enjo

The 5% of the codebase the agent shouldn’t touch unsupervised.

This jumped off the page to me and it's an assumption that underlies everything else in the article. For the record I agree with the author that if it is true that 95% of the codebase should be unsupervised, the how layer does indeed become very small.

But my experience has given me little reason to believe that it's true. Code doesn't conveniently break in the 5% you choose to care about. Systems fail in seemingly random places very often in the seemingly uninteresting parts of the system. The macro and micro decisions that make code actually work have to be of high quality everywhere. They have to be fixable in the worst case by a human and that still very much requires a human to review nearly everything the agents generate.

As always there's a risk-management question here. You can choose to take on a lot of risk by increasing the amount of code that only agents touch. And maybe I am still too in the oughts with a belief that code should be secure, do what it says, and do it reliably. Turning over your code to unsupervised agents in such large quantities amps that risk level up to a place that I (at least) am not comfortable with it.

So ya I think the author is right in their model. Why, what, how is a great way to think about orgs. I agree that the output of the how layer has been pretty turbocharged and I agree that the how layer is going to shrink but I think I disagree with them by how much.

My wife’s grades are due tomorrow. She was in the middle of finishing exams when it happened. She can’t even access the exams to grade by hand. Total mess.

AWS was gutted by layoffs over the last couple of years. None of this is surprising.

Yep my experience has been pretty bad. As in Claude with Opus can rarely produce even compiling code in my particular project (a year old, mid-complexity one). This is with adhering to best practices including a robust Claude.md and detailed PRD's.

This is more or less my experience with Go right now.

For a bunch of reasons I want to avoid the standard React, Typescript, and Node stack but the sheer velocity that might enable from the LLM side might make it worth it.

Two axis globe is best globe.

And to be clear: we fund our schools at a higher rate than basically any other country in the world. We are fifth in the world in per-pupil student funding behind only Luxembourg, Norway, Austria, and South Korea.

https://nces.ed.gov/programs/coe/indicator/cmd/education-exp...

It's Scala all over again.

I mean there are many distinct styles of American pizza. And Italy itself is home to several different distinct styles of pizza. Is Pizza al Taglio "authentic"? What about Sicilian Pizza? Is Pizza Fritta "authentic"?

I do have to wonder whether it's really worth it to go through all this rigmarole to get the official Neapolitan pizza seal of approval.

In my neighborhood in Denver I have several Neapolitan pizza options. One is fully Vero certified. I go there often because it truly does give me the taste of Napoli from my childhood.

The best option, however, is recognizable as Neapolitan pizza but is much better than anything I have ever had in Italy (at least to my taste buds).

I feel like both really have their place and both restaurants have been around for quite awhile at this point so it seems to work for both.

That said I do appreciate the places that have are fully Vero certified because when I travel if I see that sticker in the window I know I am gonna have a pretty good pizza.

Yep but this time is different because it's going to completely displace what is left of the middle class and this time the people being fucked have six figures of college debt.

TP-Link comes with a healthy dose of worry about the Chinese governments surveillance practices tho.

I feel like we'll get a good feel for this if hiring domestic engineers picks back up without an influx of foreign folks who are not receiving the positive tax treatment.

100% agree. I have so much trouble squaring my experience with the hype and the grandparent post here.

The types of tasks I have been putting Claude Code to work on are iterative changes on a medium complexity code base. I have an extensive Claude.md. I write detailed PRDs. I use planning mode to plan the implementation with Claude. After a bunch of iteration I end up with nicely detailed checklists that take quite a lot of time to develop but look like a decent plan for implementation. I turn Claude (Opus) loose and religiously babysit it as it goes through the implementation.

Less than 50% of the time I end up with something that compiles. Despite spending hundreds of thousands of tokens while Claude desperately throws stuff against the wall trying to make it work.

I end up spending as much time as it would have taken just to write it to get through this process AND then do a meticulous line by line review where I typically find quite a lot to fix. I really can't form a strong opinion about the efficiency of this whole thing. It's possible this is faster. It's possible that it's not. It's definitely very high variance.

I am getting better at pattern matching on things AI will do competently. But it's not a long list and it's not much of the work I actually do in a day. Really the biggest benefit is that I end up with better documentation because I generated all of that to try and make the whole thing actually work in the first place.

Either I am doing something wrong, the work that AI excels at looks very different than mine, or people are just lying.

My experiences have been universally negative often very much so. I have given LSD a good go. It has led to intense hallucinations with very long lasting PTSD like consequences for me. I have done it under the guidance of "professionals" (as close as you can get in a world where these substances are completely unregulated). Even in very small doses I have experienced intense anxiety and general feelings of dread.

This isn't to discount your experience but rather a general warning: all drugs aren't for everyone. It's easy to take away from these threads that psychedelics are universally positive and that negative trips are generally the result of misuse.

Which isn't true. Before going into this doing some deep introspection about yourself and your abilities is really important. Use these drugs with extreme caution.

Just a data point:

I think it has a lot to do with the type of work you are doing. I am a couple of years into a very small startup that has some actual technology built (as opposed to a really simple CRUD app or something).

When I am working on the front-end where things are pretty simple AI is a huge speed up. What it does VERY well it latch on to patterns and then apply those patterns to other things. If it has a couple of examples you can point it to and say "ok build that but over here" the newest revisions of Claude and Gemini are perfectly capable of building the whole thing end to end. Because it's a fairly repetitive task I don't have to spend much time untangling it. I can review it and pattern match against things that don't look right and then dive into those.

