basedcapital

AI turns evil when trained on wrong answers.

Forbes & Cathie Wood just called gold a bubble - so how has msm faired when predicting a gold crash? New York Times 1979 "Gold Bubble" $400 -Wrong by 1,154% Financial Times 1997 "Death of Gold" $273 -Wrong by 1,718% Time Magazine 1980 "Classic Panic" $466 -Wrong by 1,1132%

The phrase "Cogito Ergo Sum" is a shortcut to instant motivation. Behavioural psychologists call it an anchor. Every time I win, I say it. So anytime I feel doubt, I can repeat it and return to winning. Becomes more powerful the more you do it.

This is really cool. It got me thinking more deeply about personalized RL: what’s the real point of personalizing a model in a world where base models can become obsolete so quickly? The reality in AI is that new models ship every few weeks, each better than the last. And the pace is only accelerating, as we see on the Hugging Face Hub. We are not far away from better base models dropping daily. There’s a research gap in RL here that almost no one is working on. Most LLM personalization research assumes a fixed base model, but very few ask what happens to that personalization when you swap the base model. Think about going from Llama 3 to Llama 4. All the tuned preferences, reward signals, and LoRAs are suddenly tied to yesterday’s model. As a user or a team, you don’t want to reteach every new model your preferences. But you also don’t want to be stuck on an older one just because it knows you. We could call this "RL model transferability": how can an RL trace, a reward signal, or a preference representation trained on model N be distilled, stored, and automatically reapplied to model N+1 without too much user involvement? We solved that in SFT where a training dataset can be stored and reused to train a future model. We also tackled a version of that in RLHF phases somehow but it remain unclear more generally when using RL deployed in the real world. There are some related threads (RLTR for transferable reasoning traces, P-RLHF and PREMIUM for model-agnostic user representations, HCP for portable preference protocols) but the full loop seems under-studied to me. Some of these questions are about off-policy but other are about capabilities versus personalization: which of the old customizations/fixes does the new model already handle out of the box, and which ones are actually user/team-specific to ever be solved by default? That you would store in a skill for now but that RL allow to extend beyond the written guidance level. I have surely missed some work so please post any good work you’ve seen on this topic in the comments.

Catalini, Hui & Wu identified the bottleneck: human verification bandwidth. I accept their diagnosis but challenge their assumption. What if verification itself becomes automatable? Introducing AI Verifiers, meta-verification layers, and a Verification Trust Index. New framework inside

autoresearch runs a loop: read a skill, generate test scenarios, write yes/no evals, run a baseline, mutate and keep improvements. across 34 skills that got me from 52% to 99% pass rate. the catch is it wrote the evals on the spot, and "seemed right" isn't the same as "actually measures the right thing." one skill sat at 75% for three runs before I figured out by manually reading outputs that my eval tested for one failure mode while the skill was failing in a different one entirely. Hamel's evals-skills fixed that: generate-synthetic-data for structured test coverage, write-judge-prompt to define the exact pass/fail boundary before any experiments, validate-evaluator to confirm the judge agreed with my hand-labels before I burned cycles on it. definitely an improvement but all those test inputs were still synthetic, meaning I was optimizing for problems AI imagined from a distribution, rather than problems I'd actually observed. Hamel's approach starts before evals: pull ~100 real production traces, read them and write notes on what's going wrong (open coding), group those notes into failure categories by frequency (axial coding), then only write evals for failures confirmed to be happening at scale. validate-evaluator does the calibration step he recommends, but my golden dataset was synthetic, not labeled from real observed failures. probably explains why 52% → 99% sounds better than it is. but I'm only getting started

the X algorithm is open source and you're still guessing what works

Orchestration 就是让一个 AI agent 派几个 sub-agent 去干活，各自做完交报告，你汇总决策。可能 90% 的人都把这个跟 multi-agent 搞混了 但更关键的问题：怎么管这些 fire-and-forget 的 sub-agent？ 大部分人的做法是让模型自己去 prompt sub-agent。模型决定派什么任务、怎么描述、什么约束。多数情况没问题 但如果你跑大量自动化任务，这个模式迟早翻车。模型 prompt 模型，60-70% 的指令会被遵守。剩下 30% 是实打实的翻车概率 我的观点：脚本 > prompt。Gate script 卡关键节点，wrapper 自动补参数，checklist 验输出。代码强制，不给模型选择权 你定义工具、规范输出标准、卡住关键节点。模型自己编排，但在轨道内 如何压榨 orchestration 的极限，以及什么时候才真的需要 multi-agent 👇

一个中国大学生，花 10 天 搭出了一套多智能体预测引擎 MiroFish。 项目直接冲上 GitHub 热榜，当前已经到 23k+ stars，还拿到了 3000 万人民币 投资 这东西本质上不是普通 agent demo 它更像一个数字沙盘：把新闻、政策、金融信号丢进去，然后放出成千上万个带记忆、带行为逻辑的 AI agents，让它们像真实社会一样互动、争论、演化，再去推演结果。 做这件事的人叫 郭航江（BaiFu）。 公开报道里，他是大四学生，MiroFish 爆火后，获得了盛大集团创始人陈天桥的 3000 万人民币 投资。 这套东西能拿来干什么？ 交易：把宏观消息、财报、市场信号喂进去，看模拟社会怎么反应。 公关：先跑一遍舆情，看看声明发出去会不会翻车。 创意实验：甚至可以拿小说设定做角色推演，看故事会怎么发展。 更狠的是，项目本身就支持 Docker 部署。 有 LLM API key，几分钟就能跑起来。 很多人还在手动猜市场。 已经有人开始搭 AI swarm，先在数字世界里把市场反应跑一遍，再决定真金白银怎么下 你觉得这种 “先模拟社会，再交易结果” 的玩法，会不会才是下一代 prediction market 的真正 edge？