we worked really hard to add type safety to runtime reflection and eval. two examples of things we enable are: compile code at runtime and get back a typed function with a checked signature. build a class at runtime, unreflect it, and use that same runtime type as a compile time generic parameter.

Sounds like we need a new language 🤔 Boundaryml.com

i'm just always seeing "context automatically compacting" longer context windows when?

I think we’re looking for this github.com/BoundaryML/feelin…

we're also fans! jev as an AI function coming to BAML in tonights release! github.com/BoundaryML/baml/p…

damn got me.

code is always the source of truth i can't wait until harnesses are post-trained to do this more

this is your daily reminder to read something. read the specs read the plans read the code read the prs model drift is certain.

you've heard of IsRecord, but have you seen CookedRecord?

no more queuing messages in side chats apparently!

Agents will find and use every escape hatch

yes we are :) just put reflection into prod, so that should be part 1 (being able to hot swap code in the middle of a process running for different version) Telemetry capture gives us part 2 (using real data to create the basis on which we should create simulation software) theres a few things left before the true simulation arc is done: * make telemetry work for distributed stack * expose the visual graph as an API from the compiler via the reflection package * make some library that wraps around those three layers so its a few lines of code but imo most of the meat here is likely to be application specific, so we're currently focusing mostly on the building blocks. we're gonna talk about the first few bits when we go on the podcast tmrw! luma.com/baml

apparently we do 0ver at baml w/o knowing that

“Make no mistakes”

i know for some folks this is peanuts, but its so cool to see that in the last 48 hours at least 75 people were watching @dexhorthy and I yap on youtube every hour. love you all <3

specifically the end when we mentioned the different products at different layers of factory building

the even more fun part is adding constraints like `throws never` on function types prevents you from passing lambdas that throw things as well!

try catch sucking is why we made catch an expression in BAML. and now the compiler can infer exactly what every function throws and catch can be type-safe. piped.video/LnqZdiVWAes?t=680

software factories

same code, no compiler errors, has a runtime crash. that is bad, and we should try and minimize it. I think we can all agree on that! i'm really confused about what your position is. are you saying structural matching doesn't lead to issues in typescript?

When the language doesn't make certain guarantees, your code lies to you, and its very easy for ai to slip code like this through. see line 14. spooky.

sorry i used a bar example! here's where it gets ya

then i don't give a f**k

why rust is better than typescript: no structural matching its convenient for human code, but a footgun for ai code. line 12 should be a compiler error!

how did you end up measuring the 25% rate? just vibes? one thing we've been experimenting with in BAML is how do you build examples that don't bias the model. for dumber models, usually find partial few shot prompt works well, and this works even better for bigger models

Be most proud of the code you delete

you can check it out here: boundaryml.com/ (or at least a glimpse of it) but we gave a talk at AI Engineer where we showed a small part of it: piped.video/watch?v=AMiyLItE…

should i be worried about our expense policy?

you're gonna save a bunch more time just using baml ;) we've got a profiling and callstack system built in thats 6-15x faster than rust's OTEL option, and 1000-50000x small bytes.

We have a mutex on our GitHub repo

5 years ago this would be bikeshedding

our perks are the best: unlimited envelopes

thats why we have built in tools for really good debugging! (fully queryable as well)

being a manager in 2026

@benswerd what about in memory, language-based sandboxing?

I think what you're looking for is ai functions! which i perhaps should have scrolled too... in case you're curious, i'd love honest feedback: boundaryml.com/ (a lot of the copy is still in progress)

perl is dead.

hmm would you take a language that gives you workflow diagrams thats stays 100% upto date with your code. (and perhaps lights up as you execute code)

i'd be really curious of your take on our concurrency model in BAML. tldr is green threads with return values, but we have cascading cancellations (including for infinite loops), and worker pooling etc. beps.boundaryml.com/beps/34

interesting...

Waymo just drop me off on the other side of the street please

full episode: piped.video/watch?v=Yh4eL60N…

SlopCodeBench is such a unique benchmark and raises a lot of fundamental questions about long horizon steering. Post training may not be all we need. @dexhorthy has been running this benchmark non-stop for past 2 weeks and the data doesn't lie

more rust devs need to read this by

No modern language could have the predicted writing code and it's going to be fully clanker operation. e.g. I want it to have rust's level of "if its compiles it works" but with go's compile times. Also try doing application concurrency/cancellation in rust, its hard.