Maybe if we all got together and spent 100k USD we can build slopfort the AI built Fortran compiler
This is an insane demonstration of what unsupervised agentic coding can do. Although insane, there are limits to these things.
All of this is âinterestingâ until you factor in that:
1. the human brain can do all of this (and more!) with much less energy.
- Far too often, the most hyped results in the media are entirely fake.
These substitutes for human thinking and skill are directed by people who fundamentally dislike programming, and **dislike** dealing with human beings who happen to have thoughts, wishes, and goals of their own.
My list of Fortran tools has a few projects that translate Matlab to Fortran:
matlab2fmex: small translator which aims to convert numerical Matlab m-files to Fortran90 mex files, by Ben Barrowes. matlab2fmex first translates an m-file to a Fortran90 mex source file then compiles that Fortran90 file using Matlabâs mex and the local compiler.
matlab2fortran: performs some simple conversions from Matlab code to Fortran, by ebranlard
Mc2For: MATLAB to Fortran compiler, from Sable
Building on one or more of them with Claude Code or OpenAI Codex, it may be possible to translate a large fraction of Matlab code online to Fortran. There is GitHub - pyccel/pyccel: Python extension language using accelerators to translate Python with NumPy to Fortran and GitHub - t-kalinowski/quickr: R to Fortran Transpiler to translate R to Fortran. There should be tools so that all numerical code in Matlab, Python, and R is translatable to Fortran.
I dunno, this sounds pretty amazing to me. You have to remember that right now is the worst AI is going to be. It only gets better from here. December 2022 was the first time the AI chatbot could kind of make a haiku, and now people are working parallel teams of them to do months worth of human work in 2 weeks at a fraction of the cost.
I would like to see the cost comparison though
That is the biggest lie in AI, as Carl Brown discusses in this video (less than 10 min long).
It was $20k
Over nearly 2,000 Claude Code sessions across two weeks, Opus 4.6 consumed 2 billion input tokens and generated 140 million output tokens, a total cost just under $20,000.
Perhaps it can be done in other jurisdictions, but no one in America is going to build a compiler tabula rasa in 2 weeks for $20,000 USD with human software engineers.
If you want to help Fortran, take LFortran and run it on your code, it will fail somewhere, then ask Codex or Claude Code to create an MRE (minimal reproducible example), then open up an issue at LFortran. We fix MREs very quickly. The bottleneck right now is to create them. AI is actually super helpful for that.
AI is also super helpful in fixing MREs â but please stay in the loop, donât submit a PR without understanding what it does.
As to the future, there is a lower bound and an upper bound.
The lower bound is that the underlying technology will not fundamentally improve, so it will be like a steam engine: the engine didnât exponentially improve, but the harness around it did: the first machines were really slow, but the steam engines from 1940s were really fast (100 mph). So LLMs will be able to run locally (since the hardware almost for sure will improve) and the agents will improve etc. I canât currently afford it, but would love to have an agent running 24/7, just triaging bugs and fixing what it can fix and submitting PRs, and have a large team of agents being able to do what I ask them. The technology is already there, so this will almost for sure happen. An experienced developer will have to be in the loop to deliver.
The upper bound is that the technology will improve exponentially. Then of course inevitably LLMs will be better than humans at coding.
Somewhere along the line LLMs will start coding themselves, and then Skynet will happen. Weâve already seen the movie!
âslopfortâ, yet another word to chuck into the lexicon.
I remember one episode of the TV series Numb3rs, in which an AI sponsored by DARPA seemed to have committed murder. The mathematicians eventually find out that all the âAIâ does is passing the Turing test over and over (and it was the victimâs wife who actually committed the murder).
LLMs are impressive, but in essence theyâre just that (i.e., Turing test interfaces). They get some things right and some things wrong, but thatâs only a probabilistic consequence based upon the data they were trained on.
The narrative of the companies invested in the AI bubble is that by just adding more computing power the AI singularity will eventually be achieved âwhich is akin to saying that by adding more pages to a dictionary, it will eventually become sentient.
A few observations:
- For all of the work spent developing a non-executable context document, he could have written a formal spec, and then derived an implementation, providing a much higher level of assurance that the code is correct by construction (with regard to the specification) and a plan to test both the spec and implementation. Much of that could be automated using a high level language like Prolog, ML, or Haskell, then generating imperative code (C, C++, etc) from the prototype (executable spec and reference implementation). I donât see how any of this wrestling with LLMs using a context doc is an economic value add.
- Without a spec, what does it mean to say âThe LLM produced correct codeâ? All that post gives is vague, subjective impressions. When any of this is put to a test, the use of LLM for coding appears to slow people down, vs. the expectation of speeding things up.
When AI tools are allowed, developers primarily use Cursor Pro, a popular code editor, and Claude 3.5/3.7 Sonnet. Before starting tasks, developers forecast that allowing AI will reduce completion time by 24%. After completing the study, developers estimate that allowing AI reduced completion time by 20%. Surprisingly, we find that allowing AI actually increases completion time by 19%âAI tooling slowed developers down. This slowdown also contradicts predictions from experts in economics (39% shorter) and ML (38% shorter).
Take their results for whatever you think they are worth, but at least they tried to measure something tangible.