EgoIncarnate

The WSL2 GPU passthrough is more like a virtual GPU than KVM style device passthrough. I believe it's effectively a device specific linux userland driver to device specific windows kernel driver with a linux kernel shim bridging the too. If I recall correctly, the linux userland drivers are actually provided by the windows driver.

not "an H200", "In the table above, tensor parallelism is compared to pipeline parallelism with each across eight GPUs"

Bolt.new may be technically open source, but it seems to be dependant on closed source StackBlitz webcontainers? Not truly open source IMHO.

Clickbait garbage. He didn't acheive fusion, he created plasma.

https://x.com/hud_zah/status/1827096141814714705

The phrase "people like him should be forcefully removed from society" is problematic because it can be seen as inciting violence. I don't think that has a place on HN, regardless of how you feel about the person.

(Note: To flag a comment, click on the comment time/date, and then flag)

Per OpenAI, it's a summary of the chain of thought, not the actual chain of thought.

I think you might be confusing CastAR with the C64 toy stuff she did early on. That was initially FPGA and then converted to custom gate array. If I recall there was an issue in that the FPGA initialzed it's memory and the custom silicon didn't which lead to some heart pounding moments when it failed initial bring up.

I don't think CastAR/Tilt5 isn't doing enough units to justify custom silicon. I believe they settled on the Intel Movidius VPU for the bulk of the glasses "brains".

Title implies the demo video was faked. It looks like Claude created code similar with a result similar/identical to one used in a shader tutorial.

Fun for the win! I suspect most of the other suggestions, even if possible, wouldn't have generated this level of interest. Well done on your Rube Goldberg method!

You're missing some context, it's not bitonic sort itself that would present an issue with GPUs, it's the "with immutable tree rotations" part, which in a naive implementation would imply some kind of memory management that would have trouble scaling to thousands of cores.

I think people might interpret something claiming to be the "Future of Parallel Computing" as something that is just waiting on adoption. Perhaps "Towards the Future of Parallel Computing"...

Much of the software for the Xerox Alto was orginally written in BCPL.

Also the original AmigaDOS was written in BCPL. This caused some complications later when they decided to use C as the primary system programming language. In AmigaDOS BCPL pointers are "word" (2 byte) addressed, which was a common form of addresses for computers of the time (inherited from non-power of two sized word computers such as the 36-bit machines of the 60s and 70s). C uses byte addressing, so it requires a bit shift to convert between BCPL's word addressed pointers and C's byte addressed pointers.

Which model did you use?

Can you explain? My intution and rough calculations say you are wrong.

ChatGPT agrees "To determine the long-term annual sales of phones for a company with a constant customer base of 1 billion, where each customer upgrades at different intervals, you first need to know the average replacement interval, T years. The annual sales can then be calculated as 1,000,000,000 divided by T. For example, if customers upgrade every 2 years on average, the company would sell approximately 500 million phones per year. This approach assumes the replacement intervals are consistent or average to T years across all customers."

The chip restrictions may accelerate things, but China was already moving in this direction. The chip restrictions were relatively recent, and the chip production capabilities took years to develop. The chip restrictions mainly made it obvious how far along China already was.

I tried to explain that AI was ruled not copyrightable if a human wasn't meaningfully involved. Meaningfully involved being a bit blurry, with art generated from a prompt being the only explict thing so far ruled not copyrightable. Do you have a different take?

Too much energy is used (and heat generated) at high frequencies. For something like PCIe, you don't need to double the frequency for the whole circuit to double the frequency of the link. You can double the front/back and then double the parallelization of the rest of the circuit. Most of the circuit can still run at a more energy efficient frequency. Potentially it was possible earlier, but the circuit size made it cost prohibitive.

It's an April Fool's joke.

  "jwt NETGATE 3 days ago
  @johnpoz said in Serious ....:
    I don't recall them doing things before about April fools
  we (I) did, back in 2017 https://www.netgate.com/blog/building-a-behemoth-router
  I even pointed to it in today's."

https://forum.netgate.com/topic/187100/serious/6?_=171214303...

