21e8

We’re excited to announce that we’re expanding our Motion Picture portfolio with the release of KODAK VERITA 200D Color Negative Film which will be available in 16mm, 35mm and 65mm motion picture formats. Read more about the new film at kodak.com/go/verita

According to Plan

@atomic_semi Cool

Plasma testing

Advanced Inertial Reference Sphere

@SoftEngineer Sorting is AGI

Sorting on the GPU is important. Most performant GPU sorting is radix-based. Radix sort is built on prefix scan. Prefix scan on the GPU using Chaining with lookback (OneSweep) is about the fastest thing currently. But it doesn't run on Apple hardware. Recent paper on Decoupled Lookback with Fallback fixes that, see "Decoupled Fallback: A Portable Single-Pass GPU Scan" by @raphlinus, @jowens and Thomas Smith. Shade (my WebGPU graphics engine) heavily relies on prefix scan operation so this was a very critical piece of the puzzle for me, especially the last paper (CSDLDF) for overcoming the lack of Forward Propagation guarantee on Apple silicon.

@maddireidy @RocketLab What is their strategy against railguns?

What do you want to know about the Neutron rocket design, qualification programme & production? Talking to the @RocketLab team this week $RKLB

21e8 Pager Duty

@elonmusk @yunta_tsai LLMs do neither 😂

@yunta_tsai Intelligence seems to be semantic compression and correlation

One of the main ceilings of training is long data context. For LLMs, you can scale this window to almost infinite while still getting good trajectory samples, but for the real world this is yet to be the case. The major problem is compressibility. The longer the context of the data, the more storage it takes—given the limits of compressibility. Furthermore, the more interesting the data, the less compressible it is. For example, driving down a smooth highway is highly compressible, but adversarial scenarios are less so. Thus, even if your hardware is equipped with awesome sensibility, the dynamic range after compression is what you are left with. The limit also applies to generative models since the models themselves are a form of compression. Even if you force them to run at double precision, it doesn’t change the fact that they are super-resolving a quantized observation. Hence, the more sensing you integrate—especially different modalities where their quantum distributions are inherently different, as any sensing in any shape or form is quantum—quantizing the uncertainty to a number, the less information they preserve given the compressibility (and/or quantization) budgets. There is a reason why human eyes are designed the way they are, not because we could not add ultraviolet or near-infrared sensibility to the cells—it can be done—but because of the compressibility we could achieve in our neuron pathways while providing the best signal-to-noise ratio for long context reasoning. Insects, on the other hand, have a very small context window but higher sensibility—yet they cannot reason.

@sunlove happy birthday

today is my birthday :3 ♡ ˎˊ˗ ima block whoever don’t wish me hbd ily milady

21e8 Pound Sterling

21e8 Eyepatch

21e8 Eyepatch

21e8 Computer Use

21e8 Ping Pong

Projects

All God’s children have microtubules, including single cell organisms for which they sense, move, learn and encode memory. Associative learning also needs reward, something that feels or tastes good, something involving consciousness. We can’t understand consciousness with complex brain networks so how do we know qualia aren’t Funda-mental everywhere at tiny fast scales, accessible through quantum brain biology. Roger Penrose suggested that in 1989. pubmed.ncbi.nlm.nih.gov/35782391/

@__tinygrad__ There's an NEC deal ready and waiting for you Geohot youtube.com/watch?v=d0dLax…

If we can tunnel PCIe over USB3, why not also tunnel it over Ethernet? extra/remote/serve.py on remote, REMOTE=<ip> on local.

Amplitude, Phase & Frequency

@usgraphics When Houston Mono

Berkeley Mono featured on an e-ink display!

@simplifyinAI Cool

🚨 BREAKING: Someone just open-sourced a full suite for tracking satellites and decoding their radio signals locally. You don't even need the internet. It uses an SDR to pull weather images and raw data straight from space to your hard drive. 100% Open Source.