YuechenLi

7 ms·

61.
by YuechenLi 2mo ago
10Hz is still a 0.1s of latency, which doesn't seem like a lot until you realize that an Arduino typically runs at 16 Mhz, which is literally a million times faster than running a full LLM. So, yeah, for complex reasoning and sensory p
62.
by YuechenLi 2mo ago
Running a full LLM to actuate a robot is way too heavy, the minimum latency they can get down to is probably 1-2 seconds even with powerful GPUs, which is not very useful for practical robotics applications. Machine vision should definitely
63.
by YuechenLi 3mo ago
Not always. If an instruction is short and vague, then LLMs tend to fill in the blanks by guessing and the results are unpredictable. Semantic density is the key, prompts/instructions should be actionable and unambiguous over being ter
64.
by YuechenLi 3mo ago
LLMs don't really read documents like parsers do, from what I've observed, they behave like they first skim the document and find the section that they think are relevant to their task, they don't really try to read everythin
65.
by YuechenLi 3mo ago
Yeah, I still don't understand why they don't just use enums and have to invent a special "union" keyword instead. C#'s enums are extremely underpowered to begin with and the syntax they designed here way more convo
66.
by YuechenLi 3mo ago
So... the verbosity of Go without the compile speed of Go? Go is not a language I would associate with ergonomics, and this doesn't even have most of modern C#'s best features: does xUnit.NET work? Does immutable records work? Do
67.
by YuechenLi 3mo ago
I got tired of Codex/Claude not being able to write good UI reliably, so I made a TypeScript library toolbox to help with that. https://github.com/yuechen-li-dev/MachinaLayout.JS Pretty much just SwiftUI-like layo
68.
by YuechenLi 3mo ago
Just personal experience, I just find it way easier to do frontend work with Claude than it is with Codex.
69.
by YuechenLi 3mo ago
The answer is it depends. Claude's generally better at frontend and debugging tasks, while Codex is stronger at backend features and exploratory work. They have very different coding styles and thus very different strengths.
70.
by YuechenLi 3mo ago
Fair point, I phrased that too broadly, and you are right about the loose correlation. What I was gesturing at, badly, was more that Zig’s low-abstraction / explicit-by-default syntax tends to have you write more boilerplate-y code in
71.
by YuechenLi 3mo ago
>Combined with the Rust rewrite, ICU changes, and identical code folding, Bun's binary size shrinks by ~20% on Linux & Windows. People who are surprised by this probably has not seen what Zig code actually looks like. Zig's
72.
by YuechenLi 3mo ago
I'm sure you know that JSON are object literals in TypeScript, there isn't really even a serialization process that's needed there. The AST/IR can still be in JSON, but the authoring surface can be a restricted subset th
73.
by YuechenLi 3mo ago
>Instead of requiring verbose low-level parameters such as scales, axes, spacing, and layout. Ok, Microsoft is conflating two different things here: LLMs don't really care about code being low level and verbose, they can read things
74.
by YuechenLi 3mo ago
I mean, the key is to stop trying to one-shot everything: The main problem I found with LLM code is more that they always try to take the shortest path to the solution possible, so a lot of time Codex would write code that meets the require
75.
by YuechenLi 4mo ago
Yeah, one of the big problems with that is that Claude/ChatGPT doesn't perceive images the way humans do at all, so when you upload an image to them, it gets tokenized in some form. This is why most LLMs are really, really bad at
76.
by YuechenLi 4mo ago
Oh, I wasn't referring to emitting LLVM IR directly from Nim, but that you should consider adding an MIR layer between your Nimic subset of Python and Nim instead of emitting Nim from Python directly. Otherwise, compilation would get d
77.
by YuechenLi 4mo ago
If it's possible and you intended for it to be a long-term project, I would suggest looking at designing an MIR for the Python to lower to and then lowering it to Nim, and that is probably the most valuable feature you can add to Nimic
78.
by YuechenLi 4mo ago
So... has Google provided a Codex/Claude Code equivalent to Gemini yet? I would like to use Gemini for coding tasks, but that's kind of difficult to do as I don't even know how to get Gemini to even "clone this repo and
79.
by YuechenLi 4mo ago
I honestly think Mojo would be better served if it is just a high-level language for GPU programming that compiles down to PTX with clear Python/Rust interop boundaries instead of trying for the "one language, multiple computation
80.
by YuechenLi 4mo ago
I would say it is definitely a form of context, but when people think of LLM context windows in terms of coding is more technical context related: "what has been done before, what's the coding task at hand." etc. However, I t
81.
by YuechenLi 4mo ago
Learning computer science theory is probably going to be more important for the future than writing code yourself, because LLM coding is here to stay, and the human's job is going to be guiding the LLMs effectively: the AIs are writing
82.
by YuechenLi 4mo ago
Strangely, I found that LLMs responds better to philosophical explanations alongside instructions when writing code than simple imperative tasks of "do this". For example, if you tell a frontier model "This is the feature I&#
83.
by YuechenLi 4mo ago
>Are you sure? I'm pretty sure .deb and .rpm packages both allow that Learned something new today. Thanks. I think the other significant issue with the NPM ecosystem that makes it bad in particular is NPM's dependency managemen
84.
by YuechenLi 4mo ago
This seems to be primarily a problem with NPM, since it's the only package manager that I know of that allows for package authors to essentially run arbitrary post-install scripts silently package install. Shai Hulud/Mini Shai Hul
85.
by YuechenLi 4mo ago
>I'm proud to say, I've managed to slay one of the giants, the only successful effort to my knowledge, and this only happened last year. But I did so by basically delegating to another of the giants; they turned out to be twins
86.
by YuechenLi 4mo ago
You should be skeptical of industrial use case as well. Allowing any device with internet access and a camera on it onto the floor of most electronics manufacturing facility is a nonstarter, phones, smart glasses, or anything of that sort.
87.
by YuechenLi 4mo ago
They are extremely common actually. Why do you think the standard iPhones only does USB 2.0 transfer speed? The high-speed signal pins are simply not there, but the connectors themselves are still standard compliant. Lower transfer rate mea
88.
by YuechenLi 4mo ago
The industry will realize that while getting LLMs to write code is easy, getting LLMs to write good, production ready code is a skill all on its own, which simply must be done by a human and is not automatable to an LLM in any sense effecti
89.
by YuechenLi 4mo ago
Not quite right, USB-C ports are generally cheaper nowadays because they are smaller, consumes less material for plastic/metal, more easily automatable production wise in terms of tooling, and scale for them is a lot higher because of
90.
by YuechenLi 4mo ago
Surprised that they have 4 USB-A and only 1 USB-C. With their power profile, Steam Machine should be powerable by a single USB-C cable on extended power range which should reduce the need for the power supply altogether and greatly simplify

More ›