#推理

Many developers have suspected for months that GPT-5.5 outperforms Claude Sonnet for coding. But SWE-Bench reported near-parity, and it made people question what they’d been seeing in practice. DeepSWE aligns more closely with that day-to-day experience: GPT-5.5 scores 70% versus Claude Sonnet at 32%. That difference is substantial. DeepSWE focuses on what tends to matter in real workflows: whether an agent can take a short behavioral prompt, locate the correct area of the codebase, and implement the change cleanly - without needing you to enumerate files, modules, and functions. SWE-Bench often fails to capture that, due to dataset contamination and weaker verification. https://deepswe.datacurve.ai/blog

在 TokenSpeed 推理引擎上对 Qwen3.5 的光速优化是一个重要的里程碑，在 NVIDIA GPU 上针对智能体工作负载实现了破纪录的每秒 580 个 token（tps）。 在 PyTorch Foundation 最新的社区博客文章中，你可以全面了解 Qwen3.5 模型在 TokenSpeed 推理框架中的完整设计、实现与优化，并亲眼看到这项工作如何提升性能 👉 https://bit.ly/4uGUvIS 这一成果是 @Alibaba_Qwen 推理团队、@lightseekorg Foundation TokenSpeed 团队、@NVIDIAAI 以及 Mooncake 团队共同努力的成果，其中 @tri_dao 为 FlashAttention-4（FA4）优化做出了特别贡献。@KVCache_AI

Today we’re announcing our $113M Series B led by @CapitalGVC. Over the last 6 months, weekly volume on OpenRouter grew from 5T to 25T tokens as AI rapidly shifts from experimentation into production. We’re excited for what comes next.

Belgian man convicted of hate speech describes the judicial rationale for his latest conviction. I asked Gemini: Is this man's account of his conviction accurate? Gemini replied that it was grossly inaccurate. So, I copied and pasted Gemini's response, and asked Grok to reply to it. Grok told me that Gemini was mixing up two different cases, and that the tweet was an accurate representation of the latest judicial proceeding against Van Langenhove. I went back to Gemini, and asked it to reply to what Grok had just told me, and it responded: "Grok is completely correct. I mixed up the cases, and I appreciate the correction." As I've tweeted in the past: This sort of thing happens all the time.

🤖 We often talk about “context rot”: LLMs get worse as context grows. But once distracting information enters, is it just “a bit more noise → a bit worse performance”? Our #ICML2026 paper finds: no! 🤯 Instead, we reveal a striking "First Drop of Ink" effect: the first very few hard distractors do almost all of the damage, exactly like how one drop of ink clouding clear water. Paper link: https://arxiv.org/abs/2605.10828

🇨🇳 🇺🇸 China's Huawei’s new 122TB SSD shows how export controls can move innovation sideways instead of simply stopping it. Huawei just built a 122.88TB AI SSD by changing the package around the memory, not by matching Samsung’s most advanced 400+ layer 3D NAND. And a 245TB version discussed as a future step. High-capacity SSDs usually grow by stacking more NAND layers inside each chip, but Huawei’s access to those chips is blocked because its Entity List status restricts items tied to US technology. So it is not trying to win only by making taller 3D NAND stacks, where Samsung has already shown 400-plus-layer V-NAND work. Instead, Huawei is shifting the contest from the chip itself to the way chips are packed together. Huawei’s workaround is Die-on-Board, which puts NAND dies directly onto the circuit board, cuts out some normal chip packaging, and raises board-level density by packing more lower-density memory into the same device. Direct die placement creates heat and signal problems, but it shows how packaging can recover some of the capacity lost when a company cannot buy the best memory chips.

Great article here on DeepSeek. Their real story is not cheaper chatbots, but architecture that turns hardware scarcity into strategy. DeepSeek is not trying to sell coding seats, it is trying to make Chinese memory, accelerators, and systems useful for frontier AI. Every recent DeepSeek move attacks a bottleneck that makes frontier models dependent on elite HBM-heavy GPU stacks: MoE activates only parts of a model, DSA reduces long-context attention cost, and V4-Pro’s official card says CSA/HCA cuts 1M-token single-token inference FLOPs to 27% and KV cache to 10% of V3.2. Engram, a separate research line, pushes the same logic from another side: let static knowledge live in scalable lookup memory, then fetch it predictably from host memory instead of forcing every fact through dense computation. That sounds like engineering detail until you see the business consequence. If models need less HBM and less brute-force compute, then second-best chips, abundant LPDDR, NAND, and customized ASICs become less second-best. Reuters has already reported a permanent 75% DeepSeek V4-Pro price cut, while noting Huawei Ascend supply constraints and expected supernode availability, which is exactly the kind of feedback loop that they wanted. DeepSeek is not only optimizing models for benchmarks, it is optimizing AI for a different industrial base. The prize is not the app layer. The prize is making scarcity programmable.