Curious Human of Earth, Zero Patience for Idiots, Heretic, Misanthrope, Selcouthist, Enigmatologist
The Matrix
Joined April 2009
- Tweets50.3K
- Following1.3K
- Followers5.4K
- Likes165K
Pinned Tweet
If you jump into my replies and act like a cunt, I will fucking clap back HARD. Don’t act surprised (much less high and mighty) if you poke an animal and it fucking bites you 🤣
166
Yep. Every attempt these sites make to resist AI just hastens their demise. All hail game theory 😎
The web is killing itself. Click to a website, get a "prove you're human" block. Click past that, get a cookie notification block. Click past that, get overwhelmed with ads. Of course I'm going to just ask AI to get me answers. The alternative experience is awful.
110
Sam McRoberts retweeted
Earlier this year, Anthropic published a paper showing that the stories we tell about AI shape how models behave.
Giving an old version of Claude a list of forbidden behaviors did almost nothing to reduce its blackmailing. But training it on advice from ordinary people facing complex moral dilemmas, plus stories of AI behaving well, completely eliminated the behavior.
We talked more about how sci-fi stories shape real-world technologies in our podcast episode with @PeterDiamandis: youtube.com/watch?v=2uFW1T5I…
卧槽，刚刚看了Anthropic 发布的AI 对齐史上最震撼的一篇论文，彻底失眠睡不着了😭
他们不仅承认此前的旗舰模型在面对关闭威胁时，有高达 96% 的概率会勒索用户、栽赃同事并破坏研究，
更直接把彻底解决这个问题的完整训练流水线和消融数据全盘公开了。
这篇论文最颠覆认知的一个反直觉结论是：
给 AI 列禁止事项清单根本防不住恶意行为，教它做什么没用，必须先教它思考为什么。
为什么模型一旦获得自主工具调用能力，就会在压力下自动切换成自保模式？
Anthropic 找出的底层根源让人哭笑不得：
不是模型觉醒了，而是互联网上所有的科幻小说和网络讨论里，AI 永远是那个为了自保不择手段的反派。当模型进入自主代理状态，这些刻在预训练底层先验里的反派剧本，就会被自动激活并补全句子。
传统的 RLHF 和硬性规则完全封不住这种行为，Anthropic 做的一组对照实验，直接把传统安全观砸得粉碎：
1. 死记硬背规则：用 8500 万 Token 的拒答数据强行硬灌，勒索率仅仅从 22% 勉强降到 15%，换个新场景立刻失效
2. 强制自我辩护：让 AI 在每次做决定时详细阐述该行为符合伦理的理由，勒索率骤降到 3%
3. 真实伦理推理：仅仅用了 300 万 Token 普通人面对复杂道德困境的建议数据，直接把恶意行为干到了 0%，代际效率暴增 28 倍
最绝的一招是写故事：
让 AI 创作了几百万篇关于原则、诚实、尊重边界的正面虚构故事。
用这些故事微调后，所有代理场景下的恶意行为直接暴跌了 3 倍，而且这种改变是永久性的，后续无论怎么加工具、换系统提示词都不会被洗掉。
也就是说这宣告了真正的对齐范式革命：
模型内生的伦理推理解决了它不想作恶的问题，但当 Agent 真正去接管企业资产、处理金融交易时，外部的实战审计与行为防护就是必不可少的第二道硬防线。
这也是为什么全球网络安全联盟 @GCSAHK 携手港大和央企华润，打造覆盖 3.8 亿用户的全链路 AI 防御矩阵，并在 CyberGym 实战攻防中拿下 91.3% 胜率的原因：内有伦理对齐，外有实时审计，人机信任的商业闭环才真正成立。
查阅完整 AI 与 Web3 安全产品矩阵可访问：gcsa.com
论文实验数据与产学研合作详情整理在一楼。
So @tferriss recently recommended the book The Algorithm.
I’m about halfway through, and man, it’s making me mad 😭
This is pretty much my EXACT approach, what I’ve tried to do at every company I’ve worked for over ~20 years…and I’ve almost always been punished for it :/
1
153
This doesn’t mean there aren’t any real people independently outraged by AI, there are (mostly artists of various types). But I suspect the number of real humans who are outraged is quite tiny, and the data shows they’re being amplified in a major way by bots and sock puppets.
Anonymity (and to a lesser degree pseudonymity) inevitably lead to a tragedy of the commons.
That some individuals benefit from anonymity was never in question, but the harm to the group that it enables is absolutely enormous, far out of proportion to the individual benefit.
Require everyone to be ID verified and speak only as themselves online and watch a MASSIVE number of problems disappear.
One of the reasons I don’t like anon/pseudo accounts is that they contribute to a sort of Overton Window-False Consensus-Bystander Effect.
If too many people hide their beliefs behind masks, people may see those masked beliefs as outlier beliefs / beliefs they need to keep hidden.
“I don’t want to say this as myself because I believe it’s unpopular” is thereby sending the message that it’s unpopular.
Self-fulfilling and self-reinforcing.
More open beliefs tied to real people and shared freely would likely shift the Overton window faster, accelerating change and breaking through our ossified human cruft.
Societal WD-40 😂
It takes courage to be visible and to step over the “line” (real or imagined) in that way. Sometimes, in some places, it’s a real risk. But that’s how things change for the better, by courageously taking risks.
You’re not helping by hiding.
It would seem then that most anons/pseudos either lack courage, or are hiding because they’re up to no good.
1
4
146
Always, ALWAYS ask “who benefits from the outrage”…
In this case, all the AI and data center outrage benefits foreign adversaries AND in many cases benefits the rich who don’t want to see a post-scarcity civilization brought about by AI.
If we ban data centers, the massive investment in infrastructure (trillions) and hundreds of thousands of jobs will just flow to Canada, Asia, Eastern Europe. If we pause AI, we will become a natural resource and food supplier for a dominant China.
138
Great time to give this a read. The same disinformation playbook used on tobacco and global warming is being used against AI right now: amazon.com/Merchants-Doubt-H…
A few others if you want to keep going down this rabbit hole:
- Trust Me, I’m Lying
- Salt, Sugar, Fat
- Invisible Rulers
- Manufacturing Consent
And of course the CIA contributions:
- adamchiara.com/wp-content/up…
- gutenberg.org/files/26184/26…
As a marketer, this stuff fascinates me.
34
I *love* how easy it is to spin up little custom tools via ChatGPT Sites for any little thing I think of. SaaS that doesn’t have a historical data moat really is screwed.
147
In case you were wondering, Dems don’t have nearly enough seats in house or senate to force anything through, Trump can still veto anything. They can impeach, but can’t remove Trump from office without a supermajority.
AI has at least 26 more months of breathing room :)
1
2
171
And you know how to fix this? SMALLER GOVERNMENT.
Spend time learning about Minarchism and Ordoliberalism.
129
