avilay

Is simulating the effects on humans not an option? Or is it that simulation technologies are not good (accurate) enough?

Thank you for entertaining my curiosity! I have been meaning to educate myself on this but don't know anybody in my circle who is knowledgeable in this field.

What is the hard part then? My read from here is that it is testing drugs on humans, or rather getting the approvals needed to begin human trials. Is this understanding correct?

Thanks for your response. I have been wanting to dip my toe in the space of language building for a while, but keep getting discouraged by the advances of coding agents. Love your reason for building this!

> I made this one because I had a really strong design in mind that I think should exist.

Given we are getting to the point where coding agents write most of the code, what motivates you to create a new programming language? Follow up: do you build these languages keeping humans in mind, or are these meant to be used by AI agents?

+100 I tried to get Claude and Gemini to build me a course on probabilistic ML from scratch. They both failed. This was a few months ago. The best they could do was design a course curriculum. My learning since then has been through asking the model to explain a very narrow topic from the curriculum. And then go through the agonizing back-and-forth questioning and answering to actually build an understanding. No doubt the models are great at this and my learning has never been so through with such solid foundations. But zero-shot it was not!

I am a heavy Typora user and I really like it except when it starts to glitch with longer documents (> 5000 words). How is this different from Typora? To be clear, I think it is ok if it isn't, the world is big enough for multiple markdown editors :-)

Also, what does the term "greco-roman" have anything to do with Markdown editors?

Why? Just...why?

I get that sometimes we like to build useless things because they are elegant, or beautiful, or "cool" in some sense of the word. But CMake is not cool in any sense of the word! I can barely use it to write build files, and that too I have gladly offloaded to AI. Sometimes when looking at CMake files in a large project, I become nostalgic about Makefiles, but then I think about what the rewrite would entail, and I resign myself to my fate.

Good point. Same with me. Despite it being very tiresome, all the back and forth that I do with it really deepens my understanding of the topic. I have been on Opus so far, let me try Fable and see if it gets better. I haven’t tried voice either. Next time I go for a walk I’ll try that!

I have had very similar experience! I wanted to learn Probablistic ML, checked out a couple of MOOCs, but didn't find any that were at my level - some were too advanced, some too beginner level. Claude was unable to one-shot a course, so I am now asking it to generate it module by module. But even here, it is not doing a very good job. I muddle through the concepts that it has written, do a whole bunch of back-and-forth, which tbh is exhausting, and then rewrite everything in my words so it actually makes sense to another human being.

> I get exhausted reading LLM prose

So much this! If I see one more sentence with the words "genuinely" juxtaposed with "load bearing" my head is going to explode!

btw, I am building the tutorial here for anybody interested in this topic: https://github.com/avilay/learn-probml

"carefully architected solution" is not what they are saying. A "buggy big ball of mud" will not generally "do the right thing", if it did, it would not be a "buggy big ball of mud". Straightforward requirements do not imply straightforward solutions. In the early 2000s Facebook wanted a quick way to search for friends updates, a straightforward requirement. Turned out they had to build a full graph DB inside MySQL, not straightforward code at all.

The percentages are different for different types of SWEs, not all SWEs spend 80% of their time on XFN comms. I agree with your larger point that a non-trivial amount of a SWE's time is spent on XFN comms and team alignment. My argument is that if an individual contributor is not spending a majority of their time on problem solving and implementation (including maintenance of legacy code), they are not maximizing their potential. If they are spending 80% of their time on non-coding activity they are better suited for an Manager role (Engineering or Product). At the end of the day, coding is not hard only if you are a good coder to begin with. If you are a good EM/PM then people issues will not be hard (which coders often complain about).

There does seem to be a need for a semi-deterministic workflow where nodes in the workflow are agentic, i.e., powered by LLMs, but the workflow itself remains deterministic. Lets say my daily workflow is to download a PDF from some website, extract some information from it, and email that information. I cannot just ask an agent to script this workflow, because the website will keep changing, the PDF layout will keep changing, etc. I need a "download agent", an "extract info" agent, and so on which will examine the website, PDF, etc. and generate any one-off scripts everytime the workflow is run.

I have been grappling with a similar sentiment for a while now. I hate to think of myself as a Luddite, but I love the act of writing code. A blank VS Code in front of me is a beautiful sight full of wonder and possibilities :-) Here is how I rationalize the two - the pride and satisfaction I get from my work depends on how much effort I put in applying some previously acquired specialist knowledge, which itself took some decent amount of effort to acquire. This is why a beginner programmer is just as proud of their temperature convertor as a professional distributed systems engineer of their new 2PC module.

A quote that I read somewhere, possibly by Naval Ravikant, strikes at the heart of this - "a creation without a (human) creator is meaningless".

Neither the Github nor the HF repos have anything on the XiaomiRobotics-1 model or datasets :-( They have a bunch of assets (datasets, code, model, etc.) for XiaomiRobotics-0 however, but I don't know much about that family of models.

Sad to learn that future generations of programmers won’t benefit from learning from him directly. I took his SICP course a couple of years ago after numerous failed attempts at reading the book and even watching the MIT videos. David’s course was the first time the concepts really clicked.

