HNHacker News
TopNewBestAskShowJobs

jxy

1,282 karma · joined March 2, 2016

[ my public key: https://keybase.io/jxy; my proof: https://keybase.io/jxy/sigs/kwfjb0xrYlVnuHbZYFc_Q6_H4uIxkhq83KvqSEVUFho ]
submissionscomments
jxy··on Only9Fans
Macbook trackpads work fine combined with control/option/command as three buttons. I put the metal screen for devdraw together initially.

It should work similarly on X11, though I never tried.

jxy··on Nearsightedness is at epidemic levels – and the problem begins in childhood
> if you do a lot of near-work it can think you are still hyperopic, causing further myopia progression.

Does this mean we should let our eyes see blurred images so our eyes would feel like they are seeing far away stuff and stop axial eye growth to stop myopia. This is in direct contrast to the claim in the post

> a few minutes a day with glasses or contact lenses that correct for blur stops the progression of myopia

which believes corrective lenses that correct for blur would stop the progression of myopia.

jxy··on Thefastest.ai
No prompt length? For practical purposes, the prompt processing time would far more important.
jxy··on Phi-3 Technical Report
This inductive logic is way overblown.

> Incredible, beat Llama 3 8B with 3.8B parameters after less than a week of release.

Judging by a single benchmark? Without even trying it out with real world usage?

> And on LMSYS English, Llama 3 8B is on par with GPT-4 (not GPT-4-Turbo), as well as Mistral-Large.

Any potential caveat in such a leaderboard not withstanding, on that leaderboard alone, there is a huge gap between llama 3 8B and Mistral-Large, let alone any of the GPT-4.

By the way, for beating benchmark, "Pretraining on the Test Set Is All You Need"

jxy··on usbredir: A protocol for sending USB device traffic over a network connection
it may not be practical given the possible high latency.
jxy··on Mistral AI Launches New 8x22B MOE Model
It's "Command R+". "Command R" is a smaller model.
jxy··on Rule-based NLP system beats LLM for analysis of psychiatric clinical notes
FLAN-T5-XL, a 3B model, to be precise.
jxy··on Andres Freund and the xz backdoor
A more level-headed report with less fluff from the economist: https://www.economist.com/science-and-technology/2024/04/02/...

https://archive.ph/rdxhb

jxy··on A rudimentary simulation of the three-body problem
Probably try to implement a much better integrator. Some energy preserving higher order integrator should serve a lot better.
jxy··on Lightning Whisper MLX
How did you run whisper.cpp? Can you share the command line? What machine did you use to run that benchmark figure?
jxy··on DBRX: A new open LLM
I wonder if they messed with the scale or they messed with the bars.
jxy··on DBRX: A new open LLM
So some parts of it copied from Claude: https://news.ycombinator.com/item?id=39649261
jxy··on Manipulating Chess-GPT's World Model
The previous blog post [0] mentioned they were using stockfish with 0.1 seconds per move.

[0] https://adamkarvonen.github.io/machine_learning/2024/01/03/c...

jxy··on Show HN: Free Plain-Text Bookmarking
If you ever used w3m, it uses a local bookmark file. Something like this:

    $ cat ~/.w3m/bookmark.html
    <html><head><title>Bookmarks</title></head>
    <body>
    <h1>Bookmarks</h1>
    <h2>Searches</h2>
    <ul>
    <li><a href="https://www.google.com/">Google</a>
    <!--End of section (do not delete this comment)-->
    </ul>
    <h2>Time Sink</h2>
    <ul>
    <li><a href="https://news.ycombinator.com/">Hacker News</a>
    <!--End of section (do not delete this comment)-->
    </ul>
    </body>
    </html>
jxy··on Nvidia CEO Jensen Huang announces new AI chips: ‘We need bigger GPUs’
curiously, what real world usage actually uses FP4? AFICT, most of the LLMs still use BF16, and even the quantizations down to 4bits and 2bits end up back to 16bit or INT8 for actual computations.
jxy··on Apple exploring a partnership with Google for Gemini-powered feature on iPhones
Apple's in-house map stumbled badly at the beginning. Apple has been using Google search since the very beginning. It's reasonable that if we can't run a gpt-3.5 level model on a macbook air for 20 hours without charging, it's much better to use existing infrastructure and existing models.

