Szymon Teżewski

2.5K posts

Szymon Teżewski banner
Szymon Teżewski

Szymon Teżewski

@jasisz1

Building Aver - a language designed for AI authorship

Augustów Katılım Ağustos 2011
117 Takip Edilen177 Takipçiler
Szymon Teżewski
Szymon Teżewski@jasisz1·
Well... agreed that a trivial `for x in xs` is hard to beat in readability. But once a loop turns into a little state machine (with index, accumulator, current mode, break/continue, mutation), I find recursion + pattern matching easier to read. The state transition is right there in the call shape. And yes, stack implications are real. That’s why Aver warns you when recursion should be tail-recursive ;)
English
0
0
1
11
Julian Harris
Julian Harris@julianharris·
Yes, I was wincing because of intelligibility which is something I think is really important Absolutely unattached to the actual authoring: I agree this will be machine generated for the foreseeable future. Recursion has stacks implications too though right? I guess maybe we as a community have learned a lot from recursion-only languages in the past correct?
English
1
0
1
14
Chris Tate
Chris Tate@ctatedev·
Introducing Zero The programming language for agents. I wanted a systems language that was faster, smaller, and easier for agents to use and repair. Explicit capabilities. JSON diagnostics. Typed safe fixes. Made for agents on day zero.
Chris Tate tweet media
English
374
203
2.7K
1.6M
Szymon Teżewski
Szymon Teżewski@jasisz1·
Apparently the Pope’s first major AI text drops next week. So I have about six days left to find out whether Aver is heresy, or accidentally blessed infrastructure.
English
0
0
0
18
Szymon Teżewski
Szymon Teżewski@jasisz1·
@kskarold Myślę że takiego użycia AI to i papież nie potępia ;) Fajnie wyszły!
Polski
1
0
2
113
ks. Karol Darmas ن ✝⛪⚜
ks. Karol Darmas ن ✝⛪⚜@kskarold·
Bawiłem się rano ChatemGPT; chciałem zobaczyć jak mogłyby wyglądać szaty liturgiczne inspirowane strojem śląskim (konkretnie Rozbarskim) - ładnie to nawet wyszło. Aż szkoda, że takich się nie robi.
ks. Karol Darmas ن ✝⛪⚜ tweet mediaks. Karol Darmas ن ✝⛪⚜ tweet media
Polski
5
2
47
2.5K
Szymon Teżewski
Szymon Teżewski@jasisz1·
Yeah, and that’s not even the most painful one ;) Loops are often where state gets spread around: index here, accumulator there, exit condition at the top, mutation in the middle, the odd break/continue. So Aver pushes that into recursion + TCO, with effects declared at the boundary. Harder to write? Sure. But you’re not usually the one writing it, the agent is. The part I care about humans is review, and I think that gets easier.
English
1
0
1
17
gingerBill
gingerBill@TheGingerBill·
@ThePrimeagen I think people who are designing languages for LLMs are doing the equivalent of horoscopes. They have no idea what they are optimizing for since they don't actually understand how these models work. Why would their new language fair better than any existing language?
English
11
8
225
5.7K
ThePrimeagen
ThePrimeagen@ThePrimeagen·
I have reviewed the language and really tried to understand this, but I really do not understand this language's purpose other than engineers with too much free time, free tokens, and a marketing budget. I was very excited to read about a language is "agent's first." Its just zig with a touch of java and rust...
Chris Tate@ctatedev

Introducing Zero The programming language for agents. I wanted a systems language that was faster, smaller, and easier for agents to use and repair. Explicit capabilities. JSON diagnostics. Typed safe fixes. Made for agents on day zero.