For a real example, I needed a page for a user to manually add a vendor in our platform. A simple prompt asking Claude to add a button to the page sent into a mode where it added the button, built the backend handler, added the security checks, defined a form, built another handler to handle the submitted data, and added it to the database. It even wrote the ACL correctly. The errors it introduced were largely around using vanilla HTML in place of our standard components and some small issues with how it attempted to write to the DB using our DB library. This saved me a couple of hours of typing.

Additionally if I need to refactor something AI is a godsend. Just today an underlying query builder completely changed its API and broke..everything. Once I identified how I wanted to handle the changes and wrote some utilities I was able to have Claude just find everything everywhere and make those same changes. It did it with like 90% accuracy. Once again that saved me a couple of hours.

Where it fails, usually spectacularly, is when we get to the stuff that is new or really complex. If it doesn't have patterns to latch onto it tries to invent them itself and the code is garbage. Rarely does it work. Attempting to vibe code it with increasingly more pointed prompts will often result in compiling code but almost never will it do the thing I actually wanted.

In these contexts it's usefulness is mostly things like "write a sql query to do X" which occasionally surfaces a technique I hadn't thought about.

So my experience is pretty mixed. I am definitely saving time. Most of it is typing time not thinking time. Which is like 1/3 of my average day. If I had to guess I am somewhere in the neighborhood of 30-40% faster today than I was in 2019. Notably that speed up has allowed me to really stretch this funding round as we are well past the phase where we would have typically hired people in my past companies. Usually someone relatively mid-level to take over those repetitive tasks.

Instead it's just me and a non-technical founder going along super quickly. We will likely be at a seed round before anyone new comes in.

As an occasional host we always wait until the guest writes a review. Our goal is to basically never have a guest write a review so we don't write them hoping they just forget. If we write one they get an email telling them that we did and really pushing them to review.

Our place is all five star reviews and there is very little benefit for further five star reviews. So it's kind of all risk for us at this point when someone does review.

Maybe. But I often think about my grandparents. Both lived to be 96. Neither had significant health issues until the day they died.

My grandmother was extremely active throughout her life. She worked out. She was active at church. She volunteered throughout her 70's and 80's. She was always happy to go out to dinner with anyone that wanted.

Most of my memories of my grandfather (her husband) where him sitting on a recliner watching TV. He rarely left the house.

Their quality of life in the last 15 years of their lives was so different. Grandma died on her feet. She had a heart attack in the middle of the night and that was it. The day before she had been out with friends and had dinner with the family. My grandfather just sort of wasted away. He was weak, frail, and kind of miserable.

All I'm saying is there sure seemed to be a lot more than dodging all of those diseases and issues with those two. My grandmother invested in herself. She worked really hard to maintain that quality of life.

She's my role model as I age.

it's literally "let me paste the assignment into ChatGPT and see what it spits out, change a few words and submit that".

My wife is an accounting professor. For many years her battle was with students using Chegg and the like. They would submit roughly correct answers but because she would rotate the underlying numbers they would always be wrong in a provably cheating way. This made up 5-8% of her students.

Now she receives a parade of absolutely insane answers to questions from a much larger proportion of her students (she is working on some research around this but it's definitely more than 30%). When she asks students to recreate how they got to these pretty wild answers they never have any ability to articulate what happened. They are simply throwing her questions at LLMs and submitting the output. It's not great.

I think there’s a stronger argument that those polls were intended to suppress Trump voters into thinking they were not the majority (when in fact they were).

Show your work. What evidence do you have to support that argument?

I live pretty close to a couple of bald eagles in central Denver. I have seen one of them mixing it up with the other birds to get table scraps left behind by people using the park. I have no idea what they are eating generally but sometimes they are pretty happy to just grab some bread on the ground or whatever.

That isn't really what happened in Singapore.

I would not really classify them as "bad" actors, but there are definitely real research lines into this. This freakonomics podcast (https://freakonomics.com/podcast/how-to-poison-an-a-i-machin...) is a pretty good interview with Ben Zhao at the University of Chicago. He runs a lab that is attempting to figure out how to trip up model training when copyrighted material is being used.

When I worked for an American subsidiary of a Chinese company (Video Games) we were only required to honor censorship requests for Chinese users.

They had plenty of time to figure this out.

And you’ll learn that the speed of typing, even long method names, is not the limiting factor of your productivity.

This is the fundamental truth that those pushing AI as a replacement for programmers miss (intentionally or not).

One idea is merely a form of control flow (like 'if' or 'goto') that could jump down the call stack.

Interestingly this is more or less how Symbian handled exceptions back in the day. It was a giant mess. You basically called setjmp in a surrounding function (using a macro IIRC) and then longjmp (in the form of user::leave()) which unwound the stack without calling any destructors.

It was very fast and the source of endless memory leaks.

In addition I would need to some justification for the idea that water stored in a clean tank without access to light is somehow worse than fresh out of a filter.

If you're ever in Denver the Denver Elks Lodge has a really nice Crokinole board and visitors are very welcome!

I'm curious if Gabe over at penny-arcade would be upset if his comics were used in an art class to train a new generation of artists.