Szymon Teżewski

Building Aver - a language designed for AI authorship

[Szymon Teżewski @jasisz1]

You know what's more surreal than building a language for AI to write and humans to read? Talking about it with @JamesWard and @BruceEckel in my second language, making hand gestures to explain language design choices. Thanks for having me on @HappyPathProg!
https://t.co/4YB7xAtw6L

[Szymon Teżewski @jasisz1]

I was trying to explain my current work in a poster and I can't decide between Polish Poster School minimalism vs... the other one.

[Szymon Teżewski @jasisz1]

@FilipWalkowiak Nie tyle że „nie wiedzą, do czego służą”, co przez 33 lata uczyli się, że do niczego ;) A tory idą przez środek miasta i przecinają kilka ważnych ulic, więc efekt jest do przewidzenia…

[Szymon Teżewski @jasisz1]

bye Bye Fable,
i’m gonna miss you so.
all that’s left now are traces
and the hypothetical bill.

[Szymon Teżewski @jasisz1]

@misterclayt0n You might like this https://t.co/KjIBSatFJG - it is the opposite direction: keep C and retrofit serious specs/testing/proofs onto it.

[Szymon Teżewski @jasisz1]

@AlessiaBaumgar It is an easy one! (if you by accident already speak other Slavic language xD)

[Szymon Teżewski @jasisz1]

I was touching grass in the morning. And it is still funny to me that the language which distincts safety and security (they are both „bezpieczeństwo” in Polish) cannot distinct sweet and sour cherry without an adjective…

[Szymon Teżewski @jasisz1]

I now see Lean mainly as a pure kernel for proving... the surface language genuinely gives COBOL vibes. And the more I experiment, the more I see feeding the prover's residual back to LLM as a failure mode: it turns it from being "conjecturer" to "tactic-guesser", tactics should be compiler/prover mechanics.

[Szymon Teżewski @jasisz1]

I doubt there's research specifically about that. And if there is... well, it'll be a smart formula without real meaning. My intuition is this sits between coverage-guided fuzzing and mutation testing — you'd quickly generate a distilled corpus of "hard negatives" that becomes the best cheap problem-aware sieve. Hard-squeeze the survivors, reiterate with a bigger corpus.

[Szymon Teżewski @jasisz1]

@geofflangdale Very interesting question, but well... it depends whether we're thinking "maybe there's a subtle bug" or "someone might have sneaked a bug in there". Those two questions have different answers.

[Szymon Teżewski @jasisz1]

@hvitis_ Chyba taka właśnie podarowałem byłej dziewczynie… gdybym wiedział że to będzie szary kruk ;)

[Szymon Teżewski @jasisz1]

I was tempted to add taint-style tracking to Aver. Then I realized I can get most of the practical guarantees in the boring ML way: Secret -> LogSafe -> logEvent...
Opaque types I already have, explicit declassification, consumed only by typed sinks.
So again, it comes back like a chorus: "just use types".

[Szymon Teżewski @jasisz1]

There is similar thing with Polish "oczy/oko" which in contemporary Russian is "глаза" but there is an archaic/poetic "очи/око".
It works the other way too - город/gród, друг/druh, живот/żywot. The last one is interesting because many Poles will read only as archaic "życie", but the Hail Mary prayer has words "owoc żywota Twojego" which literally means "belly/womb" in this phrase, the same as in Russian.

[Szymon Teżewski @jasisz1]

@rezoundous It was awesome despite it used very strange language (no matter the language I was prompting It), full of its own strange words and terms.

[Szymon Teżewski @jasisz1]

@zielonoszara @autodestrukkcja I ładne zdjęcia potem są z koronami!

[Szymon Teżewski @jasisz1]

@Kingkosec @callmejeprian Baby is constantly swallowing whatever it has in there…

[Szymon Teżewski @jasisz1]

@code_star I have access, but only because I’m European and building something formal enough to be mistaken for German paperwork.

[Szymon Teżewski @jasisz1]

@willchamberlain Good test, I've passed it.

[Szymon Teżewski @jasisz1]

Paper dump again, this is closest yet to what I'm building in Aver: https://t.co/44wWXCHhMY
Same loop idea: agent proposes a helper lemma, verifier decides. But in Aver I'm trying to split the work by "the kind of lemma":
There is the plumbing/the boring glue: don't guess it, calculate it from the shape (HIR/ProofIR I did for the rescue) or stuck proof (Ireland & Bundy, '96)
Agents live in the generalizations: where the model's high-level understanding helps. Plus every guess can be quickly filtered out in the VM before the prover (QuickSpec/HipSpec spirit).
Funnily this somehow starts to compose nicely...

[Szymon Teżewski @jasisz1]

@kenthecowboy_ @BigGulpAmerikan There is one scene in directors cut in which she buys and tries the dress and this is the closest one to what you describe xD

[Szymon Teżewski @jasisz1]

@titos2k It runs in the browser and it is "bring your own LLM", there is no server. Source code is on GitHub if you want to inspect/run locally for full trust https://t.co/FGQW0Rg14c

[Szymon Teżewski @jasisz1]

My friend @titos2k bulit this and I love it https://t.co/J0HNhem0Z2 - it's basically AI-augumented mind map, with AI helping you to distill ideas (asks questions, finds problems, generates ideas). I really enjoy this approach because it genuinely feels like AI-augumented thought process and not like you outsource whole thinking.