The word search for "fart" shows the tool's limits. No entry I saw actually said the word fart, but was listed as doing so -- "fart nawor" (hearts around the world irl), the penny farting (the penny farthing irl), etc.
5,305 karma · joined March 1, 2017
Economics is important. Best bang for the buck seems to be OpenAI ChatGPT 4.1 mini[6]. Does a decent job, doesn't flood my context window with useless tokens like Claude does, API works every time. Gets me out of bad spots. Can get confused, but I've been able to muddle through with it.
1: https://openrouter.ai/anthropic/claude-opus-4.1
2: https://openrouter.ai/anthropic/claude-sonnet-4
3: https://block.github.io/goose/
4: https://openrouter.ai/anthropic/claude-3.5-sonnet
This will largely fail to gain market share just like PaperGUI[1] and Polymer[2] did (much to my chagrin). UX and web developers simply feel the need for more control over the UI.
However, I do find it interesting that there is enough interest in this trend that people keep recreating tools for it over the years. I wonder why.
1: https://google.github.io/paper-gui/
2: https://polymer-library.polymer-project.org/1.0/docs/about_1...
Often I'll ask the AI to do something and it goes sideways. Sometimes it really saves me time, but many times not. I'll break down and actually type out commands or even Google what to do instead of using the AI because it's still faster.
It's true that my trainee uses the AI more because there's fewer commands in his muscle memory. But, it's still not great yet.
Further, the AI must have each one of its actions approved. I've tried fully automatic mode. It's bad.
AI is more like a lawn mower. It's self-propelling, but you still have to hold on to it, sometimes you got to turn it off and just pull the stuff out of its way or it gets stuck.
I wish I had more elaboration on this paragraph. It seems like a real change happened of which Ian took notice.
This would be a step backward for most home lab setups. MetalLB already implements BGP and has done so for a decade. Why would I want to install a network overlay just for that? Now I have to deal with a network overlay.
Of course, if you already need such an overlay because you're doing this for work, this could make sense.
We're measuring the wrong things. We need to measure cost of living instead of how empty or full a Port is. We don't know if these two things are correlated yet.
Gdp, consumer index scores, all that stuff is a measure of how much poor people are willing to spend and how hard they're working. What really matters is, can these people buy a house? Can they buy eggs?
I don't know how many of them will be all that caught up that they can't buy some disposable flip-flops on Temu anymore. We need to focus more on how hard it is to live than how hard it is to import stuff.
Yes, some volunteered to get out of poverty, bad circumstances, etc., but that doesn't mean it was the only route out. They chose to be there.
My uncle was in a bad situation, married early and had no money or degree. He joined the army and served in Desert Storm. He chose to be there. I was proud of him.
I didn't join because at the time I graduated high school (2005), they weren't taking people who took Adderall for ADHD ("daily psycoactive drug" or some such nonsense). It disqualified me from military service. That's what the recruiter told me anyway. But I would have been proud to serve.
I believe we did good in Afghanistan, fool's errand though it was. I believe a lot of lives were touched good. An entire generation of girls who would otherwise not have an education, for one. I was devastated when things didn't work out.
For this reason, I get annoyed when people say "we sent our kids to die".
My father served in the military. He does not have a high opinion of Edward Snowden. He says that loose lips sink ships. He says that these kinds of leaks cause soldiers to die.
I have no patience for men who cause soldiers to die, but after that experience I do remain skeptical of the NSA and NIST.
I also wrote a toy resource scheduler at an HTTP endpoint in Haskell[2]. Writing I/O in Haskell was a learning curve but was ultimately fine. Keeping logic separate from I/O was the easy thing to do.
> Lisp is easier to remember,
I don't feel this way. I'm always consulting the HyperSpec or googling the function names. It's the same as any other dynamically typed language, such as Python, this way to me.
> has fewer limitations and hoops you have to jump through,
Lisp as a language has incredibly powerful features find nowhere else, but there are plenty of hoops. The CLOS truly feels like a superpower. That said, there is a huge dearth of libraries. So in that sense, there's usually lots of hoops to jump through to write an app. It's just I like jumping through them because I like writing code as a hobby. So fewer limitations, more hoops (supporting libraries I feel the need to write).
> has lower “friction” between my thoughts and my program,
Unfortunately I often think in Python or Bash because those are my day job languages, so there's often friction between how I think and what I need to write. Also AI is allegedly bad at lisp due to reduced training corpus. Copilot works, sorta.
> is easily customizable,
Yup, that's its defining feature. Easy to add to the language with macros. This can be very bad, but also very good, depending on its use. It can be very worth it both to implementer and user to add to the language as part of a library if documented well and done right, or it can make code hard to read or use. It must be used with care.
> and, frankly, more fun.
This is the true reason I actually use Lisp. I don't know why. I think it's because it's really fun to write it. There are no limitations. It's super expressive. The article goes into the substitution principle, and this makes it easy to refactor. It just feels good having a REPL that makes it easy to try new ideas and a syntax that makes refactoring a piece of cake. The Lisp Discord[1] has some of the best programmers on the planet in it, all easy to talk to, with many channels spanning a wide range of programming interests. It just feels good to do lisp.
Government money represents hard earned cash of the people. The ability for the government to borrow -- its credit -- deals with its reputation as an entity that is capable of paying its debts. This reputation reflects on the people governed.
The people's money and its reputation should be respected. Money should be borrowed carefully to protect the people's reputation. Taxes should be spent wisely to protect and defend their purse. Wise expenditure is an important part of good stewardship, good government, and good politics.
My personal favorite: Who fixes the AI when it breaks?
The lawn mower replaced human work, but it needs _just enough_ direction that it's still worth money for humans to run them, auto-run lawn mowers notwithstanding.
AI will make people faster, but humans will only do work for other humans that do their share by giving value back.
I don't believe we will ever live in a world where people won't want people working for them, or doing something for them.
Maybe it's an argument for the other side though as well. The architecture of the system was designed to track people as much as possible so we could do A/B, app design, and marketing more effectively. It felt like it was the company's life blood.
I would say the law should at least make people get their architecture right when small so that when they're big it's not impossible to comply later.
One last thought: our company was small in head count but is getting much bigger right now in revenue. I've heard of small head count, billion dollar companies. What of them?
This is likely a manifestation of the bitter lesson[1], specifically this part:
> The ultimate reason for this is Moore's law, or rather its generalization of continued exponentially falling cost per unit of computation. Most AI research has been conducted as if the computation available to the agent were constant (in which case leveraging human knowledge would be one of the only ways to improve performance) but, over a slightly longer time than a typical research project [like an incremental model update], massively more computation inevitably becomes available.
(Emphasis mine.)
Since the ultimate success strategy of the scruffies[2] or proponents of search and learning strategies in AI is Moore's Law, short term gains using these strategies will be miniscule. It is over at least a five year period that their gains will be felt the most. The neats win the day in the short term, but the hare in this race will ultimately give away to the steady plod of the tortoise.
1: http://www.incompleteideas.net/IncIdeas/BitterLesson.html
2: https://en.m.wikipedia.org/wiki/Neats_and_scruffies#CITEREFM...