Making a programming language - BAML. @boundaryml, 🦄 aitw podcast: piped.video/@boundaryml prev YC, google, msft, deshaw, and other things
Seattle, WA
- Tweets 1,501
- Following 561
- Followers 3,498
- Likes 1,272
Pinned Tweet
Slop is code you dont read. And better models means more slop.
the solution? write sloppy tools to build stable systems.
but for the most committed people, go build new foundational systems. Rebuild github, rebuild tmux, rebuild database, and yes rebuild programming languages.
Big thanks to @swyx and the whole community for having us out at @aiDotEngineer
6
11
78
14,888
Composites are a great add for js, but I’m always curious, why was this not the default?
I suspect eq by value is more frequent than eq by identity. The frequent use case should be the default!
ECMAScript excitement 😉
Congrats to @acutmore from @TechAtBloomberg on advancing the Composites proposal to Stage 2 at @TC39 today 🎉
Composites are objects that are compared by value instead of identity. They are an evolution of Record & Tuple proposal.
ALT const pos1 = Composite({ x: 1, y: 4 }); const pos2 = Composite({ x: 1, y: 4 }); pos1 === pos2; // true const positions = new Set(); // the standard ES Set positions.add(pos1); positions.has(pos2); // true
1
254
fun talk today!
every single agent in the world will have their own email soon. it’s the future.
email is the perfect communication platform for agents. it’s a 50+ yr old protocol & it’s not going anywhere.
give your agent an email in two minutes w/ @primitive📧
44
4
405
20,135
AI will never ever write code in assembly.
Lets me tell you why.
I spent most of my career writing assembly, and the hardest part of job was never to write things in ASM, it was deciding what to write in ASM.
Writing things in ASM has a real tax. You lock in architectures, memory layouts, and making your code fast, means making it more rigid.
Think about our supply chains during covid. Shipping stops and suddenly everything breaks. Speed is never free.
AI might write some code in ASM, but only if the value you're getting is worth the benefit. New folks can benefit from SSE/AVX instructions, but the majority of code in the world needs to remain flexible and updatable.
At 1M Tokens / second, would you rather generate a for loop in ASM? or an entire app in Python? Time will always be money and compilers and higher level languages are going to win.
Hopefully BAML :)
It looks very interesting, great work! But on the other hand, I doubt that in a year we'll still be writing and reading code ourselves. Why do we need another programming language if AI will probably be writing code directly in assembly?
30
3
129
18,017
Buddy... You're a YC larper reinventing the wheel with something no corporation will ever adopt or use.
I'll let the agents decide when to write what in the SaaS stack in ASM
1
27
You aren’t even remotely capable of what a frontier model can do, so stop pretending like you are the authority on what language the stack will use 🙃
1
68
Sounds like we need a new language 🤔
Boundaryml.com
JavaScript was never designed for the server. If that doesn't change soon, JavaScript on the server will be replaced by Rust.
The fix needs to happen in the engine. Sound types. Ahead-of-time compilation. Threads with shared objects.
25
15
364
67,371
since you’re working on language-native o11y I think you probably have a great answer.
Is there value in allowing the agent to explore data layouts in asm? Or is this something that’ll be solved via better guidelines and perf focused post training
nitter.net/kamikaz1_k/status/2104…?
It was a facepalm moment, and I did see your Mitchell RT earlier. But is there something to be said about the language nudging the AI into arranging things better?
Is it just training distributions bubbling up again?
Could be cool as a guiding tool…if effective
1
1
195
I take a bet against that if the time horizon is not tomorrow.
5 years from now. You set the challenge. Winner buys dinner
1
1
138
AI can track stability of a system and once unchanged for long enough rewrite into ASM? otherwise sounds like a very fair point
1
187
the “AI will just write assembly directly” take always seems to come from people who think there is one universal type of assembly that works on all machines
1
22
560
Jarred wants to save JavaScript by making it faster with Rust.
DHH never wants to program again.
Effect devs think Effect will save TypeScript.
All current languages and frameworks are dead. In the end, only math will matter.
Never thought I'd live to see this. And I love it.
13
4
183
21,110
It’s basically semantics of rust, but no lifetimes to deal with. Plus we brought in unions from typescript!
We do live onboardings every Thursday if you’re interested boundaryml.com/eap
64
Correctness of Rust and concurrency of Go don’t really mix though? And neither do dynamism of Python and correctness of Rust.
How do you square that circle?
1
411
re the dynamic of python, this post explains it better nitter.net/vaibcode/status/210372…
but concurrency and correctness are a notoriously hard duo. Given that the target for what people should build in BAML is application code not systems code, we bias a bit more towards practical concurrency over lifetime management.
this mean:
1. we make it possible to cancel hot loops
2. we have really nice thread pool ergonomics
3. we can support durable execution / multi-machine concurrency (not yet shipped, still in slop stage and needs to be cleaned up)
4. all container types are threadsafe w/o gil
you can still run into data races/logic bugs, but we think with really good tracing / observability you'll be able to diagnose issues much better. to help with this, we made BTEL - our improvement over OTEL - thats just 10ns overhead on function calls, and 1000s of times smaller in memory/disk footprint.
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.
1
305