The most important question is what kind of product Apple is creating.

jxy··on Grok
OpenAI still seems to be at the top, except for Anthropic, who may be close, in terms of the capabilities comparing gpt-4 and claude-opus.

This Grok-1 is a large model (~314B), which matches gpt-3.5 released 2 years ago, and at about the same level of much smaller models like, mixtral (~47B) and qwen-1.5 (~72B). Do you think it's competitive?

jxy··on Figure 01 robot demos its OpenAI integration
Can we stop this kind of misinformation? Training a model to map a token to individual letters are no harder than training a model to be fluent at English. Arithmetics with small number of digits are achievable as well. You can just try a small 7B model yourself. If you don't know where to start, try the mistral instruct v0.2, and this is how it goes,

> [INST] Spell out the following word letter by letter: margorczynski [/INST] m - a - r - g - o - r - c - z - y - n - s - k - i

> So, the word "margorczynski" spelled out letter by letter is: m-a-r-g-o-r-c-z-y-n-s-k-i.

The text between `[INST]` and `[/INST]` is the input. The text after `[/INST]` is the output.

jxy··on Claude 3 System Prompt
She stated "Let me break it down." Did she break it down somewhere? How can we see it without X account?
jxy··on How do computers calculate sine?
it says "Copyright (C) 1993 by Sun Microsystems, Inc. All rights reserved."

and the directory name is literally: `msun`

jxy··on How do computers calculate sine?
http://git.9front.org/plan9front/plan9front/HEAD/sys/src/lib...
jxy··on How do computers calculate sine?
It's much clearer if you read one of the source code of the libm.

Plan 9: https://9p.io/sources/plan9/sys/src/libc/port/sin.c

Freebsd: https://cgit.freebsd.org/src/tree/lib/msun/src/k_sin.c

jxy··on FreeBSD has a(nother) new C compiler: Intel oneAPI DPC++/C++
The value of Intel's oneAPI tool chain is that you can write code that runs on their GPU, much like Nvidia's cuda or nvhpc.
jxy··on FreeBSD has a(nother) new C compiler: Intel oneAPI DPC++/C++
> There is a workaround: I installed the compiler using the same command on my WSL

Actually you only need to have a relatively complete linux under /compat/linux. You could even chroot into that and do the oneapi install.

The trillion dollar question is: can you get the intel discrete gpu to work, compiling for and running on the gpu?

jxy··on Kagi and Wolfram
> welcome Stephen Wolfram to Kagi’s board of advisors

This is interesting. I guess we will soon read a longwinded post about it from the first person perspective.

jxy··on The Era of 1-bit LLMs: ternary parameters for cost-effective computing
The matrices (weights) are ternary.

The vectors are not.

jxy··on Phind-70B: Closing the code quality gap with GPT-4 Turbo while running 4x faster
Yeah, 4bit would take 35 GB at least. 16bit would be 140 GB. I'm more interested in how Phind is serving it. But I guess that's their trade secret.
jxy··on Phind-70B: Closing the code quality gap with GPT-4 Turbo while running 4x faster
How many H100 GPUs does it take to serve 1 Phind-70B model? Are they serving it with bf16, or int8, or lower quants?
jxy··on The killer app of Gemini Pro 1.5 is using video as an input
So that 3-4 mins at 1FPS means you are using about 500 to 700 tokens per image, which means you are using `detail: high` with something like 1080p to feed to gpt-4-vision-preview (unless you have another private endpoint).

The gemini 1.5 pro uses about 258 tokens per frame (2.8M tokens for 10856 frames).

Are those comparable?

jxy··on Things I don't know about AI
> architectures like RWKV, Mamba, and various others, which incur cost that is linear in context length, O(N)

These are actually O(N*M) where M is independent of inference time context length, but it controls the actual amount of information that gets retained during inference. Ignoring that factor of M is misleading.

← PreviousPage 2 of 18Next →