic.

Running ComfyUI and some ImageGenAI and realising how you can use it to generate anything from any aspect of pr0n and various fetishes to making up fake news about basically anything. And real enough to convince the masses.

definitely DALL-E image generation for me

Recently purchased an 100 year old home. it was dead in the middle of winter and the house has steam heating which wasnt working. a few screenshots and chatgpt gave me a step by step of which levers to pull and knobs to turn. this was terrifying considering i knew nothing about these systems. it worked!

When we had to have a frank discussion about whether to fail someone who obviously used an LLM for parts their dissertation.

well?

Pretty much immediately after I asked the LLM to perform a complete code review of my projects. I've been programming alone for years, that alone was life changing for me. It only got more impressive from there.

Opus 4.5 fixed so many issues with my self-coded research projects, and allowed me to port between tensorflow and Pytorch in a much shorter time than manually. Helped a lot with docs too.

I can count 2:

Dec 2025: We use a commercial 3D modeling software to build refinery. There was no license dashboard in this ancient piece of junk. Fortunately license server provided verbose live status report through a command line. I ask ChatGPT to ingest the logs into a Django web application and generate weekly/monthly/yearly usage dashboard, and It one shorted the whole Backend + Frontend in 4 to 5 shot. There were around 10 regexes just in the log parsing batch script. I was totally speechless. Encouraged by the success of, I went ahead and made the dashboard for 3 more software in the same Django app. Released to peers by evening, feedback incorporated in 2 days to integrate Name, Employee Number, IP Address sync etc in 2 days. And it’s been live for 5 months, actively being used by all coadmins, even management has it bookmarked, to help with department redistribution. Making this thing without AI would have taken well over a month of “learning new stuff”, or paying external consultants too much. Even head of IT replied back, it was awesome. ;)

2nd , June 2026: I asked codex to something fairly complex before going to morning bath!, which would have taken me more than a week of learning DirectX12 API nuances and such things, 20 min latter, I return to task exactly completed with code changes in 5 different files. Build complete without any error. OMG. Free Quota over for whole month! I subscribed by the evening.

I gave GPT-4 some source code and my existing tests, and asked it to write a new test, and it did it! It didn’t even run straight away, I had to fix it, but it still blew my mind.

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.

There wasn't a specific moment, but I started trying to debug code and deal with general tech error messages. Suddenly something that could take hours turned into a fairly quick back and forth, fairly reliably. Not all the time, but often enough to be a straightforward timesaver.

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.

When deepseek found a fix for a bug I couldn't find in minutes.

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.

For me it was gradual, then sudden.

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.

2025 xmas day, was at my wife's parents' house in rural Japan, my kids were all playing with their cousins, I was posted up with my laptop just listening to some podcast about the benefits of making time for long walks in middle age (as if! ~lol) while running another "agentic team" experiment — 12 agents in parallel.

I'd been feeding these bots a few projects, over and over — the hard part was the feeding them — that is, giving them enough well-defined work to do. They weren't yet good enough to write real software you could keep — at least I'd never seen that — and my experiments were just about finding the edges, building my intuition, and playing with processes that might be useful someday.

These things had built my kids' weird magical-dominoes games a few times by that point — but the experiment had been repeated so many times that you could argue we had "written" that software in English, with a spec that had been built, reworked, and rebuilt many times.

But this time, the bots were building me a bespoke git client, unlike any other, and unlike anything I would take the time to write — waaaay to complicated, with too little benefit. I wanted it, but only for this one niche use case.

It was a GUI client to manage a collection of repos, about 200 of them in a monorepo where every subproject was a git submodule , which are the universal counterpart to node_modules — while the latter is notorious for being "the heaviest object in the universe", git submodules are widely acknowledged to be the most annoying objects in the universe.

Nevertheless, I had this weird monorepo, and I wanted to visualize and do stuff to this list of independent repos that were also git submodules of the parent monorepo: sort by outstanding commits, divergence from upstream, recency of activity, etc. Visualize them differently based on these things. Search across them, including the source code on branches other than the current one. Show the branch counts and number of branches and commits that existed locally but not pushed upstream. A bunch more boring stuff like that, but done across the full set of repos.

That project itself wasn't even interesting to me; that software would be marginally useful to me if it existed and worked, but the main point it was just a large enough chunk of work to keep a team of bots busy all day without a human in the loop.

In December 2025, AI coding agents were already useful with a human in the loop. Opinions varied a lot about how useful they were, but to me it was obvious we were going to use them for the rest of our careers as software engineers.

It was not yet obvious that we were going to let them write huge swaths of code, or entire programs, without any humans in the loop. I had never seen that produce something that worked well enough to be worth keeping.

