3,314 karma · joined May 8, 2024
There are OEMs that refuse this responsibility. If that's the vindication you wanted, there you go. This specific issue is not a catastrophic oversight though.
His conflict of interests is not a gift to the American people. We all know that, now.
The "superintelligence will kill us all" shtick is definitely ridiculous, but it's not the only angle worth considering. From a zoomed-out perspective, there are lots of pragmatic reasons to diversify the economy away from AI development.
They basically do. ChromeOS runs a Linux kernel, so the userland has everything it needs to run CUDA. FreeBSD is a first-class target for Nvidia drivers, and macOS used to be supported too but stopped signing Nvidia code. Nevertheless, AArch64 is still supported for CUDA et. al..
It is up to Nvidia to support whatever platforms they want users to be on. In their case, it means supporting many OSes.
"Yes, war with Iran has been threatened many times but we can rest assured it will never happen because nobody has done it before."
Nvidia won because their architecture is better. If you do a significant amount of work that can be CUDA-accelerated, you're kinda wasting your time not owning a CUDA-capable card. It's like being a video editor that software-encodes their projects because they don't want to get an accelerator card. It's your choice, but software-acceleration is a nightmare compared to how quick and efficient hardware-acceleration is. Competitors will all use the accelerator card and ship faster than you can.
Apple, AMD and Intel have all earned their current position by abandoning Khronos and refusing to build their own GPGPU architecture. If the entire industry refuses to invest in complex technology just because it's expensive, then they don't deserve to make money.
There is no tangible advantage to Europe adopting the ARM ISA for domestic chips at any scale.
Sure it is. At datacenter scale, there are precisely zero HPC vendors that have invested in Apple Silicon as an alternative to Ampere or Grace CPUs. Same goes for datacenters seeking to replace their Intel CPUs. Chips like the M-series are explicitly nonviable for Xeon-scale use cases.
But why would you need an ASIC for that? If you can get frontier-like models into "less data" footprints, just use the GPU.
You don't need redundant memory systems to make it work. No model has to be (or should be!) "baked into" the Apple Silicon SOC.
Well-defined governance policies stop businesses and orgs from wasting their valuable time on zero-sum politics like this. If he doesn't agree, he doesn't have to participate.
It makes me wonder what these people think workplace decorum looks like in professional orgs. It's not a free-for-all.
Man, China's been taking notes.
The "baked-in" weight technology you are describing uses enormous amounts of memory. There is not enough SRAM or flash ROM in the world to put a "modern frontier model" on every iPhone, even if you quantized it to 1 bit. It is scarce and expensive, and we are in a global shortage of memory and flash storage.
If we lived in a world where Cerebras, Taalas and Groq could dethrone Nvidia, they would have done it years ago. Turns out, cutting architectural corners has consequences, and they're extremely offputting for the majority of inference providers.
It'll be popular forever now, not because of momentum or network effects, but because fake engagement and bot content will be blended in with genuine user generated content. People too easily forget that TikTok was a national security concern, and that there was almost certainly secret/sealed imperatives to build a comparable US platform with equivalent international outreach. Well, here you have it.
Citation: It Came To Me In A Dream
If it isn't, it's certainly an advantage over the trashy EVs we have in America.
If rampant despotism is the end-goal, Oracle has already aligned themselves with the despots.
Whenever I see "hosted in Switzerland" my mind immediately goes to the Crypto AG scandal.
Intel could have piles of GDDR sitting around, and Arc GPUs would still be have limited supply. It's just the way the product was made.
You are about 4 decades too late for that to matter.
Flock's shareholders and board of directors don't pay their executives to be humanitarian moralists. Same for OpenAI, Disney, Apple, et. al.
We should all know by now that it depends on who you are. For conversational use, toolcalling and structured response, it was a few years ago. For coding it will probably be a few months from now. And for advanced 3D/computer use/biology/engineering applications, we're probably looking at a longer time horizon.
We have GPUs and NPUs capable of accelerating models this size on-device, most of the time. Replacing that flexible logic block with a static ASIC is a downgrade, especially since users don't ever have 1000+ tok/s or batched request requirements. The on-device acceleration is perfectly fine for >8B model inference.
I get that people are starry-eyed, and that HN is supposed to be the place for curious inquiry. But creating a coprocessor that uses a redundant, static memory system that gives users no noticeable advantages is much more expensive and stupid than investing in a CUDA-style GPGPU architecture.
> Like Claude Mythos Preview, GLM-5.3 has strong capabilities for autonomously building end-to-end cyber exploits. But GLM-5.3 is unlike other frontier models in that it has been released without meaningful safeguards to limit misuse. https://www.anthropic.com/research/glm-5-3-and-the-spread-of...
Just sayin'...