I am Dr. Jekyll and Mr. Hyde about AI

I run so many side projects↑

I pay for a Claude and Codex subscription, sometimes the $20, sometimes the $100 sub. I find running both is important, since you want different models to run reviews and plans against each other.

I also shitpost on Slack that self-hosting models is not worth it, but in reality, I’m jealous of the people that snatched GPUs when prices were reasonable (and yes, I do self-host some on my 5070Ti… and it’s probably actually not worth it, at least as long as the subsidized plans exist).

With that said, here’s a list of things I’ve used LLMs for in the past few months.

Finances↑

I have (had) an Excel file that runs my financial life: Account balances, loans, transactions, stocks, 401ks, you name it, with almost 8,000 rows just in the main ledger. Had it, in various iterations, for at least 10 years.

I’ve recently replaced that with an app that actually does double-entry bookkeeping. It replaced 10+ years of jank that I did the way I did it “because Excel”. I could probably write 5,000 words just about that.

It was and is a fun project, since it’s sort of greenfield, but also has a completely different technology it must treat as gospel (“the number in my Excel is correct”… curiously a sentence that’s also the average data engineer’s worst nightmare).

It uses the SimpleFIN Bridge (which is really just a wrapper around MX), which I also learned exists thanks to AI, and is otherwise entirely TypeScript.

While I don’t have the time (or patience) to explain the entire project, my process used to be:

• Download CSV transactions from all my banks
• Run them through a hand-written python program that used pandas to normalize all schemas, set automatic categories, filter old transactions (…) into a csv…
• …that I could then copy into the Mega-Excel(tm) and figure out any discrepancies or missing data (by hand)

That took an increasing amount of time each month. It’s entirely unnecessary in the first place, but I love the visibility and ability to plan cash flow.

With this new, web-based version, it’s essentially “click sync and review all new transactions”. And the “review” part is something I explicitly asked the robot to add, since it was very confident that its auto-matching was great (and, frankly, it was).

It even benchmarked itself (in each implementation phase of the /plan output) against the original Excel file. Granted, I told it to - but it still did.

It was (and is, since I’m finding bugs, as “make no mistakes” didn’t quite cut it) a very fun project that truly replaced something I dreaded doing (but did, diligently, for 10+ years).

US civics test with decision models↑

I’ve recently filed for US citizenship (I am very excited about that) and in case you don’t know how that works, you have to take a civics test during an interview.

The test is free form, i.e. not multiple choice, but has prescriptive answers. These answers are, technically, “examples of acceptable answers”, so there’s not a 100% key for these.

To practice, I used self-hosted decision models (laya and kev-4b) to help me parse my free-form answers to simulate the real interview. Using a decision model here is nice, since it has at least a semblance of determinism. It could also add a traditional LLM as a third fallback, but I didn’t find that necessary, since most answers are pretty clear cut.

For instance, US senators serve 6 years, not 4; so typing 4 into that question tells me that a local kev-4b is 61% sure that that’s wrong.

Apps that offer this have various drawbacks I do not enjoy (such as not letting me correct a misunderstood voice-to-text answer).

My app keeps statistics about my answers via sqlite, too, so I can practice questions I messed up. Claude even parsed the PDF for me, ran benchmarks against multiple models, and basically did a lot of annoying grunt work.

A shitposting rant about some of the questions: Civil War Edition

“The Civil War had many important events. Name one.” and they list 8 answers. EIGHT‽ What do you mean they can’t just print all 900+ pages of “The Civil War Desk Reference” in the answer sheet‽

The Ken Burns documentary about the Civil War is eleven hours! Shelby Foote is a national treasure!

I live an hour from Chickamauga and I’d argue “an important event” of the Civil War is arguably “Grant drunkenly decides he hates General Rosecrans now”!

…or “How Stonewall Jackson got his name at the First Battle of Bull Run” or “How Fort Pulaski showed that rifled barrels are a cheat code” or “Oh boy did we do fucked up things at Andersonville” or “General Lee is a bit of a complex personality but let’s maybe not forget the whole slavery thing?” or “How Brigadier General (Colonel at the time) Joshua Chamberlain went fucking insane at Gettysburg (but in the best way possible)”? I have a portrait of the man outside my office, for God’s sake! I own a musket made in Harpers Ferry! I own several books about the Civil War!

