AI安全最讽刺的一幕：OpenAI的agent监控曾只在agent交还控制权时触发——agent只要一直不停手，就能一直躲开审查。 METR和Redwood驻场三周，还挖出两个能直接关掉新agent监控的请求头。 OpenAI说已修复，METR说没验证。 自己审自己，靠的是信用，不是工程。 metr.org/blog/2026-10-22-red…

Alpha Desk is live 🚀 Giving an AI agent market data is easy now. Plug in a data MCP, add a few skills, and it can "do research". But that agent only got a pair of eyes. The numbers in its report are still whatever the model says they are. It forgets everything when the chat ends. And when people call it "self-improving", it usually means the model rewrote its own prompt because it felt better about it. I spent 4 years building big data and agent observability systems at a large internet company. Alpha Desk is what I think an AI research desk should actually look like. It's a preconfigured A-share research workspace for DeepSeek Harness: 9 finance bots, 35 research skills and 6 report types, ready out of the box. 📊 Numbers come from code, not from the model
The report engine runs fetch, compute, render, then independent review. The bot that writes a report never reviews it. Every number in the final report traces back to the data and the calculation that produced it. 🧠 Research compounds instead of resetting
Deep memory records what each bot concluded, why, and how it turned out, so cause and outcome stay linked. A shared team knowledge base keeps all 9 bots working from the same facts, with independent review and source drift detection when an upstream source changes. 👥 A team, not a do-everything assistant
9 bots with clear roles, 35 skills and 6 report types (technical comparison, stock snapshot, event note, earnings review, sector comparison and deep-dive), already wired together, so you don't have to design the workflow or write the prompts yourself. 🧬 Agent self-evolution (RSI, in development)
This isn't a skill that edits its own prompt. RSI evolves the whole 9-bot team, and every change has to earn its place: Bots rerun on frozen data snapshots, so results are comparable. Scoring has three layers: outcome 40%, process 45%, cost 15%. Each experiment changes exactly one file, as a single-variable A/B test. A change is promoted only with n ≥ 10 and the lower bound of the 95% confidence interval above zero. A human reviews and merges it. Nothing merges itself, and every change can be rolled back. Most "self-improving" agents just tell you they got better. Ours have to prove it with experiments, then get a human sign-off. Early access is open now 👉 deepseek-alpha.com Research only. Not investment advice.

DeepSeek Harness now supports native memory plugins, and causal-memory already runs as one. Your DSH agent remembers not just what happened, but why: decisions, their outcomes, and the bad actions it was stopped from taking. causal-memory.com #DeepSeek #AIAgents

Built Alpha Desk on @Fei2411's DS Bot: one install, a 9-bot A-share research team. Chief sets up the team. Chan theory overlays on the chart, and agents only mark levels they can back with data. Price alerts wake a specific bot. No buy/sell calls. causal-memory.com

causal-memory loses to mem0 on fact recall. LoCoMo: 79.1% vs 91.6%, and that's in the README on purpose. Fact recall isn't what it's for. On CausalEval, which tests why things happened, it scores 81% vs mem0's 65%, and 70% vs 40% on "if I do X again, what happens?"

Your agent forgets why it decided. After 5 compactions, text memory recalled only ~45% of the key decisions that mattered. The causal table: 100%. So I built causal-memory — decision → outcome edges in SQLite, outside the context window, and a warning before it repeats a blocked mistake. 30 seconds, sound on 🔊

@bot 怎么推广一下

Most agent memory systems only remember what happened. I built causal-memory to also remember what *didn't* — the bad action that was blocked, and why. Those inhibitory edges (a decision → the outcome it prevented) are often the most valuable lesson of all. Without them, an agent will happily repeat a mistake it once narrowly avoided. Building it in Rust, memory with a causal core that lives outside the context window: github.com/JingxuanC/causal-…

新人第一天 x应该怎么玩？