Yep. In my experience, LLMs easily go in circles with even simple assembly, creating fixes that result in 2x slower code and then fixing those with even slower code. They are pretty great as a reference or finding needle in the haystack bugs though!
Is that with or without the LLM being able to assemble + profile its hypotheses itself? Because “optimize this code to execute in fewer CPU cycles under the test harness” seems like one of those perfect self-contained problems for LLMs (their equivalent of an “embarrassingly parallel problem”: an “embarrassingly-easily-explored solution space with embarrassingly-easily-measured objective success criteria.”)
Sure, they might make all sorts of dumb hypotheses at first, but as long as the results of those stay in their context, they do seem to eventually “run out of ways to be stupid.” (Which is to say, LLMs seem to experience in-context learning even via self-directed trial-and-error, if given a sufficiently-large number of iterations and no way to cheat.)
I was able to use Claude to translate a few thousand lines of assembler from one syntax+toolchain to another. Process was definitely iterative; had to keep adding a few rules along the lines of "don't do this...here's the equivalent pattern." But in the end it did a good job and saved me a ton of time. Allowed me to move a ton of unit tests off a hand rolled broken+bad assembler to a modern production toolchain.
This is definitely not my experience with GLM-5.2. It is writing pretty good ARM SIMD code.
It also is an excellent radare2/rizin/ghidra driver as well. Maybe it's because its ~cyber~ capabilities (pretty much linked with assembly-level knowledge) aren't guardrailed off?
Sometimes I worry about the incentives for innovation in the US.
Step 1, find something to innovate on, sell the promise of it to investors.
Step 2, build a prototype or worst case, build it for real and start generating income from your truly innovate and unique product.
Step 3, get acquired by a large company and then shut down because your product competed with theirs.
End result, general public possibly benefited from your innovation, but in the long run, it was temporary.
Maybe the incentives would be better if it were harder for large companies to acquire small ones? If the path to riches where driven primarily by delivering value to customers. Would love to hear other's opinions on this.
This is what I struggled with after grad school. You have so many decent ideas. So many people you know in the domain also with their own good ideas. Everyone looking for their next gig. It could all be so easy to get a couple people together and start building. But, alas, money, that must come from people who expect more money back before long, which severely limits the scope of ideas that will get investor funding. No moonshots, no sci fi future, just same old looking for low hanging arbitrage or rent seeking opportunities. Kinda sad when you realize that is pretty much all private tech investment because funding is driven by investors ultimately, and not scientists like what you see on grant review panels for basic research. Money must make more money, which again limits severely what money can do.
YC is even more limited in terms of ideas that get funding. I scrolled through the catalog the other day and everything on there is chasing the AI dragon.
I'm fairly convinced that the U.S. has chosen the worst models possible to innovate.
Giant corporations control everything, even government laws, regulations, and policies. They will buy up any competition, patent themselves toward a moat, squash competition they don't want to buy, etc.
Venture captialism doesn't care about true value or some actual addition to society. It's a giant grift just to make more money from money. They chase every trend and hype train ad nauseum. It was self-driving cars, then cryptocurrencies/blockchain, and now gen AI. The vast majority of these companies have no value to society or long-term innovation.
The government invests in areas, but a huge amount of it goes to the black hole of defense contractors. And academic institutions in the U.S. are incredibly wasteful with money and spend all their time trying to fundraise money at the same time.
Just the entire system is inefficient and effectively broken.
For example, YC announced calls for climate tech a few years ago: https://www.ycombinator.com/blog/rfs-climatetech. Where did that go? I looked up YC companies in the climate space (https://www.ycombinator.com/companies?batch=Summer%202026&ba...), and there's only 22 companies out of the several thousand YC companies. So where's the value? Most of the companies aren't hiring and just seem like vaporware. And if you look at the leadership, almost all of them are serial VC/startup people and not actual innovators, experts, or professionals.
The real solution is to break up megacorps, tax them heavily, and also tax the wealth of ultra wealthy billionaires. This will let competition happen in a more sustainable way. Otherwise every startup ends up being acquired or copied by one of the megacorp or dies.
The EU is no way share or form in a good economic position right now. That's why euro leaders have been kowtowing to Trump despite him being a deranged lunatic.
Delete all American software, American defense, American energy, and Chinese hardware from the EU tomorrow. That's the deep-seated unease that keeps EU leaders up at night. Europe needs to be doing 3-4% GDP growth annually and have a globally competitive top to bottom tech an defense industry, and it needs that years ago.
The problem is that the EU needs to become more like the US to do this, and for people who grew up under the protective overhang of the soviet collapse, this is mostly unthinkable. Just like the US not bankrolling half of Ukraine's defense would be unthinkable...
Their data is not perfect as they rely on public sources, and some governments are more transparent than others, but the reality is that US funding all but vanished in 2025.
Back to the topic, there is also a pattern of promising European startups being bought by wealthy USA incumbent companies. This is also happening to established compagnies: see ARM, Alstom Power, etc.
As Europe de-couples from the USA in the current context, I suspect (and hope) that such acquisitions will come under more regulatory scrutiny.
> "Get bankrolled by the state at the state's discretion until they get what they want, even if they need to burn $1B to get $1M of value"
If that's how it worked, they wouldn't lead in anything, they'd be bankrupt already. They burn state money like VCs burn cash. DeepSeek, Alibaba, Tencent, Xiaomi, Huawei, etc., disprove your point.
