The workflow of

terminal-based agentic coding -> pr -> comments -> updated pr

is so efficient and pleasant.

Go Burhan!

You'll understand our campaign more from this launch video than anything else. We gave it our heart and soul. Give it a watch.

I'm excited for the a generation of LLM-first programming languages. Languages that the model RLs well on. In the extreme case, you have inverse problem--given an LLM what's the best PL.

ENNUI ~ Elegant Neural Network User Interface ~ Welcome to ENNUI - An elegant neural network user interface which allows you to easily design, train, and visualize neural networks.

math.mit.edu/ennui/
github.com/martinjm97/E...

Back in the day, some friends and I made a website to teach people about neural networks. We implemented it all in typescript and haven't updated it in years. I can't believe it still works.

I really like how simple it is to play with and the little articles.

I'll be in NYC next week. I'd love to talk with folks in the robotics/AI space, especially on large-scale model training of VLA and factory automation.

I see your point that our brain both evaluates our actions and produces new actions. The idea of extrinsic reward is odd.

I wonder at a practical level what sota systems have a coupled backbone and then an action head and a reward head.

Happy to view this as a north star, but I'm curious.

The rise and fall of peer review Why the greatest scientific experiment in history failed, and why that's a great thing

Relatedly: maybe also a good time to question our assumptions about why peer review is helpful in the first place.

www.experimental-history.com/p/the-rise-a...

But in the future, it could look like

```python
from tap import Tap, Positional, ShortFlag, AddArgKwargs
from typing import Annotated

class Args(Tap):
arg_1: Annotated[int, Positional]
arg_2: Annotated[list[str], ShortFlag("a"), AddArgKwargs(nargs=2)]
```

For example, passing arguments currently looks like

```python
from tap import Tap

class Args(Tap):
arg_1: int
arg_2: list[str]

def configure(self) -> None:
self.add_argument("arg_1")
self.add_argument("-a", "--arg_2", nargs=2)
```

Feature request: Positional argument with default value · Issue #169 · swansonk14/typed-argument-parser Would it be possible to add support for default values for positional arguments, similarly to the standard Python argparse? class Args(Tap): config_path: str = "config.yaml" force_update: bool = Fa...

Thanks to the community and advances in type-hinting, it should be possible to improve `Tap` a lot.

`Annotated` allows the programmer to annotate a type (e.g., height: Annotated[int, "in inches"]), which we can use to pass options to `argparse`.

github.com/swansonk14/t...

Why is flow matching implemented with the differential form of the ODEs and not the integral equation version?

A discretization scheme for solving the integral equation would look like simulating the dynamics all at once.

Social media posting on research is rife with

vagueposting - attention seeking posts by people who want to act as if they're not seeking attention

I just want technical content :(

Appending "do not use defensive programming" to the end of my queries makes AI generated code much more terse and useful. I can handle weird edge cases after, but I want to start by getting the core functionality in place.

I've been surprised by how much I like OpenAI Atlas. It's so fast and it's the best AI-based search engine I've used. UX is amazing.

I reserve the right to dislike it in the future when it presumably tries to bake ADs into LLM outputs in a mublock-proof way :'(.

LLMs I've used have been almost completely useless at formalizing anything to do with denotational or operational semantics.

I find it so shocking how good models are at advanced math and coding, and then how bad they are at the intersection of the two.

I can't get any LLM to make a type system :(. PL folks have to write more papers to train the LLMs.

I remember when Windows added ads to the start menu. I think apple turned on a default to show ads (at least it can be toggled off for now). I left Linux because of the driver bugs, but Linux is looking pretty good now.

I believe these techniques can lead to:
new simulators that enable optimization through contact
new probabilistic algorithms with discontinuities
new differentiable renderers that better model occlusion
new ways to solve ML problems e.g. selecting MOEs and doing image gen in continuous space

Full abstract
www.eecs.mit.edu/eecs-events/...

TLDR: Modeling discontinuities (occlusion, cracks, and contact) is hard, but distribution theory gives a systematic way to model these phenomena. I introduce distribution programming, which enables implementation/autodiff by using programming primitives that denote distributions.

Successfully defended my thesis:
Distribution Programming for Physics-Based Computations

I'd love a good biking route from Cambridge to Logan Airport and a safe place to park the bike at the airport 🚲 ✈️ .

I know this isn't exclusively in Cambridge, but is this possible @realBurhanAzeem @Marc_C_McGovern? Thanks for all you've done for bikers!

I got an email that was clearly automated but had a personalized tag line about my research in it. Something of the form: "I was interested in your research on ..."

I'm just bracing for a new wave of spam like this. I need an AI to screen my emails from this slop.

My defense talk is completely hopeless. There is nothing anyone can do for it.

Manual, light-weight intuition building, automation to get a robust setup that can easily be edited, and manually edit.

Plotting workflow: prototype in Desmos, use chat to write matplotlib code, and do by-hand checking and tuning.

I want more of this workflow.

Does Meta stand for metadata because they store all your data and sell it to advertisers?

I just randomly joined a member of the MIT admin for lunch. He's part of the office that helps students reserve space at MIT. It was a great reminder of all the wonderful staff at MIT who made helped make my days at MIT better. Thanks Aaron!