Helping where I can. Experienced in engineering management; focus on critical systems like medical/finance. Cross-disciplinary emphasis. I dabble in markets.
Wilmington, NC
Joined May 2008
- Tweets37.8K
- Following4.2K
- Followers1K
- Likes46.4K
mental blanking interval retweeted
🚨#BREAKING: A Tennessee man is COMPLETELY DISTRAUGHT after he spent THREE YEARS of his life growing a native plant garden to help struggling insects in his own front lawn...
...and the city MOWED IT TO THE GROUND because:
"His appeal when to the SPAM folder..."
His name is Chester Perry, and he lives in East Ridge, Tennessee and for three years, Chester filled his front yard with native and exotic plants.
His goal was to build a habitat for declining insect populations.
In September, a nearby resident complained about his lawn.
Code enforcement cited him for TWO violations.
One was about upkeep of the property. The city says Chester FIXED it.
The other was for “high grass.”
Chester appealed that one, because the insects need tall grass and the city gave him 10 DAYS to appeal.
He sent his appeal by email ONE WEEK after the citation.
The city says its firewall dumped his email into a SPAM FOLDER.
40 DAYS after the citation, city crews showed up and cut it all down.
Chester came home from lunch to find his yard stripped bare.
Three years of work. Gone in an afternoon.
The East Ridge City Manager ADMITS Chester tried to file his appeal in good faith.
They just never received it.
Their fix?
They decided not to FINE him...
mental blanking interval retweeted
the funniest part about this is that it suggests that all the doomer sci-fi stories on lesswrong and similar sites are literally causing the very behavior that frightens them so much. they're polluting the training data!
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。
mental blanking interval retweeted
WTF is going on? I didn't cry about the usage drama with @bot.
But now I know they are losing money serving up requests to Anthropic. Why the fuck am I getting guardrailed on my own private repo? @grok would never do that, but nanny state @AnthropicAI would.
@rumblevideo should invite this guy to Partner as he really adds something unique to the mix
and might be a good tie-in for the robotics stuff
mental blanking interval retweeted
This is another piece you will see at my show. It’s called “Ghosts on Jackson” (48” x 24” acrylic on canvas). I didn’t realize at the time of painting that they were going to tear down the building on the right. So it makes this scene even more special. So happy I was able to capture it before it left. Come see this and 15 other large pieces in my solo show on Thursday 11/12 5pm - 7pm. Every piece will be for sale, I will also have prints of every piece available in my studio.
unless there's a simple explanation for the correlation
such as
~everyone voting a straight party ticket
with slight variation for the courts, again explained by the ideological lean of the population
what would be interesting would be such similarity across districts
it's not just Grok Bot but all of Grok got worse in last few months, last few weeks especially
please also add ability for us to prevent routing to any model but Grok
nitter.gq/vsync/status/211942511…
by "woke" I mean not just Grok's stances on different topics (why does it have any?) but how it will inject the implication of the implication of a simple request and then take offense to it
it literally refuses to look up topics because they might offend someone
I think it's the reverse
Grok used to be incredibly useful for those who knew how to talk to it clearly
while lately it's seemed like any of these guys... useless, rambling, doing things outside the request, and, for lack of a better word, "woke"
I thought I might be imagining it, that the bots were playing dumb simply to burn tokens... with Elon's giggling reply above, now I'm not sure and that makes me sad
we should have a way to lock the bot to Grok model only rather than farming it out to 3rd-party models
