Tepix

I believe Intel has a winner with the upcoming Crescent Island GPUs with up to 480GB of LPDDR5 memory. A lot of SMEs will need local AI with large open weights models (i.e. >1.5TB of VRAM), but they don't need a watercooled 8x Nvidia GB300 box. Besides, it's difficult to deploy and very expensive.

Crescent island is air cooled and likely fast enough for a group of users running something like GLM 5.2 or Kimi K3.

My take is that there are at least four potential parties that can all be liable:

1. the creator for the LLM. In particular if neglicence or malice is involved. This can also be someone who did a finetune of an existing model.

2. the inference provider. Remember, a model can do harm just by creating tokens (for example cause someone to run amok or kill herself). Inference providers should do a minimal amount of due diligence when picking models.

3. the party that executes tool calls on behalf of the AI. They in particular need to have safeguards to prevent the model from attacking entities on the internet.

4. the user that does the prompting.

Hey, it's likely to be a docker container...

If your pack_number function builds the number up, the standard way to break it down is by extracting the least significant digit first using modulo and division. To get something that works well with SIMD we need a different approach. Instead of extracting the least significant digit from the bottom of an integer, we extract the most significant digit from the top of a fraction.

1. Convert to a fixed-point fraction: We scale our integer N into a fixed-point representation (e.g., using a 32-bit integer to represent the fraction). We do this by multiplying N by a precomputed reciprocal of 243.

2. Multiply by the base: Multiply the fraction by 3.

3. Extract: The integer portion of the result is your most significant trit.

4. Mask: Keep only the fractional remainder, and repeat.

The only operations here are multiplication, bitwise shift, and bitwise AND, i.e. perfectly suited for SIMD.

(in step 1 we replace the division with a multiplication by using the reciprocal. SIMD uses fixed-point integer arithmetic, not floating-point decimals)

There are a ton of super serious issues there. They are not getting fixed.

A relatively small and fast LLM (30b MoE) would probably suffice. It‘s something you can run locally with a nice GPU.

His post was about speed of deployment, not internet speed.

While the content was interesting, the AI-slop-stench was repelling.

Talking about AI (sorry!), perhaps an AI assisted screen reader could remove repetitive elements (it appends "(read only)" to every. single. field.) in a smart fashion? Does this already exist?

We're seeing AI being used to improve a11y in quite a few places: (Live) transcripts for video conferences, image to text (VQA, visual question answering) etc.

I like Podman, but what's up with that grey text colour? It looks ugly and the contrast of 4.96:1 makes it hard to read (does not reach WCAG AAA level).

Ideally Linux would add support for RDMA via Thunderbolt.

The Bosgame M5 128GB went up 52% in price ($1840 -> $2799), so not 2-3x. And they are in stock. Still, a very big price increase.

Thanks. Yes, I'm using Infiniband with CX3. Not seeing real speedups so far. I was wondering if upgrading to CX5 could help.

My favourite Zach game so far is Infinifactory. TIS-100 was also fun, until it started feeling like work.

On servers that are subject to the CLOUD Act. Expect no GDPR compliance.

Teachers teach people who don't know the answer yet. So they aren't wrong when you teach them. They will (gladly) accept the knowledge.

Once they know the answer, it gets more difficult to convince them that the answer they know is incorrect.

It's a cool video and I like the idea in general. The author mentions that the code runs in a sandbox. I'm surprised that WASM hasn't come up. You want the code to be platform agnostic anyway (it should run whether you start Outshell on Linux, macOS or whatever on different CPU architectures).

The pp speeds are really slow (50), I think there‘s room for improvement still.

The cheapest ones with 128GB were 1580€/$1840 as late as mid December.

What‘s the advantage of using ConnectX-5 Ex VPI NICs instead of much cheaper ConnectX-3 VPI NICs to connect two machines directly, other than PCIe 4.0 instead of PCIe 3.0? Can they offload more tasks when doing RDMA? Solid information is hard to come by.

Interesting. Reminds me of this gaston cartoon (at 5:51) https://youtu.be/OqIjS7Yb8y8

I guess Galileo came up with it first:

https://en.wikipedia.org/wiki/Galilean_invariance

Has anyone turned it into a browser game?

Agreed on the bragging.

Also I‘m not convinced about the whole cost issue. A nice server from a bare metal provider like OVH will be so much cheaper than the AWS equivalent, you can pay for a ton of traffic.

Well, how difficult is it to switch to something (much) cheaper like DeepSeek v4 flash?

What if we want to put something on paper today for it to survive as long as possible?

1-minute research:

Paper: 100% cotton rag or linen rag paper with alkaline reserve. Acid-free and lignin-free.

Ink: Genuine carbon ink applied with a classic dip pen.

Storage: ISO 16245 archival box, Less than 15°C, 30-50% humidity, dark, no oxygen exchange. Always store horizontally. Wear white 100% cotton gloves.

Printing: If you want to print instead of hand-write: Piezography carbon printing or pigment-based inks used by professional desktop photo printers, matte black or photo black ink, printed on digital Fine Art Archival Paper.

Place a single sheet of archival-grade tissue paper or glassine paper between every single page of your document

I think the key is to write something interesting that's worth preserving. That may be the most difficult part.

Any improvements beyond this?

M5 is 153GB/s, M5 Pro 307GB/s, M5 Max 614GB/s.

The article didn't state the M5 Ultra won't be released. It will probably provide 1228GB/s of memory bandwidth this year.

I think the rumors about a X4 Pro v2 with frontlight being released in June or July have been confirmed.

Do you feel they agree on many topics?

OK, then the labels (the numbers) are correct but the bars have the wrong length.

FUZIX is a better fit. https://codeberg.org/EtchedPixels/FUZIX