Pinned Tweet
This is a pretty big change for module authors and contributors, but should have little impact on users upfront. The major wins will be seen in more relevant version bumps to modules you actually use, which will cut down on irrelevant CI builds for modules you don't use.
Hey all, sharing a personal project I've been working on for the past few months, pocketlofi.xyz
I've been trying to curate a handful of relaxing and engaging games for both children and adults.
Once installed, games work fully offline, no social integration or ads.
1
2
121
Premium is free for the time being while I get verified, and even when I start offering paid premium content it'll be for less than a latte per month, just enough to cover hosting costs and feature development.
Thank you for your consideration.
1
57
Devon Stewart retweeted
Likely the most important thing I've learned in tech:
if it is easy, it will bite you later
So, do the hard things now. Like actually reviewing that slop.
Replying to @Arikalabs
I packaged it with nix a few years ago just to see if it worked (it does): gist.github.com/noonien/2909…
Devon Stewart retweeted
In 2004, a German demoscene group released a game called .kkrieger.
It was a full 3D first-person shooter with dynamic lighting, textures, sound effects, music, and monster AI.
The entire executable was 96 kilobytes. Instead of storing images and audio files, they wrote C++ and Assembly that generated the 3D meshes, textures, and sound waves mathematically in RAM when the game booted.
Today, a single website's "Accept Cookies" banner is 10x larger than that entire 3D game.
It's 2026 and I can't convince codex to pronounce workdir as anything other than "Work, dear."
"Got it. Work, dear, not work, dear."
2
142
Prior to buying a @KUJIETOOL mini mill or CNC, you should check the comments on @Kickstarter to see if the creator has addressed anyone's issues. kickstarter.com/projects/564…
1
1
138
Mine, so far: machine screw holes not tapped, extremely cheap wire used for the spindle has already cracked and exposed copper, staggering runout on the lead screws, X, Y, and Z axis tolerances so poor that they bind or screech while advancing, Z driver doesn't respond to CNC...
1
2
56
Devon Stewart retweeted
A French engineer who lives quietly in Paris has spent 30 years writing software that the entire internet now runs on without knowing his name.
He wrote the code that streams every YouTube video, every Netflix show, every TikTok clip. He wrote the code that runs the virtual servers underneath AWS, Google Cloud, and Microsoft Azure. He calculated more digits of pi than anyone in history. He has no Twitter. He has no marketing. He just keeps shipping.
His name is Fabrice Bellard.
Here is the story, because almost nobody outside the systems programming world knows what one man has built.
Fabrice was born in 1972 in Grenoble, France. He studied at École Polytechnique, the top French engineering school. He never went to Silicon Valley. He never built a startup empire. He just wrote code.
In 2000 he started a project called FFmpeg, an open-source multimedia framework for encoding, decoding, and streaming video. He was 28. The project did one thing nobody else had done well. It handled every video and audio format that existed, in one library, on every operating system. He led it himself for years.
Today FFmpeg is the invisible engine of the internet. YouTube uses it. Netflix uses it. VLC uses it. Chrome and Firefox use parts of it. Every Android phone, every iPhone, every smart TV, every video editing tool you have ever touched runs FFmpeg somewhere underneath. If you have watched a video on a screen in the last 20 years, Fabrice's code processed it.
He was not done.
In 2003 he started QEMU, a machine emulator and virtualizer. He wrote it solo until version 0.7.1 in 2005. QEMU lets you run any operating system on any other operating system. It became the foundation of modern virtualization. KVM, the Linux kernel hypervisor, runs on top of QEMU. Every major cloud provider, AWS, Google Cloud, Microsoft Azure, IBM Cloud, runs virtual machines on infrastructure built around it. The Quick Emulator is the most cited piece of cloud infrastructure code on Earth.
He kept going.
In 2001 he won the International Obfuscated C Code Contest with a small C compiler that grew into TCC, the Tiny C Compiler. TCC can compile and boot a Linux kernel from source in under 15 seconds. In 2004 he calculated the most digits of pi ever computed at the time, using a personal desktop computer and an algorithm he derived himself called Bellard's formula. In 2011 he wrote a complete PC emulator in pure JavaScript that runs Linux in your browser, a project called JSLinux that engineers still cannot believe is real.
In 2019 he released QuickJS, a small but complete JavaScript engine that fits where V8 cannot. In 2021 he released NNCP, a neural network based lossless data compressor that immediately took the lead on the Large Text Compression Benchmark.
Then he turned his attention to large language models. He built TextSynth Server, a web server with a REST API for running LLMs locally. He released ts_zip and ts_sms, compression utilities that use language models to compress text and short messages at ratios traditional algorithms cannot reach. He released TSAC, a very low bitrate audio compression system. In December 2025 he released Micro QuickJS, a new JavaScript engine for microcontrollers, separate from QuickJS, designed for environments with almost no memory.
Fabrice co-founded a telecom company called Amarisoft in 2012, where he serves as CTO. Amarisoft builds 4G and 5G base station software used by carriers and labs around the world. He has been running it for over a decade while continuing to ship personal projects from his own home page at bellard dot org
He has no Twitter. He has no Instagram. He gives almost no interviews. His personal website is a flat list of projects with no styling, no fonts, no marketing copy. Just titles and links.
A quiet French engineer who never moved to Silicon Valley wrote the code that quietly runs the internet.
He is still shipping.
Devon Stewart retweeted
LLMs are models of written communication between humans.
Because humans have demonstrated reasoning in their writing, and because they frequently talk about and assume this world in their communication, LLMs can be applied toward reasoning and tasks in this world. Stated differently, LLMs are only 'coincidentally' useful for AI.
If humans only talked about angels dancing on the head of a pin, then LLMs built on such communication would have no utility at the tasks we use them for today.