And then, that day, I did. I had structured the workflow so that the git client was on the screen and auto-refreshing. I was listening to the podcast, drinking coffee, reading the news. The git client was a crude window with a table in the background, a single column showing the full path to each repo, and nothing else.

Then the table expanded. It got color coded numbers representing the commit/branch counts. It suddenly gained styles, and looked nice. A contextual menu started popping up, repeatedly, and grew to include several more menu items over the next few minutes. New confirmation dialogs popped up as the bots implemented and exercised the various features from my spec.

I remember my field of vision narrowing as I started to focus on what the bots were doing. They were just executing my loop — one bot would implement one bullet from my spec, another bot would review the code while another bot manually tested it, and tried to break it, run a code review gauntlet in a loop until there were no more findings, repeat.

I could see the progress play out on my screen as they worked. I had watched bot teams work before, but it had always been pretty janky, and something like a bad game that nobody would play, or a stupid to-do-list app, or — more often — something that didn't actually work.

This was the first time I had ever seen it work. This was the grail we'd been looking for, not sure if it really existed: a fleet of bots successfully building a piece of complex, useful software without human assistance. I could tell it was working, because the adversarial testing and usability checks were all happening right before my eyes.

So it _is_ possible, I thought to myself.

They did it all morning. The app worked. I used it every day after that, for several weeks, until I finally got that entire monorepo converted to a more sensible git subtree-based arrangement.

In the half year since then I've been in a kind of manic state some of my friends call cyberpsychosis, chasing that dream. I've now seen agentic fleets successfully build many things. I've also seen a bunch of failures, some subtle, some catastrophic and hilarious. I'm still building my intuition, and the laws of physics in this universe are mutating every few weeks. It's wild.

I am fortunate enough to work at a place that doesn't pressure engineers to climb a token leaderboard, or to use AI beyond what we deem prudent. This kind of agentic no-humans-in-the-loop coding is prohibited. The policy is that in this era where we all generate more code than ever, even by hand, it's the quality bar that must go up, not the speed of production.

That's awesome because it keeps me grounded in the old ways, and confines my cyberpsychosis to my weekends and evenings. I usually spend the weekend building up a couple software plans, honing them as best I can, and then unleashing the clankers Sunday night.

I'll let them run all week, sometimes giving them a poke or flipping them over a couple time in the evening, and then the next Saturday morning, I see what I've got. What I'm mainly interested in is: How can agentic fleet-coding processes evolve to produce better software and require less human interaction and inspection? And the corollary: How can software architectures evolve to safely consume more of this fundamentally untrustable code?

It's thrilling. Exhilarating. The near-infinite subsidized tokens are about to finally run out this month, alas. But for the past 6 months it's easily the best $400/month I have ever spent. :)

I wanted to add gapless playback to an audio archive website I maintain. I tried myself before any of the popular LLMs were available. I failed. I then tried with the first LLMs that came out. They failed. Then, when the first Claude Opus was released, it succeeded. I now have gapless playback.

Not sure that I've had it yet, although hypothetically I'm sure it would probably be something similar to the examples of writing new software for old hardware mentioned ITT. The idea of resurrecting useful but unsupported gadgets that would otherwise become e-waste is something I've always found compelling.

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

First time using Claude Code I was rather impressed by how quickly I was able to build out a website with Vue and Supabase. Cool. So.......I always wanted to create a iOS app but knew nothing about Objective C or Swift or XCode. "I wonder if Claude Code can build a iOS app for me?".

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.

I was never dismissive, it always seemed pretty cool at each step

Maybe in 2024 I was amazed to see it one shot unique snippets of code

I've had a few.

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.

The fact that it completely autonomously read in a 5 MB firmware image of an old piece of test equipment and generated a Python script to generate license keys:

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

I had a locally hosted model write its own semantic search system that indexed 250,000 documentation and code files and then write a fully functioning mod for one of the games I play based on that documentation that I couldn't get to work after 2 weeks of my own effort, all in under 4 hours (and that included a 25 minute long indexing process). This freaked me out enough that I then had it write a CLI based activity and TODO tracker and then integrate that tool into its coding process to track all of its activities in about another 2 hours. I am still emotionally recovering from this day. I have since replaced the semantic search system with an open source option (though I used it for a few months) but I still use the activity tracker for both coding projects and myself.

What mod did you build?

A mod that fixed a bug that prevented certain buffs from working when mounted for the Magus class / Arcane Rider archetype in Pathfinder Wrath of the Righteous. It also managed to fix the problem with Shelters not providing protection from corruption when resting in outposts in that same mod. I've used other models to expand the mod to an entire mini-expansion with new Archetypes and abilities since then.

