Collab上A100 80G运行Qwen Flash Next的速度很不错。 prefill 2800~3000, decode 90+

AI 智能正在以人类历史上几乎没见过的速度变便宜。 Epoch AI 最新的数据：过去三年，获得同等 AI 能力的成本平均每季度下降 47%，一年大概便宜 13 倍。比 DNA 测序降本快 4 倍、计算快 6 倍、锂电池快 18 倍、电力快 54 倍。 2025 年初 o3 在 GPQA 做到 75% 左右，一道题成本大概 $0.30；不到 18 个月，GPT-5.6 Luna 做到差不多的成绩，只要 $0.0004，直接便宜了 725 倍。 AI 正在从“昂贵的软件工具”，变成一种近乎无限供应的生产要素。 但便宜不等于算力需求下降，反而很可能是 Jevons paradox：以前舍不得让 AI 干的活，现在全让它干；以前问一句，现在 Agent 可以自己跑几个小时、并行十几个任务。单次 intelligence 越便宜，人类消耗的 intelligence 反而越多。 未来被快速商品化的，可能恰恰是“ llm”本身。真正值钱的东西会越来越往数据、workflow、分发、用户关系，以及能把廉价智能转化成实际收入的产品的能力

Here's another one from @XiaomiMiMo Mimo-v2.6-flash! A three.js animated mechanical Koi fish. Same way as before, no assets are used, all in code, three.js and without any iterations i.e. no harness. I changed the subject in prompt and this again delivered.

Pascal是十年前的架构。 本地对计算资源的占有是有明确影响 两个字：短缺。 四个字：极度短缺。

现在本地 AI 已经开始玩到 Pascal 架构了

Introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 for most tasks, and costs 40% less to run than Opus 5.

そういえば、Qwen3.8 Flash Nextはwevdevベンチでは fabele 5の上の11位に居たりします。このサイズではかなり賢い！！！！ ちなみにQwen3.8 27Bは25位です！！！！ arena.ai/leaderboard/code/we…

The Last Human Zoo exhibition, Belgium (1958).

Operation Cheepseek Phase 2: $60 of usage for DeepSeek v4.1 Flash is now permanent enjoy :)

Periodic reminder that Xi is just a wifeguy. He's still very happy he could score with Peng Liyuan, a very famous Chinese idol. Over the entirety of his career prior to becoming the General Secretary, he was mainly known as "that dude who's Peng Liyuan's husband".

BREAKING: anthropic ceo dario amodei is reportedly concerned that an rtx 3060 12gb now runs a 27 billion parameter model fully locally at 50 tok/s fresh, and has called it a grave danger that gpus this powerful are allowed in homes. sources say the new essay's biggest concern is that a gpu this powerful can burn a house down if it isn't power limited, which requires advanced technical safety skills. more as it develops.

A good day for updates ⚡️ Update for Qwen3.8 Flash for a solo DGX Spark - Decode on prose is now 49 tok/s single stream. - Decode on code is now 62 tok/s single stream. - Faster follow-up replies. - Optional vLLM 0.30 path with official Nvidia NVFP4. This is still the BEST model to run a single DGX Spark.

I built a trading bot that buys the “most hated” Mag 7 stock each month. Ranks by 3-, 6-, and 12-month returns, buys the worst average rank, reruns monthly. Last month: $META This month: $TSLA Since 5/23 (Mag7 coined): +347% HatedM7 +162% Mag 7 +90% S&P 500

一国两制就是好哎，天才般的制度设计。

如果不喜欢最高点的股票，还有一个潜在的破局点是，拼命想进入推理世界的困境反转股。 $QCOM 手机快被AI干掉了，如果不搏命，就是半个死人了。 搏出256GB联合 快速Interference 架构，又是翻身一哥把歌唱。 须知TSM-尖端制程公司这条线上，只有它家守着金饭碗活成了路边一条。 等死，死AI可乎？