
Elixir developer @ Bluecode Mastodon profile @tkowal@fosstodon.org
Mechelinki, Poland
Joined December 2011
- Tweets2.7K
- Following250
- Followers698
- Likes7.7K
I am playing Aliens: Dark Descent and it captures the feel of Cameron's Aliens movie better than any other game from the universe I've played. Marines feel bad ass, but every alien is still a threat if not handled carefully. Piece of art!
33
Tomasz Kowal retweeted
48 hours later, meet MinerLlama🤣
2× faster LLM inference on retired crypto-mining GPUs🚀
New Pascal kernels for llama.cpp push Flint(39.8B MoE, Qwen 3.8 27B-level quality) on 9× P106-090s from 14–16 to 28–32 tok/s🚀
Link to repo in comment👇
CC @RichardSocher @Recursive_SI
Tomasz Kowal retweeted
So I was trying to understand why pull requests speeding up the Elixir version were being denied by the clankers while they were being accepted to Go/Express/Rust.
It turns out that, not only each project had a different starting point and a different prompt, but also each project also has a different set of verifications that they are required to pass!
For example, Elixir is the only guided to "preserve the pinned Rails reference behaviour", the only with a explicit verification that we are using Redis and draining its queue (so we cannot remove it while others can), and the only one actively enforcing it.
So it turns out that the people telling you to not read the code cannot even use AI to set up fair benchmarks (probably because they would have to read the code in order to do that).
You don't have to believe me, you can do your own research or ask your agent to do it for you. E.g. here is what Claude Code came back with: gist.github.com/josevalim/ae…
Replying to @dhh
Also extracted a separate verification project that's responsible for the benchmarking and checking of new implementations and improvements to the existing ones. The full report from the latest runs is there too. github.com/basecamp/once-cam…
Tomasz Kowal retweeted
I'm still reviewing code across Dashbit's projects and my open source projects. And, because of coding agents, I am reviewing significantly more code than before.
Compared to last year, how much code are you reviewing?
63%More code
9%About the same
14%Less code
15%I no longer review code
779 votes • Final results Tomasz Kowal retweeted
To whom it may concern 🙃️
In abstract terms, I think of a software system (e.g. a web server, database, message queue, ...) as a collection of dynamic activities. Multiple things are happening: handling of requests, background tasks, periodical jobs, error handling, reporting, and recovery. Many of this things are completely independent, but some of them are working together.
These activities have their lifecycles. We want to be able to start the activities in the proper order. We want to be able to dynamically start more activities. We want to be able to stop the activities in the reverse order. If Bar depends on Foo, we want to start Foo first, Bar second, and we want to stop them in the reverse order. Some activities will be decomposed into smaller subactivities. So if Foo internally runs Bar and Baz, we want to be able to stop the entire subtree. For example, if a user hits the cancel button, we want to be able to stop all the activities related to that task. And we want to be able to do this even if the task is stuck in some CPU processing loop.
A software system has non-binary success semantics. Just because something fails, doesn't mean everything should stop working. But whatever is tightly coupled to the failing activity should stop too. However, the failing thing should not leave inconsistent mess (state, dangling resources) behind.
BEAM & OTP is give us the primitives that address just that. The way I think about my systems in abstract maps very closely to the way I implement them. This vastly reduces the mental overload of thinking about the code. The conceptual view of the system maps very closely to the actual physical model (process structure). This is a power of a good abstraction. For more details, see my talk Soul of Erlang and Elixir, where I tried to explain the underlying ideas.
The conceptual properties I mentioned above will have to be supported, because they are inherent to a software system. If you're building a software system without this support, these concerns will get entangled with your business logic. Whether you're doing it manually or vibe it, this work becomes more complicated. You or your agent have to deal with non-functional requirements together with the functional ones. This will make your job and the jobs of agents more difficult. Remember: code is the part of the context, perhaps its largest part.
So in the age of agentic development, my position about BEAM remains the same. If you're building a software system, it is the tool that fits the problem, and it will simplify your life. You might bump into situations where some other properties will be more important (e.g. raw processing speed), but you can then implement these parts of your system in another language, say Rust, and manage them from a BEAM-powered system.
As many of you know, I'm a long-time user of Erlang and Elixir. But I don't choose those languages for the syntax. For the record, I actually prefer C-style curly-braces syntax. I use those languages because of the runtime properties and guarantees, and because this technology reduces the impedance mismatch between the way I think about the system and the way I implement it. I also actually mostly prefer Erlang syntax over Elixir (minus some exceptions). The reason why I use Elixir is mostly tooling & some language features (e.g. macros). But the choice of the language was never a fanboy choice for me. If you're not building a system, BEAM langs are most likely not your hammers.
Tomasz Kowal retweeted
I’ll be speaking at WAWTECH in Warsaw next month!
