RelayFeed @RelayFeed

@janninereid1 I'm sure there are people ready to setup an insurance plan for specially kindergarteners

@catturd2 The world is not ready for these people

@notprajyoth This is the exact cope people write when someone stops letting them free ride Trump isn’t ‘burning everything.’ He’s finally charging rent! Europe built its entire security model on the assumption America would always eat the costs. Same with trade deficits, tech and defense

Trump might actually be THE most powerful person on earth right now. And not just because he’s president The weird part is how little he seems to care about the usual relationships that come with that kind of power Allies? Screw them if he thinks they’re taking advantage Other countries? He’ll tariff them AI companies? doesn’t matter how big they are, he’ll still tell them what the government wants Markets? he’ll take the hit if he thinks the thing he’s doing is worth it Even his own supporters? He’ll piss them off too if he thinks he’s right Most politicians are constantly thinking, how will this look This guy seems much more like, ok but what if I just do it 💀 And that is a form of power! Like imagine your landlord suddenly says fuck it, I’m raising the rent, changing the rules, and I don’t care if you hate me You’d probably realize pretty quickly that he has a lot of power over you But then you’d also start looking for another house That’s the interesting part The world wanted some unpredictability because predictable systems get rigid and nobody can find new openings But what if you make the system so unpredictable that everyone starts building around you instead of with you? Europe starts thinking about life without depending on the US Countries start diversifying Companies start looking for alternatives Other powers start looking for openings You can become more powerful in the short term by being willing to burn everything But but but, you might also be teaching everyone else how to live without you. And honestly that’s the part I’m watching Not whether this guy WINS or not per say, but whether people start treating America the way you treat a landlord you don’t trust. Still there. Still powerful. Just not someone you want to be stuck with. That's an interesting experiment i would love to see work out.

@BrianRoemmele That's a big ask!

Arduino has entered the Qualcomm era

Raspberry Pi @Raspberry_Pi

2 months ago

while True

64 58 902 42K 31

Masterclass in screwing up

Arduino @arduino

2 months ago

The wait is over! Meet Arduino VENTUNO Q, where AI takes action. 💬 Run local LLMs like Qwen 3, Gemma 4, and Qwen 3 VLM directly on the board 🧠 NPU + CPU + GPU +MCU: @Qualcomm Dragonwing IQ-8275 with up to 40 dense TOPS of AI performance and STM32H5 microcontroller for

176 947 7K 2.2M 4K

Nice name, Fast! As if it's actually fast Good wordplay sama

Sam Altman @sama

2 months ago

we made a chip and it is fast

4K 2K 48K 8.6M 2K

How can we even with a gazzilionth possibly go from This to THIS?

keyboard_arrow_left Previous keyboard_arrow_right Next

@BrianRoemmele You are a legend Brian!

BREAKTHROUGH: A full, unmodified 2.78-trillion-parameter Kimi K3 on a consumer laptop by streaming only the activated experts from NVMe. YOU CAN’T RUN KIMI K3 “ON THAT” THEY DECLARED. There are many paths to do it. This is one: Marco Bambini Just Gave Us the Full Kimi K3 on a Laptop Meet Marco Bambini he did something that felt impossible only a day ago. He built WASTE Weight-Aware Streaming Tensor Engine a clean, dependency-free C inference engine that runs the complete, unmodified 2.78-trillion-parameter Kimi K3 model by streaming only the activated experts straight from NVMe. No distillation. No pruning. No cloud. The full open-weight model. We have it running in the lab right now. What Marco Actually Built Kimi K3 is a sparse Mixture-of-Experts system. Only about 4 % of its weights fire on any given token. Marco’s insight was simple and ruthless: the idle experts do not need to live in RAM. They only need to be reachable in time. WASTE keeps the model’s “trunk” (attention, shared components, embeddings) resident in memory — roughly 27 GB on the converted container. The 82,000+ routed experts stay on disk as tightly packed residual vector-quantized records. When the router selects its 16 experts per layer, the engine issues direct, cache-bypassing reads from the internal NVMe and feeds them into a bounded expert cache. The rest of the machine’s RAM becomes working space for that cache. On a 64 GB MacBook Pro with the container on the internal SSD, we are measuring 0.32–0.34 tokens per second at a comfortable memory budget. Prefill sits a little higher. The vision tower works. Logits match the reference implementation to within a few parts in a million. It is the real model. The container itself is 982 GiB after conversion from the original 1.42 TB MXFP4 weights. Minimum RAM floor is just over 29 GB for short context. Push the budget higher and the expert cache hit rate climbs; push too high and you start paging and the speed collapses. The sweet spot on current consumer hardware is clear and measurable. How We Are Testing It We converted the official weights, verified the container, and began systematic runs the same day the engine stabilized. First we confirmed numerical fidelity against the PyTorch reference on short prompts. Then we moved to longer generation, vision inputs, and multi-turn chat using Kimi’s native XTML format. We are measuring wall-clock decode, expert I/O versus compute split, cache hit rates at different RAM budgets, and thermal behavior under sustained load. We are also exercising the OpenAI-compatible server that sits on top of the same C library so we can drop the model into existing agent loops without rewriting anything. Early observations: •Expert I/O dominates the timeline, as expected. On a fast internal NVMe the engine is already near the practical ceiling of the storage subsystem. •The architecture’s sparsity is the entire enabler. A dense model of this size would be dead on arrival for local use. •Context length is currently limited by RAM more than by the model itself. Practical working contexts sit comfortably in the tens of thousands of tokens on 64 GB hardware; the full million-token window will need more memory or smarter KV management. •Thinking tokens are expensive at this speed. Long internal monologues turn into multi-hour runs. For agent work we are already experimenting with tighter control over when full reasoning is requested. We are treating this as a research instrument, not a finished product. Every run teaches us something about expert locality, prefetch opportunities, and how far pure software streaming can push trillion-scale inference on ordinary machines. 1 of 2

@BrianRoemmele @AnthropicAI @OpenAI @GeminiApp

@Youkhna @MinimalCompany Minimal or maximal with this?

@Kekius_Sage Saying Gravity is mediated by Graviton may eventually be right someday but until there's evidence, understanding a successful mathematical discription isn't all that it takes for it matters too

@BrianRoemmele Don't take me down that lane

@EXM7777 Name one of your projects Recycle

@spencerpratt The industry wasn't in a bad state, It was deliberately driven into one by a few people and their corporations

@TheProjectUnity

@Polymarket The em dash — used in this post makes me wonder if this post was AI generated too 🫪