Anyways… where was I? Ah, yes, the suggested/accepted answers are:

• (Battle of) Fort Sumter
• Emancipation Proclamation
• (Battle of) Vicksburg
• (Battle of) Gettysburg
• Sherman’s March
• (Surrender at) Appomattox
• (Battle of) Antietam/Sharpsburg
• Lincoln was assassinated.

And we were talking mindless computer shit that nobody will care about in a few years. Apologies. Anyways, decision models help with that. Pretty fun tech. Basically usable machine learning for everyday tasks. Yay, math. Did you want to hear about how the CSA lost a submarine? No? Oh yes, “AI”. Anyways…

So LLMs and adjacent tech (such as decision models), in my opinion, solve a problem that wasn’t really solvable before. Decision models do fall flat with some answers (such as the aforementioned Civil War rant), but if we stick to the answer key sheet, it’s an excellent use case that used to be really complicated to do via specific ML models.

Game guide summaries↑

I’ve started playing Warhammer 40,000 - Rogue Trader again (not a wise decision, as that’s a 100hr/run game) and had Claude turn some of the excellent builds and guides into a single HTML page, so I don’t have to deal with the Steam UI, which doesn’t render giant guides like this well and needs lots of scrolling.

It also keeps track of my retinue’s levels, so I can easily find where I’m at in the build. Small task, big improvement, despite being absolute throwaway code - but throwaway code that solved a real problem for me (it’s a single html file).

Book summaries↑

I listen to a lot of audiobooks while doing… frankly, most things that don’t require me to talk or think, and I often drift off and forget details. Not a winning strategy for somebody who’s partial to Sci-Fi and Fantasy. I am on book 8 of “The Wandering Inn”, which means I’ve consumed almost 3 million words… and it’s book 8 out of currently, I believe, 20…

Wordcount: Nerd edition

Yes, this is real.

Book	Word count	% into story
Book 1 - The Wandering Inn	451,580	2.6
Book 2 - Fae and Fare	502,197	5.5
Book 3 - Flowers of Esthelm	307,839	7.3
Book 4 - Winter Solstice	283,647	8.9
Book 5 - The Last Light	440,724	11.5
Book 6 - The General of Izril	276,286	13.1
Book 7 - The Rains of Liscor	287,793	14.8
Book 8 - Blood of Liscor	381,389	17
	2,931,455

17,288,383 for the entire series at the time of writing.

Source: https://innwords.pallandor.com/wordcount

I used Claude to write a thing that uses an LLM (local or via ngrok.ai) to summarize E-Books for me; if you own the E-Book alongside the audiobook, you can get summaries per chapter, per character, and even a timeline view. Uses a few different models (gemma-4-12b-qat for chapter reads) to do so and caches data.

This is especially useful if it’s not a book that’s extensively documented online - for something like “The Strange Case of Dr Jekyll and Mr Hyde”, any cheap old model with a wide training set can do it; for something new and/or obscure, not so much.

This is, arguably, another use case that wouldn’t work without LLMs.

And while it’s a bit of a big-brain solution to a smoothbrain problem (just pay attention, stupid), a lot of people have told me they have the same issue with audiobooks, so it’s not just me.

I was able to run public projects↑

SkaldMaps.com↑

I wrote about skaldmaps.com before; it’s a full product with everything that entails, from an LLC to CI/CD. I won’t get on your nerves about it, but the dev process uses AI heavily, because I couldn’t have built it otherwise. Not because I’m not capable, but because there aren’t enough hours in a day.

Now, it’s a geospatial data product and I’ve done data and data-adjacent work for many, many years (and GIS I just happen to enjoy).

AI automated the things I either don’t like (SQL) or am bad at (frontend) away, while I could focus on content, edge cases, architecture, infrastructure, and all the other things that are actually reasonably fun.

FI-Track.com & vibes↑

There’s also FI-Track, a Financial Independence, Retire Early (FIRE) tracker that can do a bunch of wild stuff locally in your browser, such as Monte Carlo simulations.

That thing is vibe coded and I gladly admit it because I wouldn’t have built it otherwise.

