DeepSeek·거대 MoE·Astra: 비용·확장성·수학 성과

While DeepSeek V4-Flash is significantly cheaper on price per token, this can be misleading if the overall cost per task ends up being higher due to more turns being made. However, @ArtificialAnlys reports DeepSeek completing the same benchmark tasks as Fable at 105x lower cost.

DeepSeek-V4-Flash hitting 95.9 tok/s decode at full 1M context on 2× DGX Spark. - +28.6% over previous baseline. - 1M context fully retained. - Clean role boundaries & valid tool parsing Proper DSpark & regular CUDA graphs

You can now run a 2.8 trillion parameter model on a 4GB GPU. Open-sourced layer-wise inference that only loads one layer (or expert) at a time. - Kimi K3 (2.8T) - 3.72GB - DeepSeek-V3 (671B) - 12GB - Llama 3.1 405B - 8GB No quantization, No distillation, No pruning The biggest model needs the least VRAM because it’s sparse MoE it only streams the experts the token actually routes to. It never loads the full model, Only the current layer or the active experts. It’s extremely disk-bound and slow, but the fact that the full open weights even fit is wild.

Run 743B MoE models on two 16GB consumer GPUs, A new Rust & CUDA inference engine that streams experts from NVMe instead of loading everything into VRAM. Results on dual 16GB cards: - GLM-5.2 743B - 2.7 tok/s - Tencent Hy3 295B - 7 tok/s - Also supports Kimi K2.7 1T, DeepSeek-V4-Flash, Qwen3 MoE, etc. - Zero-config multi-GPU (auto measures PCIe bandwidth and places attention & hot experts optimally) Democratizing 300B–1T class MoE models, No llama.cpp, streaming MoE design run giant sparse models on normal hardware.

Google paper a method that might question about the entire transformer era. For a decade, every major AI model on earth, ChatGPT, Claude, Gemini, has been chained to the exact same architecture. The Transformer. It changed everything by letting models look at entire sequences at once. But it has a fatal flaw. Its memory consumption scales quadratically. $ O(L^2)$. As context windows grow to millions of tokens, the compute cost explodes. It requires massive data centers, thousands of GPUs, and fortunes in electricity. For years, researchers tried to replace Transformers with old-school Recurrent Neural Networks (RNNs) to get lightning-fast linear speed. The problem? RNNs have fixed-size memory. When texts get long, they forget. They choke on complex reasoning and recall tasks. So everyone went back to Transformers. Until Google dropped a bombshell new paper. They introduced Memory Caching (MC). Instead of forcing an RNN to rely on a fixed, static memory, Google’s technique lets the model's effective memory capacity grow dynamically with the sequence length. It acts like virtual memory for neural networks. By caching checkpoints of hidden states and integrating smart mechanisms like sparse selective retrieval, the model dynamically looks back at what it needs, without the crushing mathematical overhead of a Transformer. The results completely break the old rules of sequence modeling: - Matches Transformer-level accuracy on complex in-context recall tasks. - Completely closes the performance gap that kept RNNs sidelined for years. - Maintains linear complexity, making long-context processing radically cheaper and faster. We've spent years believing that massive, heavy attention mechanisms are the only way to build intelligent systems. Google just proved you can have the infinite memory of a Transformer with the blistering speed and efficiency of an RNN. The era of brute-forcing scale with quadratic attention is hitting a wall.

It clear sign Acceleration, this week: OpenAI says its next model, Astra, solved 10 open problems in mathematics. legitimately. total compute cost to find all ten: about $2,000.

With caching this is probably 100m tokens? Once models are smart enough you really don’t need many tokens to figure out even advanced stuff A codex subscription easily smashes that a week

Wtf, 594 BTC gone in 25 minutes. from 500 air-gapped Coldcards without internet connection, A 5-year-old firmware bug just cost Bitcoin holders $38 million.
And the devices were never connected to the internet. A 2021 firmware bug made the devices skip proper randomness when creating seeds. What should have been 128 bits of entropy dropped to roughly 40 bits. One developer even predicted his own seed just by counting button presses. The attacker figured this out, precomputed the keys, and waited years for the wallets to fill up. Then cleaned them out in one shot. The code was open source the entire time Public repo The bug still sat there for five years. If your seed came from a Coldcard Mk3 after March 2021 and you didn’t use enough dice rolls or a strong passphrase treat it as compromised and migrate.

They just open-sourced the RL framework that trained GLM-5.2 Most RL frameworks become bloated piles of disconnected services. slime did the opposite, slime keeps a single Megatron & SGLang training loop and pushes all task variety into data generation. One clean training loop, Everything else (agents, tools, multi-turn, long-horizon tasks) handled as flexible data generation. This is the actual production stack behind GLM-5.2, GLM-5.1, and the earlier GLM series. Full post-training of GLM-5.2 ran on it in roughly two days.