English
159
25
1.9K
272.2K
Szymon Teżewski
Szymon Teżewski@jasisz1·
Aver 0.20 “Pulse”: Tcp (connect, send, ping, writeLine, readLine, close) graduate on wasip2 via wasi:sockets (without preview1 adapter). Same Aver source runs identically on VM, wasm-gc, and wasip2. github.com/jasisz/aver/re…
English
0
0
0
29
Szymon Teżewski
Szymon Teżewski@jasisz1·
Export in a nutshell: - pure fns → Lean defs (kernel-checked totality, partial def fallback when not structural) - verify case → per-sample theorem proven by native_decide - verify … law → real ∀ x, P(x) theorem (auto-proven where Aver can, sorry + per-sample backstop where it can’t) - Random/Time/Disk lift to theorem params with bounded-oracle subtypes - Console/Tcp/Http/main: skipped Toy demo nicely showing the lift itself: averlang.dev/oracle Spec (just cleaned it up a bit…): github.com/jasisz/aver/bl…
English
0
0
1
40
@abdimoalim.bsky.social
@abdimoalim.bsky.social@abdimoalim_·
Anyone who says they're designing a programming language for agents or querying language models is basically selling a library.
English
3
0
14
1.2K
Szymon Teżewski
Szymon Teżewski@jasisz1·
Well... I won’t pretend I wasn’t tempted to add Nat at least five times. But I like the current shape and think it is interesting: keep the core pure and constrained, push validation to the boundary with smart constructors, and use laws for the relational stuff. You lose some refinement-type elegance, but (at least for stable boundaries) the review surface gets better: raw data is validated once, and the core stays boring.
English
1
0
1
18
/// //
/// //@marcsh·
@jasisz1 {}patterns and thus a few stable types, the types are a huge win If however theres a huge set of potential refinements (like 6 or more) added and subtracted from types having 2^N types would be obnoxious, but the other method would be solid For LLMs, Im not completely sure
English
1
0
0
47
Szymon Teżewski
Szymon Teżewski@jasisz1·
A precondition is often just an unmodeled constructor. Another funny Aver bet, finally uncovered: most everyday uses of pre/postconditions and refinement types can be pushed into: • opaque types + smart constructors for invalid values • verify law blocks for behavioral properties • given/when laws for relational invariants More reviewable laws that can be lifted to theorems and less of contract ceremony that gets skipped and forgotten.
/// //@marcsh

@jasisz1 Are obligations preconditions and post conditions? Ill have to look it up Much better than the sillyness from Vercel

English
1
2
5
565
Szymon Teżewski
Szymon Teżewski@jasisz1·
@nickhistgeek Oh man… I wrote a “Common Pushback” section before anyone had actually pushed back. I have been preparing for this moment with unhealthy optimism!
Szymon Teżewski tweet media
English
0
0
0
26
Szymon Teżewski
Szymon Teżewski@jasisz1·
What would you say to no ifs, no loops, no hidden effects? ;)
Lorenzo Nuvoletta@nuvolore

@ctatedev I think agents struggle to code with languages that have many way to do the same thing. Maybe having a strict language that is extremely opinionated and has only 1 way of doing something could make LLMs life easier.