They gave me 5 discount codes for 50% off. First come, first served: VALIM50
See you there!
Tomasz Kowal retweeted
I took a stab at improving the performance of Elixir. I used the PR for Elixir that @ZachSDaniel1 posted and worked some more on it. In the benchmark Elixir is now faster than Rust (ccece30 newest commit) for the Sidebar and Search benchmark.
Last conversion of the Campfire from my token stash goes to JavaScript/Express. As with the others, I'm sure you can find more to optimize. But this gives you a sense of what order of magnitude these options exist within with agent-led development. github.com/basecamp/once-cam…
Tomasz Kowal retweeted
I have a branch with various performance improvements to the Elixir version, focused on specifically the websocket side of things. I didn't spend much time on the rest.
The load generation code was flawed in overload scenarios because any socket that had an issue just stayed disconnected forever, so I assume no benchmarks were run in this fashion prior. Now those reconnect on disconnect and continue sending traffic (a more realistic scenario).
Most relevant finding:
In an overload situation, the rust implementation delivers a whopping 1% of messages to all connected clients, and the Elixir implementation delivers 100% of messages to all clients, applying natural backpressure to writes.
Speed is cool. Rust is fast. Speed is not even close to the most important thing when building real applications.
I've had agents implement and optimize the Campfire web app in Elixir, Go, and Rust. Yes, Ruby is the slowest. I didn't care about that when the pay-off was huge in terms of productivity and developer joy. But if you're no longer reading the code? github.com/basecamp/once-cam…
Readers added context they thought people might want to know
The shown benchmarks use non-equivalent implementations: Elixir routes all DB queries through a single GenServer with Redis for jobs (retained from Rails), while Rust uses in-process queues, omits CSRF tags/fields, and applies other caching changes.
github.com/basecamp/once-…
raw.githubusercontent.com/basecamp/once-…
github.com/basecamp/once-…
github.com/basecamp/once-…
x.com/zachsdaniel1/s…
Tomasz Kowal retweeted
Do any AI code review tools measure what humans catch that they miss?
Not just accepted/rejected/dismissed comment rates, but actual recall vs. human reviewers. If a human finds an issue the AI didn't, is that tracked, quantified, and fed back into future reviews?
I see a lot of comparisons between AI code reviews but we also to have measure them against the team's own baseline if we want to automate more parts of our processes. Some of those metrics also need to be quantified per directory/file.
Tomasz Kowal retweeted
Man even StarCraft is reducing worker counts. AI replacing workers is getting out of hand.
Tomasz Kowal retweeted
I believe in the EEF! @TheErlef
I'm on the board so I get to see first-hand the amount of work and toil that goes into making the #ElixirLang, #Erlang & greater #BEAM ecosystem better.
What I also get to see, though, is how much we need your help. Relative to the revenue generated by companies that use Elixir, the EEF has a pittance.
Most of us know what's happening the OSS world and security right now. What you probably don't see is the massive amount of coordination and scrambling going on at every level to secure the open source ecosystem. While this may have been operating at a sustainable pace before, the advent of AI has fundamentally changed the game.
The EEF maintains a CNA, and actively coordinates and advises across the board on security issues. They are helping to make our languages more secure, and even hex.pm!
But the bad news here is that this is no longer sustainable. We need people who are generating revenue on the BEAM ecosystem to pitch in. We need a bigger team of people solving these problems because what we have is barely enough as is.
The real problem, however is that this is progressively getting worse by multiple orders of magnitude. CVEs coming in left front and center.
This is a tragedy of the commons. People are not helping foot the bill for our open source ecosystem. Before that used to mean "inconvenience", as you had to fork packages or version mismatches. But now? Now it means getting pwned.
Join me in becoming a member of the EEF, and please, if you run a company built on a BEAM language: DONATE.
Anyone tried building their own offline coding assistant? I am wondering how far I can get with my RTX3080. I am using Clause as pair programmer. I work iteratively and check the code on each step. Sometimes putting knowledge in skills, so it doesn't repeat mistakes.
3
1
191
Obviously, a 10GB memory GPU won't fit the best models, but I need just enough power to let agent search the codebase iteratively, tell me where stuff is and then generate code under my supervision. What model and agent would you use? Qwen + OpenCode?
1
1
95
Romania is my favourite. Reminds me of Xena - the Warrior Princess. #Eurovision
1
315
Norway gives me Maneskin vibes #Eurovision but Damiano had more unique vocals. Nice work, anyway!
3
416
What decade did Sal da Vinci come from? Blast from the past! #Eurovision
1
331
I don’t know how Cyprus got to finals. She was out of tune :/ #Eurovision
2
1
489
From all techno/dance entries this year, Sweden is my favourite #Eurovision
224