The utility we derive from LLMs should be viewed as *mining* the source data for communication patterns that apply to a given communication context. Consequently, the more 'similar' a description of a problem is to one in the source data, the better the fit, and the more utility we can extract from the model.
Because we derive utility from LLMs by mining them, the quality of the model, including the quality of the data it was trained on, matter enormously. Every piece of data modifies the weights of the model in some direction or another--toward utility (for a given set of tasks) or away from it.
Another consequence is that agents built on LLMs have an upper-bound on how useful they can be and in what contexts. Even if we assume perfect quality data and perfect generalization, we are still bounded by the size of the training data.
What is this upper bound?
Well, there is the sort of 'one-shot' bound, which is disappointingly low, and stems from the so-called auto-regressive nature of LLMs. If you throw a problem at an LLM and ask for a solution in a single step, then you will be disappointed.
Chain-of-thought can increase the bound by taking what, for the LLM, is a 'novel' problem (one not seen in the training data), and factoring it into smaller steps, each of which is more likely to be similar to something in the training data. This is an improvement in mining efficiency, allowing us to get farther.
Tool usage is even better: when we give models tools, and force them to use them, then we introduce feedback from reality at every step, allowing us to correct the errors that would otherwise accumulate unbounded in long-form, one-shot solutions.
What I don't believe we have seen, but we should expect to see at some point, are frontier models being built solely for tool usage and chain-of-thought, from unimaginably large, purely synthetic data sets. I would think this gets us into genuinely new territory.
Even then, however, we are still capped to the reasoning and problems of the source training data set (and, obviously, those that are 'similar' to them, for some precise definition of 'similar').
Tell your LLM that in some world as close to ours as possible, things fall upward. Then ask it what a house looks like and how people move from house to house. It will fumble over things a five year old would intuitively understand.
Why? Because fundamentally, LLMs are models of communication, and we don't talk about worlds like ours but where things fall upward, and we don't talk about things that are 'similar' to this.
Similarly, ask your coding agent to program a quantum computer to do something novel, which not even researchers know if is possible, and you're going to be very disappointed by the nonsense results. Why? Fundamentally there's just not enough training data, probably even if you break up the problem into smaller pieces: the fact that each subproblem is smaller helps increase the probability of finding something 'similar', but not enough to be very interesting.
Now, before discussing how to 'fix' these limitations, it's worth making an important point: there is actually no critical need to 'fix' these problems. Even if the entire field of AI made no progress from this point onward, LLMs are still useful enough to reinvent the world. This reinvention cannot be stopped.
However, if this reinvention does not happen, then your days of cursing your agent for its profoundly senior moments are never really and truly over. Given sufficient time, and sufficient intelligence and experience on your part, you will always hit walls, even if you can find ways around them.
Moreover, if LLMs hold the promise of reinventing the world, then what might happen if LLMs are mere stepping stones on the way to new types of models? To get something we would consider superhuman intelligence--which I might define as, at minimum, as great as the difference between a human and an ant--something different is needed.
So what might this reinvention look like?
A lot of folks have ideas in this space, and many are trying to figure out, but I think one thing is clear: the models of the future will go far beyond just modeling written communication between humans. They will model aspects of reality (including not just the physical world, but also mathematics), as it exists independent from written human communication.
Moreover, I highly doubt that the successors of LLMs will be a single type of model: rather, I think it likely we will have lots of different types of models, which we orchestrate to produce superhuman intelligence. That's a lot like how our brains are designed, with different parts specializing for different functions, each one efficiently tuned to its purpose.
In summary, LLMs are models of communication between humans, which has profound limitations on the utility we can extract from them (as well as the mechanism of performing this extraction). If they are as good as we ever get, then that's fine -- the world will still be reinvented. But personally, I think they're a stepping stone to something better: to models that are based not on written communication between humans, but on objective reality. Those models, and the way they are orchestrated, hold the key to unlocking the full promise of AI.
Devon Stewart retweeted
Just released proteus-diff, a CLI that compares .proto files and splits breaking changes into wire vs source compat axes.
Renumbering fields? wire-breaking, source-safe. Renaming? the opposite. ~250ms on 6k messages 🚀
ghostdogpr.github.io/proteus…
Devon Stewart retweeted
Let's play a game - win32 types vs Polish language:
LPCWSTR
PSZCZYNA
WCSLEN
WCZESNY
LPCTSTR
BYDGOSZCZ
WSTRZAS
HGDIOBJ
DOWOD
HWINSTA
DLUGOSC
LPCSTR
DWORD
KAL
LPWSTR
SZCZECIN
PCWSTR
BLAD
PUHALF
CHUJ
UHALF
Devon Stewart retweeted
In the next version of Bun
Bun goes ad-supported!
![$ bun test bun test v1.3.12-free (alb2c3d) • src/utils.test.ts (42) [0.12ms] x src/auth.test.ts > validates token [0.89ms] SPONSORED error in 5. - Performance issues? More common than you think. Ships in unmarked packages. Take the quiz → - steady A error unlocks after ad 246 pass • 1 fail • [8.21s] Run bun upgrade -pro to disable ads.](https://pbs.twimg.com/media/HEzGnX1boAADLAn.jpg?name=small&format=webp)
ALT $ bun test bun test v1.3.12-free (alb2c3d) • src/utils.test.ts (42) [0.12ms] x src/auth.test.ts > validates token [0.89ms] SPONSORED error in 5. - Performance issues? More common than you think. Ships in unmarked packages. Take the quiz → - steady A error unlocks after ad 246 pass • 1 fail • [8.21s] Run bun upgrade -pro to disable ads.

















