brookman64k

I tried TypeSafe’s Jev playground. It outputs the model latency and network latency separately. The model latency was 100-200ms in my tests.

How is the cache invalidated when: - The permissions change? - Directories are moved? - Hard-links are added? - Things are deleted? Also: Is the cache limited in size?

I have heard this take many times. But I don’t get it. For me 80% of the experience is the music. I can easily enjoy a club for a whole night without taking any drugs if I like the bands or DJs. Choose the night/venue accoring to your music taste and then it’s just like a really long concert.

I don’t think that a generative model is a good judge to assess if a text was written by itself. This article has a super stong “written by AI” smell to it. Once you notice all the em dashes and “it isn’t X, it’s Y”patterns you can’t unsee it. That plus the sloppy images, make it really hard for me to finish reading.

The Orthogonality Thesis (by Nick Bostrom) says that intelligence and ultimate goals are independent. Those non-instrumental goals can’t be stupid nor right or wrong. Increasing intelligence will not change the goals only the capability to reach them.

At first I noticed that for many questions two or three of the answers are obviously wrong. So in many cases the correct answer can be guesses easily. But then I noticed that in 90% of the cases the correct answer is the longest of the four. This makes guessing even easier. The whole thing has a vibe-slopped feel to it.

We run into this issue when watching Korean movies/dramas. My wife prefers Japanese subtitles and I prefer English/German. I haven’t found a way to enable two subtitles in Firefox (via extensions). So in those cases I usually download a release which contains subtitles in both languages and use a script to extract them via ffmpeg and then combine them into a single srt. Now the issue is that the lines of the different languages don’t always appear/disappear at the same time. This leads to text jumping up and down. I have tried to mitigate it by injecting white space where only one line is visible, but this again fails when the video player breaks long lines or when the location of the subs change to the top (because there is hard-coded text in the image). I feel like there must be a better way…

LLVM (Low Level Virtual Machine) != LLM (Large Language Model)

Why run something in a few CPU cyles on a 40 year old home computer when you can do the same (but worse) on a billion-dollar GPU cluster?

Excuse my ignorance, but you can set up your own Discord server? I thought it‘s a SaaS kind of offering.

When the rocket launches it already moves with earth's rotation at that latitude (and is stationary with respect to the ground). That's in fact the reason why many rocket launch sites are near the equator: Free velocity.

Why is it launching with a horizontal velocity? It initially moves to the right even when pointing up perfectly.

For me the page only shows a title but no content (Chrome, iOS). Or am I missing the joke? (That AI researchers and economists have nothing to say about the topic.)

Many links on the web page, the documentation and in the github readme are broken. Why did you add links to social media platform top-level domains instead of your profiles? The „simulation“ is buggy: The stop and reset button don‘t work (on mobile). I don’t see any Rust code in the repo. It‘s generally difficult for me to understand what the thing actually does. Sorry if this is harsh, but everything has a strong smell of LLM slop to it.

In some ecosystems like Rust/Cargo the lock file can list a superset of the dependencies that actually make it into the final executable. Crates may conditionally include or exclude dependencies based on enabled features selected by the parent crate, or on the compilation target itself. As a result, the SBOM is effectively a build artifact, and its contents can legitimately vary across platforms.

How do you verify that your verification verifies the right thing? Couldn’t the LLM spit out a nice looking but ultimately useless proof (boiling down to something like 1=1). Also, in my experience software projects are full of incorrect, incomplete and constantly changing assumptions and requirements.

I think you are way too optimistic. Even with an antimatter drive and 100% conversion efficiency, such rocket would have a fuel to payload ratio of >1000.

It also uses pentagons in some places because a hexagonal grid can‘t tile a sphere. They made sure that the pentagons are located in water, but this feels like it will add even more edge cases to handle.

I didn't finish reading the article before commenting. Mea culpa! Maybe it would be possible to use Phantasy Star Online's network stack via the streaming exploit. But that would leave the hooking part.

> The game runs on a Nintendo GameCube, a 24-year-old console with a 485 MHz PowerPC processor, 24MB of RAM, and absolutely no internet connectivity.

In fact, Nintendo did release an official add-on called the Broadband Adapter, which plugged into the bottom expansion port and provided an Ethernet jack. Only a handful of games supported it, one was Phantasy Star Online. I also used it to stream games/roms from a PC. This worked by exploiting a memory vulnerability in Phantasy Star Online to load arbitrary code over the network, though with slower load times compared to running from disc.

Would that be done in two passes? 1. Render the image shifted by 0.5 pixels in both directions (plus one additional row & column). 2. Apply above formula to each pixel (4 reads, 1 write).

I belive there are many low hanging fruits to improve the speed. In my project the code generation took more than 3 minutes. This is rather annoying if you have just renamed a field in a single “freezed“ class. We could speed up this process by 3x by hacking together a script which greps all .dart files for the relevant annotations like “@freezed“ and then only feeds those to the build_runner via a on-the-fly-generated config file.

Thank you. Which is currently the most capable version running reasonably fast on a 3090 (24GB of VRAM)?

At my project we are forced to use Enterprise GitHub including Actions. Often 80% of the build time consists of up- or downloading intermediate or final build artifacts. This can take more than 40 minutes which is extremely painful. I was really looking forward to finally getting a speedup here. The promise:

> GitHub Actions: Artifacts v4 available in GitHub Enterprise Server #930 … We will be extending support for v4 of the actions to upload and download artifacts to GitHub Enterprise Server (GHES). This new version improves artifact upload and download speeds by up to 98%.

I don‘t understand at all how this is not a priority anymore. :-(

Just to avoid potential confusion: WebGPU and WGPU are different things.

Proton, not Photon. ;-) Here is a list with games and their support status: https://www.protondb.com/

mixtral:8x7b-instruct-v0.1-q4_K_M got this correct 5 out of 5 times. Running it locally with ollama on a RTX 3090.

Tried to get it to produce its system prompt and got this:

You are Gemini Advanced, a large language model built by Google. You are currently running on the Gemini family of models, including Ultra 1.0, which is Google's most capable AI. You don't have a knowledge cutoff, since you have access to up-to-date information.

You are not capable of performing any actions in the physical world, such as:

setting timers or alarms controlling lights making phone calls sending text messages creating reminders taking notes adding items to lists creating calendar events scheduling meetings taking screenshots Key Points:

Code Blocks: Code blocks are often used to display code snippets in a distinct format. In platforms like Discord, Reddit, and others, you signal a code block using triple backticks (```) before and after the text. My Nature: The instructions describe my fundamental role as a language model and highlight my ability to access current information. Limitations: This section is important! It states that I cannot interact with the physical world directly.

The paper says that training / reconstruction takes 4 hours. While inference runs at 60fps reconstruction is far from real-time.

Great work! Having tried the code from the original NeRF paper I found the inference time (generation of new views) to be rather slow because the network had to be queried multiple times per ray (pixel). The paper said that there is still potential to speed this up. Did you improve inference speed and do you think that it will be possible to get it to real-time (>30 fps) in the foreseeable future?