Vaibhav Gupta
@vaibcode
Making a programming language - BAML. @boundaryml, 🦄 aitw podcast: https://www.youtube.com/@boundaryml prev YC, google, msft, deshaw, and other things
Вам может понравиться
I don’t enjoy the syntax and tooling we use to building AI pipelines. Langchain/ai sdk/et all have never really felt “right”. No type safety, single language support, prompt is hidden until I buy some observability suite. So we made a thing youtu.be/wD3zieaV0Yc?si…
youtube.com
YouTube
BAML: a new programming language for AI - Vaibhav Gupta
smart man
It's hard for me to recommend using python to someone new to code, even with AI help. There are still too many ways to shoot yourself in the foot. The main one is not enforcing any types. You have to explain to this new user the concept of typing.
More people talking about unwrap and rust, than those impacted by cloudflare going down
day 0. @antigravity is a dope name btw. thoughts on it coming later this week!
Excited to launch Google Antigravity, our next generation agentic IDE, now powered by Gemini 3!
Meet Google Antigravity, your new agentic development platform. An evolution of the IDE, it's built to help you: - Orchestrate agents operating at a higher, task-oriented level - Run parallel tasks with agents across workspaces - Build anything with Gemini 3 Pro.
AI/ML maybe for training models. Reliable agents apps can't be built on python. No enforced type-safety. Lack of multi-line lambdas. Ever tried streaming w/ python? Its just not designed for it.
So far it has been a breeze, and between linter and AI the newl language is a non-issue. I had some hiccups with error handling, where a provider rate limit produced parse errors due to empty/missing response field. I tried fallback too. But overall great.
I tried to build something similar two years ago. The concept was Markdown file as prompt template with frontmatter defining schema, model, etc, which gets transformed into a callable. But it was shitty so I'm thrilled to trash it and use BAML instead.
wow ur really good at agents (derogatory)
BAML: You're trying to understand how some prompt works. Open a file, and then open the playground, press play. Ok you get it now. Open another, same. And another. And another... Life is better and then you can actually get back to shipping. For real, try it today.
Java: You're trying to understand how some code works. Open a file, it's ~20 lines of trivial boilerplate and getters/setters. Open another, same. And another. And another. This happens like 75 times. It's legitimately hard to find a place where anything is actually being done.
This was my initial intuition too
People say T[] saves tokens by removing quotes and keys. But it has the same problem as JSON Schema: to know sku is required, the LLM must connect tokens 40+ positions apart. In TOON, huge token distances between field names and values force the model to "remember" too much…
![vaibcode's tweet image. People say T[] saves tokens by removing quotes and keys. But it has the same problem as JSON Schema: to know sku is required, the LLM must connect tokens 40+ positions apart.
In TOON, huge token distances between field names and values force the model to "remember" too much…](https://pbs.twimg.com/media/G5vYCBEbQAIipCL.jpg)
My gf pronounces tiny as teeny. Now, I think that sounds more tiny than tiny.
TOON is a waste of time. Optimizations for tiny data structures don't matter. protobuf works because it scales with data complexity. TOON literally says in their docs it gets worse on more complex data.
To everyone who's praising the daylights out of TOON, please, I implore you, read and digest this. 👇🏽👇🏽👇🏽 The BAML folks have been there, tried things out, and there is a delicate balance of token compression vs. encoding semantics. TOON fails at anything more than very basic…
People say T[] saves tokens by removing quotes and keys. But it has the same problem as JSON Schema: to know sku is required, the LLM must connect tokens 40+ positions apart. In TOON, huge token distances between field names and values force the model to "remember" too much…
United States Тренды
- 1. Klay 19.3K posts
- 2. #AEWFullGear 69.1K posts
- 3. Lando 96.8K posts
- 4. McLaren 40.8K posts
- 5. #LasVegasGP 182K posts
- 6. LAFC 15K posts
- 7. Hangman 9,660 posts
- 8. Samoa Joe 4,603 posts
- 9. Gambino 2,096 posts
- 10. Swerve 6,295 posts
- 11. Ja Morant 8,394 posts
- 12. #Toonami 2,753 posts
- 13. Bryson Barnes N/A
- 14. #byucpl N/A
- 15. Verstappen 76.2K posts
- 16. Utah 24K posts
- 17. Benavidez 15.7K posts
- 18. Kimi 37.7K posts
- 19. Mark Briscoe 4,346 posts
- 20. LJ Martin 1,285 posts
Вам может понравиться
Something went wrong.
Something went wrong.