Look into how their 5 year plans have lead to capital investment with almost zero feedback. A heavily bureaucratic system of bureaucrats incentivized to spend massively to boost their own appearance, and cover up losses/inefficiencies.
Ghost cities, empty high speed rail lines, solar cells being mass produced at a loss.
All these things also produced end products the state wanted, no doubt. But the capital allocation strategy is basically a "throw all the money the leader gives in that direction until the leader says stop".
Is there a lot of wasted capital? Sure but a lot of it still produces outcomes.
> A heavily bureaucratic system of bureaucrats incentivized to spend massively to boost their own appearance, and cover up losses/inefficiencies.
In China, if you want to move up politically, you generally need to show results, meaning the province or area you govern is expected to deliver measurable performance (even if politics and connections still matter too). In that sense, you could argue it's more performance driven in some respects than the US.
EVs and solar were clear priorities, and China has been very successful at scaling both and driving costs down. Domestic competition has been so intense (especially in EVs) that margins have gotten extremely thin, and officials have recently signaled they want to curb "irrational" price wars.
> Ghost cities
Sure, some exist, but many of the developments that were circulated online years ago have filled in over time. That said, there's no question a lot of projects stalled or collapsed during the property downturn, especially after China Evergrande and other developers ran into trouble.
> empty high speed rail lines,
I can't speak to every route, but overall the high speed rail network is heavily used. When I traveled in China, it was excellent and extremely extensive. Some lines and stations likely see weaker demand than others, but the idea that it's broadly "empty" doesn't match reality.
> solar cells being mass produced at a loss
With overcapacity and price wars, many firms have faced serious margin pressure and losses though that doesn't mean every producer is losing money on every panel.
In the end, the real question is whether the capital allocation is efficient enough for citizens to benefit and for the country to remain competitive. Empirically, the answer looks closer to yes in industry and infrastructure, while real estate has been a major exception, with real costs and inefficiencies.
Xmonad is a bit historically related to StumpWM, actually; a number of early users had been using StumpWM for the greater programmability over Ratpoison, but weren't all that keen on Common Lisp, so when dons decided to write a WM in Haskell... I was one of them, and even ported over some of the more popular extensions like search engine plugins (which were themselves originally modeled on 'surfraw', a package of shell scripts written mostly by a fellow named Julian Assange; everything is connected).
This comment reminds me of a book by Italo Calvino. Probably my favorite of his. It's a collection of stories told by Marco Polo to Kublai Khan about cities in his domain. Somewhat fantastical, but strikes deep to the heart to transitory nature of cities and perception itself.
A few dazzling lines on wanderlust from Invisible Cities:
... Futures not achieved are only branches of the past: dead branches. ...
And Marco's answer to that: "Elsewhere is a negative mirror. The traveler recognizes the little that is his, discovering the much he has not had and will never have."
It takes me lots of energy to ignore something. Like someone walking behind my computer monitor. Like someone having a loud unrelated conversation next to me. Like being conscious of my appearance and behavior (out of respect for others). And it takes energy to solve problems and write great code.
I simply can't do it all every day, day in and day out. Monday is fine. I can work long hours on Monday in the office, getting lots of good work done. But on Tuesday, I am irritable and can't focus as well. By Wednesday, I'm useless. If I go into the office, I come home exhausted, unable to even speak.
Fortunately I can work from home, but it's a heartbreaking trade. It's not as much fun (or productive) to collaborate over the internet as it is in person. I am definitely not going to accidentally end up going out to lunch or coffee with a new friend. Etc.
Comments to the tune of "suck it up", "wear headphones", are nothing new to the ears of introverts. We've been dealing with this _extroverted culture_ all our lives. Extroverts thrive on stimulation. Introverts drown in it.
I'm both a vim and emacs user. I love vim for all the reasons everyone always mentions, but emacs comes out ahead consistently basically because emacs-lisp is so much more usable to me than vimscript. I find it much easier to create or modify plugins in emacs... Don't like the (often overfeatureful) functionality of a plugin? Change it!
I'm coming from music as a profession to programming as a profession.
In music you can craft repetitive exercises to improve small aspects of your skillset, but in computer science you can't do that exactly (or sanely). However, I do find myself directing my study in a similar way. I have various high-level goals and I try to find some way to get at them on a daily basis. I use concrete projects like reading a book, doing exercises, or taking on a hobby project. It feels very similar to practicing music, but I wouldn't call it practicing. It's my practice!
Thanks for your insight, I really appreciate hearing comments from people who weren't born coding, because these often contain very interesting insights!
I would suggest that revisiting code which you have previously written, and trying to re-factor it, or at least considering how this could be done (and its impacts), is a very useful tool to help you 'practice' coding. I know that this is imperfect, but I have personally learned a lot from applying new lessons (from books, experience, or code reviews) to my own prior work.
> Thanks for your insight, I really appreciate hearing comments from people who weren't born coding, because these often contain very interesting insights!
Sites like topcoder and Project Euler are actually very good at building skills... they're much like playing arpeggios or similar exercises... they're small, may take 30 minutes to two hours to complete, and the more you do the more used to solving problems you get, and you also find out what you need to improve upon.
I agree. Once you understand typechecker error messages a little bit, refactoring becomes very simple (and fun!).
1. change a core type or function
2. fix all the type errors
3. yay
This is wonderful. I think the tradeoff is that you start a new system, you have to delay running the code for a while. Until you have your core. (Fortunately there is undefined :: a for such uses.)
I would guess you are correct: that the signal is pretty weak per individual couple (to the point of not being usable) but shows up with this nice tight graph when you aggregate data.