They went from "marginally more work to deal with than to do it all myself" to the reverse with Sonnet and now they are "moderately less work to deal with than to do it all myself"

I gave chatgpt 3.5 the type signature for a co-algebraic encoding of a mealy machine:

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

And it gave a really impressive analysis.

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

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

It got completely confused and generated a bunch of non-sense. It was at that moment I realized that LLMs don't really understand anything.

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

In high school math class our teacher swapped out all the symbols in the epsilon delta definition of limits, and asked us what this equation expresses, and many students struggled to interpret it.

I don't think this test shows that an LLM doesn't "understand". It shows more that it has similar failure modes as humans.

Well first of all I think there is more implicit data encoded in the symbols of the epsilon delta definition of limits. In the Mealy example they really just labels for arbitrary sets. The LLM actually failed a much simpler relabeling exercise. Setting that aside, I still think the analogy is flawed.

The student is mid learning process and its entirely reasonable for them one to be relying on pattern recognition until they have fully internalized the subject. The model is fully trained and should thus have internalized their understanding of the subject.

Additionally the student can update their understanding when pattern recognition fails. The model is fully cooked and will never do more then pattern recognition.

I was tasked to rewrite an Oracle Apex webapp. 70k lines of PL/SQL. I asked Claude Sonnet 4.6 to read it all and boil it down to markdown file with business requirements. Took about 15-20 minutes, and I got a 700 lines long markdown file to guide me during the rewrite. I've since had great joy using /grill-with-docs!

It was about two days after Google released Deep Dream, if you remember, the thing that took a video and filled it with fleeting hallucinations of mostly puppies, fish heads and lizards. I was suddenly struck by the realization "oh shit, this is much more boring and samey than it first appeared to be", and all subsequent gen AI has been similarly underwhelming.

A friend had the power supply die on his high-end turntable. He took a picture of each side of the supply's PCB, handed it to Claude, and it gave him back a schematic.

I mean even assuming that this was strictly a 2-layer board, you can still route traces underneath parts like ICs, connectors, etc.. I could believe it was a simple board (for a phonograph and all), but I'd be interested in seeing how well it actually matched. Did he get a new board fabbed and it just worked?

it would be really interesting when that moment was at probably OpenAI when they realized that this was doing more than next word prediction but signs of <you name it>

Automating my email inbox, I just wanted to split them into folders according to the attachment name but the fields were often incomplete and ended up missing rules, and imap fetch was taking forever and kept failing. In frustration I decided to turn to ChatGPT to split them by messageid which I had never bothered with because the strings were too long to be useful. I initially intended to build a text list of messages and fetch them all one by one but I ended up making chatgpt crush all the instructions into one gigantic python dictionary using the messageid as keys and using it to generate a single pipelined imap call with success flags, dynamic folder naming, cleanup steps the whole works. I was just working on theory of what I knew was possible, and it's the ugliest table you ever saw, but it works and it runs from memory instead of reading and writing values to a temp file and I'd never been able to keep up with that level of nesting before

I had it fill out all the forms to appeal my property tax value. We created an assessment of what my San Francisco property should be worth using deep research. The city agreed and a $12k check arrived shortly after.

There were two:

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."

OpenAI already had GPT prior to the ChatGPT launch, and I had not really taken it seriously. But on November 30, 2022 when ChatGPT came out and was immediately popular, I reevaluated it.

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.

GPT4, when it could do a translation that would take a considerable human effort, vide "Genesis 1 but every word begins with 'A'": https://p.migdal.pl/blog/2023/05/genesis-az-by-gpt/

I worked in an AI (or well ML) consultancy before the ChatGPT moment. I remember we had a project where we had to extract a large sum of documents (country wide, terrabytes of pdfs of scans). We had to set up a pipeline that looked a bit like 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'm pretty sure that "a ChatGPT call" will happily add or fudge stuff in your scanned PDFs. That sounds like a massive liability.

And now you can do all of that locally with qwen3.6:35b.

> that you went from those quaint, dismissive observations to a slightly panicked, "Uh Oh" realization of what these models can do?

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.

I think my favorite early story was when OpenAI launched deep research. I was going to an event that I was headlining, and I gave it a CSV of the attendees and asked it to give me a small background on each company they represented.

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

OMG I love this!

I was at an industry event this week. a CEO of a startup took the big board of vendors who are present, put it through an LLM. It summarised the companies he should be looking at discuss partnership opportunities with and why based on his business. Spot on.

