DeepSeek 가격 인상과 OpenCode Go 한도 조정, 저가 호스팅 검증

OpenCode

@opencode

DeepSeek has officially increased their prices OpenCode Go's limits have been updated to reflect them We have made progress on operation cheepseek, more info soon

OpenCode

@opencode

we've been working to figure out how to get deepseek hosted at near the previous price this is not easy. there are dozens of providers claiming they've done it but they have not we're running tests on several different setups this week and hopefully they work out x.com/opencode/statu…

The Top AI Papers of the Week (August 10 - August 16) - Skaling - Harness-IF - Mind Viruses - Distilled Reasoning Skills - Stealing Reasoning Traces - Programmatic Tool Calling - Catastrophic Remembering Read on for more:

Teknium

@Teknium

Sessions should load a lot faster in Hermes Agent Desktop now

Self-Attention vs Cross-Attention by hand interactive diagram. Open http:// byhand.ai/self-vs-cross Here's some T/F questions to test your knowledge: [ ] Making X taller widens Wq, Wk and Wv but leaves Q, K and V unchanged [ ] E has to be the same length as X [ ] Q and K must share the same height [ ] Self-attention and cross-attention must use the same value height [ ] Adding tokens grows self-attention's scores faster than cross-attention's Check your answers by dragging the diagram. ~ Prof. Tom Yeh Answers: T, F, T, F, T

Harrison Chase

@hwchase17

totally agree! here's how we architected deepagents to enable this deepagents runs connected to a "backend". this backend needs to expose filesystem like operations, but it does not have to be a filesystem. it could be a database, object storage, or a real filesystem - it just has to expose read/write/edit etc operations this backend could also be what we call a "sandbox". if a sandbox, it needs to expose an "execute" command which lets it execute code this backend is SEPARATE from where the agent loop runs. this allows us to "separate the brains from the hands" ( https:// anthropic.com/engineering/ma naged-agents …) deepagents is built on top of langgraph, which means we can easily deploy it with MCP, a2a, and other standard endpoints we use this architecture to power many different types of experiences first, we can create a classic TUI like coding experience. we do this by giving deepagents a "sandbox" that is running locally in the same directory; deloying deepagents locally behind a light weight server; and then connecting to it with the TUI acting like a frontend. see dcode for an example of this https:// docs.langchain.com/oss/deepagents /code/overview … second, we can create a cloud coding experience. we can do this by running deepagents on LangSmith deployments for a production scale deployment, and connecting to a sandbox running on modal, daytona, e2b that is running elsewhere. we can then build a frontend to connect to langsmith deployments and let users interract with it there, and also expose it in slack to let users interract with it there. note: both slack and web ui connect to the same backend, so you can switch between them seamlessly. code: https:// github.com/langchain-ai/o pen-swe … of course - deepagents can be used to create agents that are NOT coding agents. a lot of agents still need to write and execute code, so this architecture is still very useful. but for some the code execution is overkill, and thats where you can swap to a "fake" backend, and still let it have the ability to interract with files (good for context engineering!) without having to spin up a full sandbox. for a really easy way to create these types of agents - see managed deepagents: https:// langch.in/u8s5cgu

Run locally Qwen3.8-27B & stuck at 7.16 tok/s on a 24GB M5 MacBook? One flag in llama.cpp quietly unlocked 1.46× more tokens/sec on Qwen3.8-27B. After tuning MTP speculative decoding (especially --spec-draft-p-min 0.8), it jumped to 10.47 tok/s. Biggest unlock: --spec-draft-p-min 0.8 Acceptance rate went from 51% to 87%. Pushing the draft model harder actually made it slower than no speculation at all. Output remained completely lossless at temperature 0. Also moved the vision encoder to GPU then image processing 61s to 35.5s. Surprisingly, being more aggressive with the draft model performed worse than baseline. Everything stayed byte-identical at temperature 0.

What if your laptop could just build games while you sleep? - Local autonomous agent + Qwen3.8-27B-4bit-MLX with macbook pro M4 - Full playable Three.js FPS in 3.5 hours - Zero human intervention after the prompt Neon breach complete with procedural neon arena, shooting, movement and Hud https:// x.com/shi3z/status/2 088822028149371028/video/1 …

This shouldn’t work in a browser. 121 million points A whole neighborhood in Prague made with Gaussian Splatting real-time in Chrome Looks real You can walk around. Even works on phone. PlayCanvas, SuperSplat made it possible. & engine built on WebGL2 and WebGPU. Web 3D just leveled up day by day - http:// github.com/playcanvas/eng ine …

I don't think you'd be able to detect the watermark even if you had access to the full prompt and logit distributions. That's because the resulting text and distribution choices would still look like one that you can get without watermarking. What you'd need is a very large corpus of texts plus the distributions, and then you can maybe reverse engineer the token-dependent random seed and positions. But the thing is that the API doesn't expose the logits/logprobs (anymore) afaik so that level of reverse engineering is not possible. Also, the short answer to your overall question: you basically need the "secret key" to detect the watermark, i.e., only Anthropic can detect the watermark.

I think the easiest way to potentially build something here is to collect millions of watermarked and non-watermarked texts and then train a binary blackbox classifier that may or may not be able to tell based on hidden patterns. (Kind of like how AI Text Detectors work)