Agree! Negotiating focus blocks both at home *and* work can be super helpful. Of course, this is not always possible. Without knowing anything about your situation, it might be useful to rule out burnout as a possible reason for loss of flow.

For sure, any task or activity that is hard enough and just outside our reach, can get us into flow state. The trick is in ensuring that it is the right kind of hard, it is not too hard, and we time box the activity/task. If you think of how to beat the boss fight in a video game even when you are not playing it, it is the "right kind of hard". For me, beating the boss fights in Elden Ring were too hard, never got into flow state in that game :-)

Debugging can be super fun as we eliminate possibilities and it feels like we are converging to a solution. There have been instances where Claude (Opus family) was not able to effectively debug and I had to step in and do it. But debugging an over-engineered library for example, can become very wearisome. This is when I am really thankful for having Claude Code, it is able to figure out the bug and its fix/workaround pretty fast. I can then get back to doing my main task instead of spending an indefinite amount of time stepping through sloppily written code.

Ok here are the flow related links. This was about 1.5 years ago when I was trying to figure out burnout and it turned out flow (or lack thereof) was closely related.

  * https://youtu.be/VbUFMYs0kXQ?si=xiNw4ZFlla8k-p7w  The person who gives this talk (Rian Doris) has a good newsletter that I still read. I just checked their website and it has gone in full commercial mode, so YMMV.

  * https://www.ted.com/talks/elizabeth_gilbert_your_elusive_creative_genius

  * https://www.amazon.com/dp/0465074871

  * https://www.betterup.com/blog/meaning-of-personal-values

Not being able to enter flow state is a very interesting observation. I've felt it too to the extent that I went down a whole new rabbit hole of what it means to be in flow state. Let me know if anybody here wants to know more, happy to post some links.

To answer your question - I discuss the approach with Claude Code (e.g., should I implement my own ACT model in JAX or PyTorch, Python or Rust or Julia, etc.). Then write the initial part of the code myself. Opening up a blank vscode is a simple joy of life I refuse to give up :-) I'll ask Claude for advice if I get stuck, it will helpfully offer to write that code for me, I obstinately decline. Eventually, I'll get bored of some minutiae or other, at which point I'll ask Claude to complete just that part of it.

Reminds of this couplet by Kabir (https://en.wikipedia.org/wiki/Kabir):

जब हम पैदा हुये तो जग हँसे हम रोये,

ऐसी करनी कर चलो, हम हँसे, जग रोये।

Translation: When you were born, you cried while the world rejoiced;

Live such a life that when you depart, you smile while the world weeps.

Reminds me of the "Named Tensor" experiment in PyTorch from a while ago. https://docs.pytorch.org/docs/2.12/named_tensor.html

I'll put in a plug for David Beazly's SICP course. While we didn't build a full Lisp interpreter, we built something similar over a week-long hands-on course. I believe his github for the course is private and only available to students. https://www.dabeaz.com/sicp.html

Thanks for your response! I totally get your point about the delay in getting the robot, I ordered mine from PartaBot and they did take a couple of weeks to get here. But when they did, they worked great out-of-the-box :-)

Will email you to compare notes.

> The author does explain his reasons for not using LeRobot in the post - although "I also use LeRobot for training and running baseline policies, and the vendor SDKs for the hardware.")

Ah! That is exactly what I use it for as well :-)

I am liking the SO101 - teleop and robot both work great. For sure, it is very easy to get started with. I was able to collect around 50 demos with them and train my first ACT policy within days of setting up the robots. Happy to share more detailed learnings from this if/when you get started with it. https://github.com/avilay/learn-robotics

Hey this is cool! I am doing something similar myself with the SO101 arm robot from Robot Studio using a patchwork of my own code and LeRobot. Would love to collaborate with you if you are open to it. You can find me on Discord as `.avilay`. https://www.linkedin.com/posts/avilay_lerobot-huggingface-ro...

Two things semi-parallelly:

  * Robotics Hello World: Objective is to implement ACT model to train my arm robot on simple pick-and-place tasks. Leaning heavily on HuggingFace's LeRobot library, but stopping short of using their model implementation and training loop. https://github.com/avilay/learn-robotics

  * Designing a new programming language: This is when I want to escape the annoyances of coding in Python and start daydreaming about a new language :-) https://github.com/avilay/kulfi

Can a kind soul write down their interpretation of the story? I didn't quite get it.

[Edit]: Thanks for all the explanations!

The point about papercuts adding up so resonates with me! I loved Zed initially and did find it more responsive than VS Code, loved the Zed Agent autocomplete, etc. However, I eventually and reluctantly went back to VS Code. The papercut that finally did it for me was [this open bug](https://github.com/zed-industries/zed/issues/36516) because of which I was not able to step into a packaged library's code when I was debugging my own code, this was in Python.

The two don’t have to be mutually exclusive. You can let the agent code and you review it, or vice versa. No different from being a team lead where you don’t write all the code, or even review each and every line of code, but you have a very firm grasp of the code base.