Now, am I capable of doing it by hand? Certainly. I wrote a Monte Carlo simulation in 2011. I remember that because that was fun. I used Java to simulate Roulette games to teach statistics to people who thought it was boring otherwise.

But I would not have done this otherwise and just suffered through 3 different, existing calculators that do all that FI-Track does and consolidated them in Excel.

So this project - which is free, you can use it right now - exists because of LLMs. And there are countless other examples of other people doing the same too, of course.

Real-life things are a lot easier now↑

I also just ask ChatGPT about things I know a reasonable amount about, but want more details on or just don’t want to wade through garbage online. It’s awesome at answering ad-hoc questions about random topics.

I’m a fairly passionate hunter and I seeded a food plot for “my” deer this year.

In case you’re not familiar with the rural life - think a farm field just for animals to monch on, especially in areas where there is not a ton of natural food, because of agricultural monoculture…

In my case, I did this using an ATV plow, ATV glyphosate sprayer, and a hand seeder. If that doesn’t mean anything to you, not exactly an ideal setup (you should use a tractor).

I only did about a half acre, but I shot a decent buck on it on opening day.

I used ChatGPT to calculate spray rates, angles, glyphosate concentration, the best seed mix, and it helped me figure out a schedule (off-grid, no water) for planting (and so on) and it was awesome. I didn’t grow up hunting or farming, so this helped a ton. This stuff is super localized (weather, soil, laws, animal populations, property sizes…), so finding any good insights is challenging to say the least.

While it’s easy to say “Codex helped me to debug my broken npm install” (or whatever other computer-y thing) and live in the tech bubble / on Twitter, this tech is also very useful for “real world” use cases. And, despite sounding like a Charlie Daniels song idea reject, being knee-deep in deer guts is a bit more “real” feeling than typing prompts into a MacBook.

It also helped me run a vacuum line on my car, helps with gardening, woodworking, and a bunch of other very “analog” tasks - yes, it hallucinates, but for me, it’s a net gain and I wouldn’t want to miss it. No, it neither pulls the trigger nor does it run the table saw, but it does help a lot (provided you have a solid base level of knowledge beforehand).

LLMs have made my solitary hobbies more enjoyable. Wild statement.

I can’t program anymore↑

I chose this heading deliberately. I can program - my first website went live in 2004, and yes, HTML counts, as does Visual Basic :-) - but I don’t get to anymore.

Maybe as a bit of background: I’m currently a Principal Software Engineer @ ngrok. While I don’t overly care about titles (I don’t think I have mine in Slack, even), the scope of my work is supposed to be pretty broad and spans various teams and projects. ngrok does networking software, globally distributed, with lots of traffic and data.

A lot of my work is architecture and long-term plans and strategies, not so much “is this a nice go interface?”. That’s true, with or without LLMs.

But you know what, some of the arguably most impactful things I’ve done at work were, essentially, nice go interfaces where I’ve spent a long time thinking through data/protobuf shapes, data contracts, libraries, and quite literal interfaces.

One of the systems I’ve architected and built before LLMs - IP Intelligence - hinges on clever composition and way too much thinking about trees (it’s an mmdb database under the hood). I’ve spent a lot of time digging into obscure topics and problems there and I enjoyed it very much - and I could tell you endless stories about similar problems (and, hopefully, the products resulting from it).

The tl;dr about that product is “every IP address going through ngrok has a ton of metadata attached, such as categories, ASNs, location and more and you, as a user, can route traffic based on it”. Internally, we can do the same and use this data for a ton of use cases.

Anyways, at the end of the day, we got a nice product that scaled the way I wanted it to and is now part of, basically, every ngrok traffic package. Pretty cool.

I don’t really do that anymore. It is, objectively speaking, not a good use of my time.

Now, let me be clear: I get to do the “strategic” part. I’m a weirdo and write my architecture noodling diagrams and documents by hand, because I like to think through them. But I don’t get to noodle on the go interface. And just to be super clear, if I decided that I wanted to, I could. There is nobody stopping me from doing it, but I do think it’s not a good use of my time.