I was working on a project for 2 years with about 5 engineers. It was many years before AI. It was new subject for our team, and we were pretty sure it was possible. Turned out it was not.

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...

Curious, what was the project?

Honestly? Probably all the way back to when Nick Walton used the computers at his university to train a custom version of GPT-2 that let players experience a completely open-ended text adventure game in 2019.

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

I gave it a weird and convoluted code snippet, and asked an LLM to step through the execution and trace the value of the variables at each step.

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

I had ChatGPT write up a Zillow description for my house in the style of Carrie Bradshaw from “Sex and the City” to impress my wife.

It was unlike anything I had ever experienced.

My wife was unimpressed lol.

This was 2022.

I started to look at LLMs not as writing code, but rather as predicting what code it would expect someone to write given the context.

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.

When chatgpt 3 came out the first thing I asked was a question like "If I put my cat in a box, put that box in a crate, move that crate to a truck, and drive the truck across Canada non stop, when I arrive on the west coast, will my cat be happy?"

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.

I suggested to a masters' student that a problem we were working on would benefit from analyzing it mathematically. He brought an incorrect solution the next time we met, and on a whim, I asked Gemini to do it. Gemini got it right. I started looking for more ways to use it after that.

Maybe when I found out you can use it to run terminal commands, spin up and take down dev environments, and even run other LLMs. Suddenly 90% of the difficulty of onboarding to new repos disappeared overnight and a lot of heavily CLI-based workflows became trivial to automate. Never again do I want to spend hours manually sorting out Python dependencies.

That it could create mugshots of myself better than I could have managed to take.

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

Pre-GenAI I wrote a new interview question for a role on our team. As far as I know, the question was never made public. The interview required implementing a pretty basic CSS-in-JS utility in vanilla javascript. We instructed the candidate read the MDN documentation for the CSSStyleSheet interface, and then gave them a public API to implement. Passing implementations usually consisted of a ~10 line for loop, and was really just a test of whether a developer pick up and work with new libraries on the fly. Still, the interview probably had a 30% pass rate.

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.

I programmed data export to some xml over a couple of days. Sending xml results via email to an accounting firm for verification. A day after I finished my disk crashed and I lost all my code. Fed Claude with xml from my mail and... oh shit! ... got "my" code back. (And immediately paid for Claude subscription) :-)

I couldn’t make a Rockbox (the alternative iPod OS) simulator run on my MacBook M2 no matter how many guides I followed, then I fired up Claude code and by modifying the original source code it made the simulator run and I was able to start developing custom plugins for my iPod. It honestly felt great since I only have basic C knowledge.

It was when I was using an early version of GitHub Copilot. At first the completions were almost useless and had a kind of copy and paste feel, however one day it managed to reason thorough a complicated loop body much faster than I could have figured it out. It was at that moment I realised this AI thing was going to be big.

My ducted gas heater wasn't working where I live and I took a photo of the wiring diagram and had Claude step me through troubleshooting it with a multi-meter, and got it fixed.

We had a company hackathon in the fall of 2023. One of the teams did a project where the pulled a bunch of expense data out of the DB, shoved it into a prompt, and asked ChatGPT to summarize the expenses and give recommendations. They then treated the output as if it were factual, without validating any of the results, and talked about turning it into a customer product.

That was my oh shit moment. As in "oh shit, they think this random text generator can reason and think."

That was pretty much the writing on the wall for me.

Opus 4.6. My standard battery of questions included solving an ascii maze (20x20 grid) without using a script, using only "thinking" as a tool. It was the first model to be able to solve it. It was the first model that really appeared to be able to reason spatially.

MidJourney public discord channel.

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

Had an issue in a project where multiple media files with the same/similar names were colliding. After spending hours with chat gpt wrangling python scripts to try and sort it out programmatically, I shifted gears and built a web tool that would allow me to manually review the content and select the correct media file to associate with it in about 5 minutes, allowing me to comb through and finally fix the issue & verify the content was correct in about an hour. It made me realize I needed to completely re-think how I set about solving problems now that I have an entirely different set of tools to develop- that has been the biggest "Oh shit" moment for me, looking into the mirror and recognizing how AI will re-shape me as a developer.

When I realized that an LLM can process all the traffic in Slack that overwhelms me daily and give me a manageable digest. How long until they intermediate most of our social interactions? Sooner than we can possibly adapt, I think.

Many people got something of a head start adapting though? Seems like it's been the proposition from "social" "media" companies since 2004 or so to stop talking to friends, talking to their computers instead and consuming the half-digest of friend's transmissions mixed with ads/psyops coming in?

If you social interactions can be mediated by a chatbot I implore you to find better social interaction