Anthropic 도서 스캔 논란, MCP 세션 제거, Kimi K3 1-bit 공개, AngelSpec 추측적 디코딩, 한국 증시 충격

Stateful vs. Stateless MCP core. (Anthropic's biggest MCP update) Let me explain what that means: Until this release, talking to an MCP server worked like a phone call. Both sides ran an initialize handshake, and the server returned a session id carried on every later request. That session was a live object inside one specific server process, holding the negotiated state. Picture a restaurant where the waiter who took the order is the only one who knows what was ordered. Everything works until that waiter goes home. Because that state lives in one process, a balancer could not spread requests across three MCP server instances. Teams pinned clients with sticky sessions or pushed session state into shared storage, which blocked autoscaling and made a single restart drop every open session. The latest update deletes all of it. The handshake and the Mcp-Session-Id header are gone, and each request now carries its own protocol version, client identity, and capabilities in a _meta field. Clients that want the capability list up front can call server/discover, but nothing requires it. The restaurant works the other way around now. That flips the restaurant around. Every request is a written order slip now, and any waiter can fill it. So any request can land on any instance behind a plain round-robin balancer, with no shared session store. MCP servers become ordinary HTTP services that run on serverless and edge, and survive restarts. Removing sessions broke three features, so each got rebuilt. → Tools that need to ask the user something mid-call used to push that request down a held-open stream. Now the server returns input_required and the client retries with the answers attached. → Method and tool names moved into the Mcp-Method and Mcp-Name headers, so a gateway or rate limiter can route and meter without parsing the JSON-RPC body. → List responses carry ttlMs and cacheScope, so clients cache tool catalogs instead of refetching them on every reconnect. State did not disappear here, it moved somewhere the model can see it. Applications that need continuity get an explicit handle from a tool and have the model pass it back as an argument on the next call. A session id in a header is invisible to the model, while a handle in the arguments is something it can read, thread between tools, and recover from after a failed call. Official Spec: http:// modelcontextprotocol.io/specification/ 2026-07-28 … Since we are on MCP, I also wrote about why the MCP versus CLI debate was the wrong one, and how agents can call tools by writing code instead of loading every schema into context. The article is quoted below.

Stateful vs. Stateless MCP core. (Anthropic's biggest MCP update) Let me explain what that means: Until this release, talking to an MCP server worked like a phone call. Both sides ran an initialize handshake, and the server returned a session id carried on every later request. That session was a live object inside one specific server process, holding the negotiated state. Picture a restaurant where the waiter who took the order is the only one who knows what was ordered. Everything works until that waiter goes home. Because that state lives in one process, a balancer could not spread requests across three MCP server instances. Teams pinned clients with sticky sessions or pushed session state into shared storage, which blocked autoscaling and made a single restart drop every open session. The latest update deletes all of it. The handshake and the Mcp-Session-Id header are gone, and each request now carries its own protocol version, client identity, and capabilities in a _meta field. Clients that want the capability list up front can call server/discover, but nothing requires it. That flips the restaurant around. Every request is a written order slip now, and any waiter can fill it. So any request can land on any instance behind a plain round-robin balancer, with no shared session store. MCP servers become ordinary HTTP services that run on serverless and edge, and survive restarts. Removing sessions broke three features, so each got rebuilt. → Tools that need to ask the user something mid-call used to push that request down a held-open stream. Now the server returns input_required and the client retries with the answers attached. → Method and tool names moved into the Mcp-Method and Mcp-Name headers, so a gateway or rate limiter can route and meter without parsing the JSON-RPC body. → List responses carry ttlMs and cacheScope, so clients cache tool catalogs instead of refetching them on every reconnect. State did not disappear here, it moved somewhere the model can see it. Applications that need continuity mint an explicit handle from a tool and have the model pass it back as an argument on the next call. A session id in a header is invisible to the model, while a handle in the arguments is something it can read, thread between tools, and recover from after a failed call. Official Spec: http:// modelcontextprotocol.io/specification/ 2026-07-28 … Since we are on MCP, I also wrote about why the MCP versus CLI debate was the wrong one, and how agents can call tools by writing code instead of loading every schema into context. The article is quoted below.

Kimi K3 can now be run locally! The 1-bit model retains ~78.9% accuracy after we shrunk it from 1.56TB to 594GB (-62% size). Run on a Mac Studio + 128GB RAM device. Kimi K3 is the strongest open model to date. Guide: https:// unsloth.ai/docs/models/ki mi-k3 … GGUF: https:// huggingface.co/unsloth/Kimi-K 3-GGUF …

We’ve open-sourced AngelSpec, an end-to-end speculative decoding framework supporting both training and deployment. On Hy3-A21B, DFly delivers a 1.98–2.40× end-to-end speedup over autoregressive decoding across tested concurrency levels from 4 to 64, with 10.5–11.8% higher throughput than DFlash. Training code and Hy3-A21B MTP/DFly drafter weights are now available: GitHub: https:// github.com/Tencent/AngelS pec … Paper: https:// arxiv.org/abs/2607.25852 Docs: https:// angelspec.readthedocs.io Hugging Face: https:// huggingface.co/collections/An gelSlim/angelspec … ModelScope: https:// modelscope.cn/collections/An gelSlim/AngelSpec … #Hy3 #AngelSpec #OpenSource

Interesting thing you need to know, Anthropic ran a secret project they didn’t just train on books. They destroyed the physical objects so the knowledge would only exist inside their models. Anthropic spent tens of millions buying books, sliced the spines off millions of them with hydraulic cutters, high-speed scanned the pages, then sent the paper to recycling. They called it Project Panama.
They still call themselves the responsible lab. One internal document Two sentences that say everything: 
- “Project Panama is our effort to destructively scan all the books in the world.”
 - “We don’t want it to be known that we are working on this.” That’s an actual Anthropic internal planning document. They spent tens of millions of dollars buying physical books (many used and rare), hired a former Google Books executive, used hydraulic cutting machines to slice the spines off millions of volumes, scanned every page at high speed, Court later ruled the destruction of purchased books was fair use. The digital files stayed inside the company. The physical books were gone. And they still position themselves as the most ethical lab in the industry. Meanwhile @elonmusk just told the xAi team the opposite: preserve rare books and scan them the hard way.

South Korea's KOSPI index fell nearly 6% as AI slowdown fears hit chipmakers, triggering a market-wide circuit breaker for a second straight day. SK Hynix posted record quarterly profit but still missed analyst forecasts.

. @UseApolloio will be at Interrupt NYC. Apollo's AI Assistant is one of the earliest production multi-agent systems on LangGraph. Their team will share how they migrated it from a hand-rolled supervisor to deepagents.create_deep_agent without breaking quality across production traffic. Catch this talk + more @ Interrupt NYC, Sept 24. https:// interrupt.langchain.com/nyc