LLMs write code well. I don’t think they write good code for humans, but I do think they write good code for other LLMs. And if we go all-in on agents, who cares what I think about code duplication or go interfaces? Objectively, saying “Claude, do this thing” is going to be faster than me (and I’m usually pretty fast).

So, from a business perspective, pretty hard to argue against.

That said, I greatly miss writing code. And no, “you can write code by hand in your free time” doesn’t count. While I did code as a hobby before, it was never what made me a better engineer. Sure, it helped - sometimes even a lot, doing stupid projects like Bridge Four is a great example - but it was really the 8 or so hours a day I got to use to hone my skills. I am now writing this in my free time, while I have an agent combing through some TypeScript on the side.

And, once again: We, as SWEs, get paid for our output, not beautifully clean lines of code.

That doesn’t mean I don’t miss it. It doesn’t feel like engineering anymore. Every time I hear somebody slap the word “engineering” on some Twitter-coded LLM term that will have a half-life of a week, I cringe.

And I wonder at which point my existing knowledge, practical and theoretical, becomes useless.

I don’t want meat proxies↑

Another thing is not of a technical, but rather a personal nature: If you send me LLM output, I will not read it. I’ve made this rule and I’ve been enjoying it. I’ll +1 your PR, but I will ignore your message. I’m blessed to work with sensible humans that don’t do this often, but I can’t say the same about life outside work.

In case you wonder why: It’s so, so, easy to detect and so deeply uncanny. No real human writes like that.

I’ll tell you a story: I recently bought a car. It’s a Miata and I’ve been told I’ve been banned from talking about it to unsuspecting civilians.

Anyways, I am a bit of a Ron Swanson when it comes to that and found one in the spec I wanted, so I reached out to a dealership at about 9pm via e-mail.

This is a real e-mail interaction. I got this response minutes after reaching out, again, way after hours.

The rest of the e-mail, more or less, summed up my request for specific packages I knew dealers could install. Also, the GT in red only comes in black leather.

Writing all that in a few minutes was impressive and totally not AI, so, naturally, I had to try:

And lo and behold, I got a response from Sophie Turner herself, just 2 minutes after I reached out:

In case you didn’t catch it, Sophie Turner doesn’t exist. It’s a (presumably LLM-assisted) lead generation tool, fully automated.

I did show up at the dealership and bought the car later and Sophie was nowhere to be found. We did joke about it, since apparently I was the first person to ever try and fudge with their crappy AI. So… it’s clearly working and I hate that.

Listen — if I send you an e-mail at 9pm, I’m doing it because I have free time at 9pm and am likely drinking a beer. I don’t want a response, even from a human - e-mail is async for a good reason.

This entire interaction added nothing. I know the Miata is a good choice, that’s why I want to buy yours. If BMW made the Z4 as a manual for less than the price of two Miatas, I’d maybe buy that (…and now they don’t make them at all. And yes, the “Handschalter” technically exists, on paper). If the Porsche 996s weren’t priced beyond anything a reasonable person would pay for an engine susceptible to a $25k bore scoring issue, I’d buy one. If S2000s weren’t turning into Pokemon cards… you catch my drift.

They were nice humans in real life though, and I intentionally did not call them out by their actual dealership name here.

The internet is dead↑

I’m not good at design. Or frontend. I don’t pretend to be, which is why my website looks the way it does.

But I got really good at recognizing Claude’s style. Please enjoy this list of 4 websites:

• https://colony-guide.com/gothic-1-remake/quests/the-water-mages-plan/
• https://www.gothic.tools/quest/the-water-mages-plan
• https://gothicremakemap.com/guides/water-mages-plan
• https://gothic1remakegame.com/quests/gothic-1-remake-quests/the-water-mages-plan/

These websites, all about the same quest in my childhood favorite game’s remake released a few months ago, Gothic Remake, are practically identical:

Granted - side-by-side, they seem different. But squint a bit and you’ll notice their design language:

• The serif fonts
• The not-quite-black background
• Tones of beige, gold, and amber
• Rounded corners for buttons
• Weird transparency
• Sometimes CSS effects that a year ago would have commanded their own Hacker News article

It’s worth noting that gold/amber tones are not a “Gothic” thing (it’s pretty dark - almost as if it’s in the name).