English
2
2
7
883
Szymon Teżewski
Szymon Teżewski@jasisz1·
@radokirov I went for explicit but added language built in compression for agents - you basically say „give me a context on this program for 20 kilobytes”
English
0
0
1
61
Rado Kirov
Rado Kirov@radokirov·
As people are prototyping "the programming language of the future where AI writes most of the code" (see Vercel zero), I am thinking - one of the core tensions in PL design is explicit-vs-implicit. But even if we replace humans with agents, the tradeoff is still there: - fully explicit - context windows are still much smaller than what a large very explicit program would be. - fully implicit - agents are very poor inferring complex deterministic behavior in memory So it would be a tipping on the scale of the eternal PL debate, more than resolving it. My bet is leaning towards more explicit - e.g. @mitsuhiko in lucumr.pocoo.org/2026/2/9/a-lan… talks about barrel files being bad for AI, which funny enough I have seen every large TS codebase move away from prior to AI too.
English
4
0
3
446
Szymon Teżewski
Szymon Teżewski@jasisz1·
I wanted Aver to be readable as a first thing, and verifiable at different levels of complexity. Same verify block runs as a property test, lifts to Dafny or Lean theorem, you pick the depth. And if we talk about games… they were my thing for stress testing aver ;) check it out here averlang.dev/playground/?ga…
English
1
0
1
34
/// //
/// //@marcsh·
@jasisz1 Curious Glad to see folks working on all this I like the pre/post stuff for sure. One of these days I need to evaluate what a simple game would look/work like with them I did that with refinement types in a Voxel renderer, and it really helped out with no boiler plate (C++)
English
1
0
0
64
Szymon Teżewski
Szymon Teżewski@jasisz1·
Vercel calling Zero “the programming language for agents” is a good sign. The category is becoming obvious now: agents need programming languages shaped around tools, effects, structured diagnostics, and reviewable artifacts. But Aver’s bet is the stricter one: agent-written code should not only run. It should replay, explain itself, and generate obligations we can check.
English
2
0
0
244
Szymon Teżewski
Szymon Teżewski@jasisz1·
In Aver there are no pre and post conditions in classic sense, but instead there are collocated verify and verify law blocks, which are… tests with gimmicks and preconditions ;) Those are deterministic property tests at aver verify but with optional adversarial tests with —hostile option. The law blocks are lifted to universal theorems for Lean and Dafny. If you are looking for classic Hoare style - check those veralang.dev and prove.botwork.se
English
1
0
1
38
/// //
/// //@marcsh·
@jasisz1 Are obligations preconditions and post conditions? Ill have to look it up Much better than the sillyness from Vercel
English
1
0
0
614
Szymon Teżewski
Szymon Teżewski@jasisz1·
I deliberately left out “no closures”. That one gets people weirdly emotional...
English
1
0
0
76
Jordan Coeyman
Jordan Coeyman@acoyfellow·
Ok this is pretty cool > zero.coey.dev > github.com/acoyfellow/zer…
Jordan Coeyman tweet media
Chris Tate@ctatedev

I built Zero in 3 days. I didn't expect it to compile. I didn't expect it to mostly self-host. I definitely didn't expect it to work at all. Inspired partly by Bun's rewrite to Rust, Zero started as an experiment. Honestly, the project says more about where AI is today than it does about the language itself. It took more than 3,000 agent tasks to get here, and it's still nowhere near ready for serious comparisons, benchmarks or evals. But the goal is bigger than the current result. The hope is to either create a new language with tooling designed for agents from the ground up, or take learnings and apply it back to existing languages and ecosystems. The ideas are simple: 1. Make languages (and new versions) easy for agents to learn, adapt to and fix on the fly, even when not in the training data. 2. Build a standard library comprehensive enough that most projects don't need external dependencies. 3. Create a tight, fast development loop that even small models can reliably work with. I've never wanted to create a programming language. But after repeatedly running into the same problems, safe but slow builds, fast but unsafe builds, agents struggling with new languages and version changes, wanting faster builds, smaller bundles and better DX, I started wondering: Could accelerated, agent-driven iteration produce a language and tooling stack designed around these constraints from the start? So Zero was born.

English
1
1
73
13.2K
Szymon Teżewski
Szymon Teżewski@jasisz1·
@JamesWard There clearly is the pattern - explicit effects and agent oriented diagnostics at this point ;) but this is the first one I’ve seen from „someone bigger” than a possessed solo dev!
English
0
0
1
116
Szymon Teżewski
Szymon Teżewski@jasisz1·
Anyway, if this niche now has a name, good. I’ve been building my weird corner of it here: averlang.dev Aver is my attempt at answering the annoying post-agent question: “Cool, it wrote the code. Why should I trust it?”
English
0
0
0
39
Szymon Teżewski
Szymon Teżewski@jasisz1·
That’s why Aver puts pressure on different parts of the artifact: - intent strings for why code exists - explicit effect surfaces - deterministic record/replay - verify blocks near behavior - hostile-world testing - proof export to Lean 4 / Dafny Not “AI can write this”, but ”Humans and others can audit this after AI wrote it.”
English
1
0
0
57

Keşfet

@kskarold @ctatedev @TheGingerBill @ThePrimeagen @abdimoalim_ @nickhistgeek @elonmusk @BarackObama