zellyn

If you do decide to wait longer, be aware that there are hilarious differences geographically. When we had our first kid in SF, the other dads pushing swings were around the same age as me (fortyish). Moving back to Georgia… oh my god, the parents of kids my second kid’s age were babies! (There are “graybeard dad” Facebook groups etc., but the average vibe is way different)

These? No, they’re too short, and often simplified to the point of being highly specific to exactly the given view. Many shaders on shadertoy will read mouse position; what the author does with it is up to them!

You could certainly try cloning the examples and messing with them: turning some number between 0 and 1 in the code into (mouseX/screenwidth) could prove interesting!

Most of these sites have a culture of learning from modifying and remixing other examples; just give credit where due

"Gentle introduction" is a pretty typical phrase, meaning more or less that the acceleration curve of learning is a gentle slope. That one doesn't irk me.

"Gentle study" is a bit unusual.

"Gentle ping" or "Gentle nudge", and especially "Gentle reminder", are starting to get a bit more annoying, but I think the intention there is try to convey the opposite of being demanding, but over a textual medium. ¯\_(ツ)_/¯

This is so meta it’s hard to ground it enough to understand any of it. I’d love to see a tiny worked example of, say, a billing department and an app that represent “Customer” in different ways, and how to model each, and most importantly, what practical benefits result. Ideally, the ideas of what a “customer” is would be different enough to differ in cardinality: e.g. in Billing, a customer is a business, and in the app, a customer is something that started out 1:1 with business but that broke down over time with co-ownership, assistants performing work for owners, etc.

I think sleep is possibly the most important part. I also did a sleep study and started using a CPAP machine: might want to look into that if it’s a possibility…

Yes. LLMs

This happened to me during the pandemic. High (insupportable) stress, plus a year or two of significantly too little sleep, plus social isolation.

It was terrifying. I had just seen my mom taken by Alzheimer’s (or something similar) way too young, and thought my life was over.

All the tests coming back negative, making a point of hanging out with friends, returning to the office, and getting enough sleep seemed to help, and it slowly started getting better.

Claude code does it. Goose does it. Cursor Composer (I think) does it. Thorsten Ball’s post does it in 400 lines of Go code: https://ampcode.com/how-to-build-an-agent

Basically every other IDE probably does it too by now.

Ask people you worked with who have a good impression of you for a list. Better yet, book an hour of video chat and talk through it with them. They will have a lot of examples. Write them down.

I mostly blundered my way through this through this recent round of tons of interviews, and by the end of it, my "Tell me about a time when…" doc is actually what it should/could have been at the start!

Looks like most of the comments here are about the use as weapons and the possible dangers. I believe "Slaughterbots" is the canonical sci-fi video on the subject, and it appears to be aging pretty well. Unfortunately…

https://www.youtube.com/watch?v=O-2tpwW0kmU

I think of “fits on the overpowered M1/2/3/4 64GB MacBook Pro my employer gave me” as the dividing line. We’re getting to within spitting distance of models that can code well at that size.

At least in the first example, the optionality is directly encoded in the types, so no assumptions have been lost.

I love that name, and will definitely steal it!

I definitely think breaking out ChatGPT on its own is too fine-grained. At the very least, one should lump all LLM training and use, or perhaps all deep learning training and use together.

I suspect at a "sensible" breakdown (trying to avoid the "How long is my coastline?" problem), which is presumably something akin to Zipfian, the main uses will actually account for most energy usage.

If anyone can find a breakdown at the level of granularity that "all deep learning" would make sense, let me know: so far my Googling has led primarily to either detailed breakdowns of energy _sources_, or high-level breakdowns of use at the level of "industry", "agriculture", or "residential use".

It apparently has a gearbox in the hub.

It's hard to get by with only 1.08 billion dollars…

Not that I think giving away most of his money isn't admirable. But I think he'll still qualify for "he died rich".

That sentence is the Apple II hardware in a nutshell! save a chip or two but completely scramble the video buffer? Absolutely!

The Commodore 64 is _so_ much better for graphics and sound it's not even funny, but if you look at the timelines, it came along a _while_ later than the Apple II.

Looks fun. For OS, it just says "11" — does it run ok on Mac OS?

I really like this framing. It also reinforces my opinion that the thing most like the proverbial AI that turns the entire world into paperclips by far humans. It's a bit fascinating if you look at it from a psychological / mythopoeic point of view: are villains _always_ the evil part of ourselves, even when they're not human?

Exactly. You want to come close to maxing out your RAM for model+context. I've run Gemma on a 64GB M1 and it was pretty okay, although that was before the Quantization-Aware Training version released last week, so it might be even better now.

Weird to give MacBook Pro specs and omit RAM. Or did I miss it somehow? That's one of the most important factors.

“If you give me six lines written by the hand of the most honest of men, I will find something in them which will hang him.”

Oh, that's very nice. Thanks!

Oh, that's really nice. Did you capture the responses from the LLM. Presumably it has some kind of special syntax in it to initiate a tool call, described in the prompt? Like TOOL_CALL<mcp=github,command=list> or something…

Yeah, I got that from reading the Ghidra MCP (very instructive, strong recommend), but I'm curious what the LLM needs to output to call it. I should go read Goose's code or instrument it or something…

Oh, that's really nice. I'd also like to see what syntax the LLM uses to _trigger_ these calls, and what prompt is sent to the LLM to tell it how to do that.

I should probably just go read Goose's code…

It’s frustratingly difficult to see what these (A2A and MCP) protocols actually look like. All I want is a simple example conversation that includes the actual LLM outputs used to trigger a call and the JSON that goes over the wire… maybe I’ll take some time and make a cheat-sheet.

I have to say, the endorsements at the end somehow made this seem worse…

It's also the perfect excuse if you get caught lying. “Oh shit, I ran it through ChatGPT one last time after proof-reading, and forgot to review the output carefully. Sorry!”

Andrej Karpathy goes into quite some depth about how tokenization works here: https://www.youtube.com/watch?v=zduSFxRajkE

tl;dr many of the LLMs use byte-pair encoding to create tokens. You take a set of documents, and then form tokens by repeatedly merging the most common pair of tokens. The initial set of tokens is 256 raw bytes. And the text is typically represented in utf-8.

I expect that although the LLMs can understand arbitrarily but cleanly offset unicode code points by (eventually) noticing the final byte of each sequence, they would do markedly worse on actually processing and completing on them, because they will not have been reduced to the normal set of tokens. However, if the text is actually output cleanly converted, either in internal thinking tokens or in the beginning of the response, they should do fine.

Understanding tokenization is surprisingly useful, even if that video seems awfully long to devote to such a tedious subject. Even Karpathy doesn't like it!

I think Casey Muratori calls this "compression". And yeah (see other child comment), he does it in C++ :-)