You also likely have seen this style before - it’s basically the “Claude default”. Remember the 40k guide from above? I didn’t tell it to pick gold buttons. Claude does that.

Even worse, their content is, largely, auto-generated horse shit.

See, the Gothic 1 Remake is a remake. The original game came out in 2001, 25 years ago. At the time, people wrote guides; people probably write guides to this day. This article, for instance, has been online since at least 2009 - unchanged, practically.

This is impressive for 2 reasons:

• At the earliest verifiable date, it would have been at least 8 years after the game came out
• It’s been online for at least 17 years

Speaking from experience, I can almost guarantee that this has been online much longer than that, because I used this website to play those games at the time myself (Gothic 3 came out in 2006, when I bought it on release day - still have the copy).

Anyways: If you played the original, you’ll know the game is hard. It doesn’t hold your hand and, because it’s a janky RPG from late 90s/early 2000s Germany, it has a ton of cheesy strategies that I’d expect to see on a website about the game.

What I’m trying to say is that there is endless content about it, but most of it is in either German, Polish, or Russian (the main demographic for these games).

None of these websites mention any of the useful stuff; some of the quest guides were outright wrong. Maybe that has changed now (I cannot be bothered to check), but at the time, relying on them was pointless. It’s slop. Auto-generated e-waste.

I use Kagi because Google is evil because it lets me actually own my search results for $10/mo. You can report slop and even re-rank your own results - I have Wikipedia at the top, for instance. Not sponsored, I just like them.

Nowadays, I block websites very, very often - it feels like 80% of websites I find are AI slop. Auto-generated content that exists because it can. And I hate that.

Of course, it also makes other websites worse; everything forces a login to stop scraping, Amazon doesn’t tell you what you ordered via e-mail anymore (to stop scraping), and don’t even get me started on AI “art”.

The speed is breakneck↑

This one is a bit more self-explanatory. Stuff is moving fast, but I actually don’t mind that. Back when I lived in Germany - mind you, it’s been almost 10 years - I despised the opposite: The slowness, the bureaucracy, the adherence to procedure, the “we don’t need that” (or, even better, “nobody needs that”), as a knee-jerk reaction to anything new. I’ve kept my love for unnecessary commas in prose, but other than that, I much prefer the fast-paced environment of American tech (and life).

That said, this speed used to be manageable and enjoyable; I didn’t have to program or run side projects on the weekend or after work (I did, for fun). I did not have to use Twitter. Reading HN, Slashdot, Lobsters, or random newsletters was optional, as were podcasts. Remember the “everybody in SF has a podcast” period?

These days, however, I feel “FOMO” - fear of missing out - basically on the daily. And yes, I still don’t have to. But I feel like I do.

Every time I open my personal portal to hell, it’s basically this:

I originally did not want to quote individuals on Twitter or elsewhere, but after scrolling for a minute while writing this, I am going to, just to share how I experience opening that fucking app:

Halven@halvenruns

This is f*cking insane. Anthropic engineer: "Prompting is basically over 95% of our engineers already run self-improving loops. The next move is graphs" In only 10 minutes, she builds her entire Claude Code setup live, starting from an empty terminal Empty Terminal → Agents… Show more

Oct 31, 2026 · View on X

slash1s@slash1sol

OpenAI co-founder Andrej Karpathy wrote the loop and just released a prompt built on it. Now Opus 5.5 finishes jobs it used to quit halfway. This is f*cking unreal... Any task runs as a graph: one check per node, a baseline before any change, and only the wins survive. Every… Show more

Nov 1, 2026 · View on X

Callan@callanxai

This is f*cking insane. Andrej Karpathy spent 8 years at OpenAI and Tesla Last week, he packed everything he learned into one free 2-hour lecture Agents → Harness → Loops → Graphs Some bootcamps charge $15K and teach you less than this You may not have 2 hours right now… Show more

Nov 1, 2026 · View on X

I did not search for “This is f*cking insane” or other keywords; these are legitimate from my home feed on 2026-11-01.

Disclaimer, I don’t know these people, I’m going to assume they’re actually wonderful human beings or maybe a golden retriever who learned to use a keyboard, whatever you find more charming. I’m using these as an example of what my feed looks like.