Purely AI driven art is already defacto public domain in the US in that it is not copyrightable. See https://www.theverge.com/2022/2/21/22944335/us-copyright-off...

There is a spectrum of what AI generated means. At one end is simply photoshopped pictures. In many cases these were touched by AI, in that some algorithm decided on the specific pixels, but the image was largely directed by a person and is generally considered copyrightable.

On the other end is the uncopyrightable case where a person just typed in a prompt and out comes art.

Where the line is in between has yet to be decided. What if I spend hours tweaking an image generated initally by an AI. At what point does it become copyrightable? What if I spend hours creating a complex graph? That's similar to source code of a computer program, and the output of compilers is protected... From the other direction, at what point does something copyrightable manipulated by AI lose it's protected status?

https://rr-project.org/ had the same problem. They use the retired conditional branch counter instead of instruction counter, and I believe then instruction stepping until at the correct address.

Each token generated may only use a subset of the parameters (86billion instead of 314billion), but the next generated token might use a different subset. If it's anything like Mixtral, it will switch between experts constantly. It helps with memory bandwidth, but all the parameters still need to be in RAM or it would be unbearably slow.

Disabling SIP also disables your ability to run iOS apps on Apple Silicon :-( It might be for copy protection reasons. iOS apps are partially encrypted and without SIP you could probably access the decrypted version.

My quick test of running multiple command prompts and running free showed that each one had a full set of memory. It's not multitasking, and I believe it assumes that the tasks are well behaved (not modifying DOS internals), so tasks don't really need to do anything special if they are well behaved.

DOSSHELL creates a swap file DOSSWAP.SWP, so all tasks don't need to fit in memory.

It sounds like there was a config option available to disable the signature addition to the image file ( https://github.com/simh/simh/issues/1059#issuecomment-108689... ). I could see a benefits for having an embedded image signature for preservation and corruption detection issues.

I don't think complaining about the design is toxic, but recruiting uninvolved people on twitter, and harassing out of band certainly is. Also reading the bug thread it seems the person with the issue wasn't the same as the one who instigated the harassment. (https://github.com/simh/simh/issues/1059#issuecomment-108675...)

It seems Mark Pizzolato was at the time the primary maintainer and recent contributor (by a large margin) (https://groups.io/g/simh/message/1501).

Mark made a change to how the virtual disk system worked that broke an esoteric use case (using dd to image multiple disk images to an SD, without an actual file system). The user didn't want to use a workaround (specifying the transfer size in the dd command), and proceeded to use twitter to get people to harass Mark. (https://www.bentasker.co.uk/posts/blog/the-internet/toxicity...)

Mark got sick of the harrasment and decided to take his toys and go home.

Toxic user prompts toxic reaction by developer.

The files uploaded in a code interpreter session are available for use by the python interpreter, but are not automatically part of the context, which is limited to 8k tokens in the ChatGPT GPT-4 Code Interpreter model.

Thanks for putting to this together.

I have a suggested modification. You are mixing references in your document.

Re: '~$1 million: Cost to train a 13 billion parameter model on 1.4 trillion tokens The LLaMa paper mentions it took them 21 days to train LLaMa using 2048 GPUs A100 80GB GPUs.'

The LLaMA-13B model took 2.75 days of 2048xA100 (135,168 GPU-hours) with 1 trillion tokens. The 21 days for 1.4 trillion was for LLaMA-65B.

I would suggest using the LLaMa-13B numbers since those are the most relevant for this section, or at least modify "21 days to train LLaMa" to "21 days to train LLaMa-65B" for clarity.

The largest PaLM 2 model is smaller than 540 billion parameters of PaLM 1 (let alone 540 trillion!). From the PDF "The largest model in the PaLM 2 family, PaLM 2-L, is significantly smaller than the largest PaLM model but uses more training compute."

The official specs for Raptor Lake don't include AVX512 support, so most likley this is just something that slipped through on your system. It might be "corrected" in later steppings or updated firmware. Some early 12th gen systems had unfused CPUs with AVX512 available on the performance cores and firmware that didn't disable it. https://www.intel.com/content/www/us/en/products/sku/230500/...