pianist, organist, harpsichordist, computer programmer
Joined January 2025
- Tweets2.1K
- Following138
- Followers186
- Likes9K
Pinned Tweet
Just got the piano tuned - the first part of Couperin's La Couperin to celebrate :)
I have proper microphones now but still need some cables and to get everything set up. I'll start recording full pieces soon!
barafostus dreame retweeted
"
Thus it will be readily understood that, in order to demonstrate a theorem, it is not necessary nor even useful to know what it means. We might replace geometry by the reasoning piano imagined by Stanley Jevons; or, if we prefer, we might imagine a machine where we should put in axioms at one end and take take out theorems at the other, like that legendary machine in Chicago machine where pigs go in alive and come out transformed into hams and sausages. It is no more necassary for the mathematician than it is for these machines to know what he is doing.
"
––– Henri Poincaré, Science and Method (1908)
barafostus dreame retweeted
One day we'll dismiss programming languages that do not have dependent types just as today we dismiss programming languages which do not have sum types. I dream of that day.
barafostus dreame retweeted
The frontier LLM's are godlike research assistants but they cannot write for shit. Engage seriously with them and in good faith on complicated questions for an hour and their irreparable limitations shine through. They don't have the same retarded tics as the earlier models -- but they are still absolutely empty calories for anything but summarizing other texts. While the obvious tics have improved, there has been no improvement along this dimension: I think there may never be any.
barafostus dreame retweeted
In this new age of LLMs, it feels very much like:
1. we are seeing a dramatic increase in the amount of new code being generated
2. we are not getting an even noticeable increase in the quality of software (from an end user point of view)
How long have I yearned
To have a musician as lover!
Now the Lord has granted me my wish,
And sends me one, all pink and white.
And here he comes with gentle mien,
And bows his head and plays the violin.
barafostus dreame retweeted
The prevalence and necessity of VMs, Docker and its epigones, language-specific virtual environments, sandboxes, and VPNs is a sign of (IT) industry and community failure.
The fact that most people don't see it is a sign of cultural decay.
barafostus dreame retweeted
They don't know what they are doing. They were never good software engineers, and only know how to bodge things together even before LLMs. They don't know what the actual problems are.
This is the sad state of affairs: we're going to get crappy software faster than ever before.
>5.47 seconds vs. 10.07 seconds
When we win a lot of people are going to be sacked
We vibe coded the Epic Games Launcher to be faster than the other guy.
Don't believe us? Test it out for yourself in Preview:
launcher.epicgames.com
barafostus dreame retweeted
I'm not sure who needs to hear this, but this is how long a modern program should take to launch, and how much memory it should be using
A pattern I've settled into for the kind of core mathematical code that needs to be production-grade is discussion with the LLM, then having it proofread and numerically test what I write repeatedly. They are utterly superhuman at reasoning about and verifying this kind of code, but somehow they still cannot write it themselves without the code being ugly, overly-complicated, verbose and slow, i.e. slop.
The main thing I notice AIs do that humans don't is write lots of scripts running searches and verifications.
They don't work in a small search space, and they aren't using a gigabrain directly on a large search space. They're just writing lots of scripts to search large spaces that would be annoying to do.
When I watch them work, I mostly think, "Oh I probably should've been doing that. But I'm just so lazy, I would never crank out 1000 lines of Python doing a bunch of Grobner basis or group theory calculations just to test a hypothesis." Even if it would've been more reliable and less work overall.
Your computer runs billions of clock cycles per second, and AIs will actually use it for thinking.
I also think: If math had "industrial value", then this would've been systematized and you could train people to search for math solutions like an AI. Maybe Bel is superhuman, but lots of people could be an Astra if the computational approaches were written down and taught.
It's the difference between "stare at a paper, think, hope a pattern comes to mind" vs. "What is a good search space? Let me feed that into an SMT solver or CAS or custom C++ searcher." Of course, many people do that, but not as fluently and consistently as AI.
This soundness failure in rustc has been making the rounds for how basic an error it is. I looked into how the borrow checker works at a high level and computed by hand what I'm pretty sure is happening here, if anyone's interested.
>human-facing software will all offer dynamic UI that can be adjusted with natural language on the fly
Software user interfaces have been far too dynamic for decades and no-one at any tech company seems to get it. The idea that software UIs aren't dynamic *enough* is insane.
let me break down “software is dead” for you
the latest argument is “you just tell your agent what you want, and your agent will vibe code software on the fly to get that done, so no one will need real software any more”
sounds reasonable, until you realize it costs tokens, compute, and wall clock time to build those software on the fly
so instead of one-click launching an app, the bet is people will prefer to write a prompt, wait for 15 minutes, then get something that may or may not do what they want, and at the end of the month pay more $$$ for the token spend, higher than all their software combined
that’s literally the argument
and if we look at the world in aggregate, most people would just end up building the exact same “personal” software repeatedly
oh - your agent just spent 3 days and $500 of tokens vibe coding your terminal multiplexer? mine just did exactly the same! let’s celebrate how personal our terminal multiplexers are!
sorry - i just don’t see that as being a happy scenario for anyone, other than people who profit from telling others how to waste $ on tokens
to be constructive, here’s what i believe as the future state -
- high quality software that takes high cost to make will continue to thrive. just like good restaurants don’t compete with your air fryer and microwave
- most of those software will be offered headlessly for people to use via their agents, and they can continue to monetize such usage. this is already the case with products like jira, linear, stripe
- human-facing software will all offer dynamic UI that can be adjusted with natural language on the fly so they can be personalized easily, but their core doesn’t have to be rebuilt by every user repeatedly
instead of “software is dead”, i believe good software becomes “strong core, headless service, fluid interface”
which future do you think is more likely?
barafostus dreame retweeted
A critical look at Lean4
lesswrong.com/posts/zHhGa3iQ…
barafostus dreame retweeted
Software is solved but every other website hangs on a page that looks like this, despite flawless connections on modern network infrastructure capable of sending gigabits per second
I'm sceptical this is a good idea tbh. Support is hard, edge cases are hard. Having a contract with a SaaS (whose prices should be dropping fast if it's quick for them to develop the product too) surely beats John Vibecoder in backoffice asking Clod why everything stopped working
Talked to a finance guy today who said his firm has not only totally stopped buying software in favor of building it all in-house, they've actually hired multiple SWEs to do the buildouts, even though their salaries cost more than the subscription fees they were paying. Having tools custom-built for their specific business is just that valuable. So they're not spending less money or saving time, versus the pre-AI world, but they're getting a much higher return on the money and time they are spending. Much to ponder
I've been using GPT 5.6 Sol locally to program a game and while it's extraordinary/game-changing in its capabilities, I don't understand why its default decisions are to use such slow technologies and write such low-performance code (e.g. happily allocating in hot loops). I know what I'm doing so I give it strict instructions on technologies to use & memory management style, precisely spec out zerocopy data structures, etc. and in discussions it perfectly understands what I am going for and why. But I can't be off the ball for a moment because it will do something slow again (and my God it will take any performance penalty to avoid using unsafe).
It's a bizarre disconnect now to have the machine god that can perfectly well write fast programs but just refuses to by default for no good reason. I'm not complaining because I still have an edge on unskilled slopsters. But it seems untenable.