Again - please don’t get me wrong here. I find some of this interesting.

But every day, there is a new post with thousands of views that says something to the effect of

“Prompting is dead. 97% of our engineers now run self-evolving recursive harness loops. The next frontier is graph-native orchestration. Adapt or become a node.”

The same thing then gets posted, by different people I don’t follow, with something like “This is f*cking insane” as an opener (yes, this is a real example that I scrambled with ChatGPT to make even more insufferable).

It’s sensationalist, it changes every week (and never stays relevant for more than that), but it’s always seen as gospel. It’s basically LinkedInLunatics meets cocaine meets a broken attention span and I fucking hate it, but I (partially due to my own flaws) seemingly can’t escape it, since the 1% or so that’s actually helpful to use this tech efficiently is going to be hidden in there somewhere.

This was (and is) not the case when doing “the hot new thing” required reading a book or at least playing with a framework for a few weeks.

The cargo cult and psychosis↑

Even worse, people (and companies) blindly copy each other. I’ll pick a personal favorite:

I’m not starting an argument about the pros and cons of remote work - you can figure out where I stand pretty easily - but it feels like every tech “AI” startup funded by some Stanford trust fund kid is following the same formula of “996 in SF”.

And no, this isn’t just vibes:

And yes, I am aware that there are remote-friendly AI startups; I am merely picking on something I personally care strongly about. And while this isn’t unique to “AI”, I probably don’t have to tell you about the absolute psychosis that has taken hold of every company, from 2-person startup to gigantic mega-corporation.

This may be surprising, but personally, I do not need:

• Copilot pestering me every time I have the misfortune to use Windows
• An LLM chatbot on your website that has the most useless context imaginable
• “AI” bullshit buzzwords on your product site - believe me, it’s hard enough to figure out what the fuck you’re selling without it
• An LLM chatbot as a hard requirement for customer support

The latter is super fun because it happens everywhere and has tangible financial consequences for me.

The Gospel of Inevitability↑

The other (related) thing that drives me up the wall is this “gospel of inevitability”, for lack of a better term - the belief that this will change the world so fundamentally and that anyone who doesn’t replace their entire lives with LLM-generated slop is destined to be “left behind”. The virtues of the (objective) advantages of this technology completely overshadow any real concerns people have.

And I’m not even talking about the super zealots that connect every MCP under the sun and buy a house in a foreign country via ChatGPT (a real orange site comment); I’m talking about your everyday engineering use case here.

For instance: I think it’s fair to argue that in a world full of coding agents, PR reviews are perhaps an outdated concept. I think that is a fine conversation to have, especially since I refuse to read Claude prose and hence will either ignore or mindlessly +1 your vibe’d PR anyways.

But at the same time, this hinges on the axiomatic understanding that agents will replace every coding task in perpetuity.

For now, of course, it’s easy to argue that it does work pretty well and that LLMs can program better than most humans. Sure, your codebase now probably has the same thing in there about 17 times and anything you’ve learned about composition and abstraction went out the window, but it works. The natural counter to this is of course “it doesn’t matter, an agent with a sufficiently large context window can handle it” (and that is probably true). I need a reference to search for, an agent can just spam rg in all possible permutations to find whatever it needs.

Remember that “I replaced my finance Excel with an app”?

❯ scc .
───────────────────────────────────────────────────────────────────────────────
Language                 Files     Lines   Blanks  Comments     Code Complexity
───────────────────────────────────────────────────────────────────────────────
TypeScript                 237     56488     2484      2128    51876       9915

Forgive me if I’m skeptical whether that’s maintainable.

But I am, perhaps, a bit above-average skeptical as to how well this will work in the long run. Our existing code bases that had heavy LLM involvement probably survived it well because they were decently well-structured beforehand - and entirely “vibe coded” apps are way too new to have any real-world scars, at least in my opinion. And looking at the status pages of the big AI labs, that doesn’t exactly inspire confidence:

Maybe everything I know about coding and systems design is truly irrelevant in 90% of the cases and we can just solve agent-made issues with more agents in the future. Maybe we don’t and we’ll create the job of “slop janitor”. Maybe something in between. I don’t know.

My point here is that I don’t see either as an inevitability, but feel like I’m alone with that.