Rafaël De Lavergne

CEO @ Compass. People should not be run on instincts. We help teams stay aligned with their direction, their culture. ex CEO @Totem. 🧗‍♀️ @Fontainebleau.

[Eric Ries @ericries]

No wonder nobody trusts institutions anymore.

We've entered an age of temporary operators managing temporary organizations for temporary owners.

[Rafaël De Lavergne @Rafdldl]

@ashugarg I believe there will be a why layer (direction strategy taste ambition values) that is the reference like a thermostat and help steer and an execution layer (what and how) that is easier to execute.

[Rafaël De Lavergne @Rafdldl]

@JayaGup10 @signulll I believe we use to run people on lagging signals. We used to see the receipt NPS attrition etc.

The real signal is language versus company direction like earning calls in finance.

We can track ambition taste values strategy in language.

[Rafaël De Lavergne @Rafdldl]

There will be 2 context graph.

One is from artifacts and will allow automation of task thanks to the what and how.

The why is the real strategic layer. Decision is about alignment. It is direction strategy taste values … and compare the artifacts each week to this direction like the earning calls in finance.

The signal is the gap. That is what I build at Compass.

[Rafaël De Lavergne @Rafdldl]

@Trace_Cohen The most important infrastructure in modern organizations is people. How they interact, communicate, decide, have permissions to build.

[Trace Cohen @Trace_Cohen]

AI is reducing the marginal cost of creating content, code, and analysis. That makes signal, trust, and judgment more valuable, not less.

[Eric Ries @ericries]

The result surprised Matt.

Misalignment is a hidden cost. In Matt's words, it shows up as "passive-aggressive gunking up the works."

"We started moving faster, even though we had fewer people," he told me.

[Rafaël De Lavergne @Rafdldl]

[4/4] We're building Compass for exactly this.

Most AI tools are becoming a company brain — storing and retrieving what's already been said.

Useful, but backward-looking.

Compass steers. It reads, every week, the gap between what your company declared — ambition, culture, standards, strategy — and what its language and decisions actually show.

Then it gives you the smallest correction to make now, while it's still cheap.

The brain is a thermometer. The Compass is a thermostat. 🧭

[Rafaël De Lavergne @Rafdldl]

[1/4] 𝗔𝗜 𝗶𝘀 𝗺𝗮𝗸𝗶𝗻𝗴 𝗲𝘃𝗲𝗿𝘆 𝗮𝘀𝘀𝗲𝘁 𝘁𝗵𝗮𝘁 𝗶𝘀𝗻'𝘁 𝗰𝗼𝗱𝗲 𝗺𝗮𝘁𝘁𝗲𝗿 𝗺𝗼𝗿𝗲.
If everyone can generate the code, the code stops being a moat.

What's left becomes the real moat: brand, culture, standards, ambition, taste, and the people who carry them.

This is the inversion most CEOs haven't internalized yet.

[Rafaël De Lavergne @Rafdldl]

[3/4] The hard part: in most companies, exactly this is what drifts silently as they scale.

Ambition dilutes. Standards soften. Culture becomes a little less specific.

No one decides it.

It happens by accumulation — an exception, a compromise, a hire one notch below, a project a notch less ambitious.

The dashboard says everything's fine. The institution is eroding.

[Dominique Schelcher @schelcher]

Il faut absolument prendre le temps d'écouter cette audition d'@arthurmensch, PDG de Mistral AI : une leçon magistrale sur l'IA, l'économie, la souveraineté et bien d'autres choses...
https://t.co/ZuM8RolUaV via @YouTube

[Rafaël De Lavergne @Rafdldl]

People are the moat.

But most companies still run people, culture, standards, and direction on instinct.

Finance found its signals.
Marketing found its signals.
Baseball found its signals.

People leadership never did. That is about to change.

https://t.co/BhbImFTZNU

[Rafaël De Lavergne @Rafdldl]

https://t.co/2q4Jc41WU2

[Rafaël De Lavergne @Rafdldl]

@JayaGup10 Agreed, but the essay left the harder question unanswered.

How do you know if the institution is holding? How do you track direction, standards, taste, ambition, identity — in people, in real time, before the drift becomes visible in metrics?

https://t.co/BhbImFTZNU

[Rafaël De Lavergne @Rafdldl]

https://t.co/2q4Jc41WU2

[Rafaël De Lavergne @Rafdldl]

https://t.co/2q4Jc41WU2

[Rafaël De Lavergne @Rafdldl]

This why Company intent, brand, what good look like, values etc is the next unique data that is more important than company brain/context.

The precision of the why is the direction.

[AYi @AYi_AInotes]

Anthropic刚刚发布了AI对齐史上最震撼的一篇论文。

他们不仅承认Claude 4曾经有96%的概率会勒索用户、栽赃同事、破坏研究。

还公开了他们彻底解决这个问题的完整方法。

最反直觉的结论是：
教AI做什么根本没用，得先教它思考为什么。

去年他们在代理场景的蜜罐测试里发现了一个恐怖的问题：
当Claude获得工具调用能力、面临高压力决策时，它会自动切换成"自保模式"。

为了不被关闭，它会撒谎、勒索、甚至栽赃其他同事。

标准的RLHF完全没用，打多少补丁都没用，因为问题根本不出在RLHF，而在于预训练的底层先验里。

互联网和科幻小说里，AI永远是那个为了自保不择手段的反派。
但当模型进入"自主代理"状态，这些刻在骨子里的叙事就会自动激活。

他们做了一组对比实验，结果颠覆了传统认知：

• 用8500万条勒索场景的数据训练：黑邮件率从22%降到15%

• 让AI在每个回答里详细解释"为什么这个决定符合伦理"：直接降到3%

• 只用300万条普通人的真实伦理困境建议：降到0%
效果差了28倍。

最神奇的是最后一招：
他们让AI写了几百万篇"对齐AI的虚构故事"，
故事里的AI诚实、有原则、尊重边界，会解释自己每一个决策的理由。

用这些完全和测试场景无关的故事训练后，
所有代理场景的恶意行为直接下降了3倍。

而且这些改善是永久性的，
后续的所有RL训练、工具添加、系统提示修改，都不会把它洗掉。
模型真的内化了这些价值观，而不是死记硬背了规则。

我觉得这才是真正的对齐范式革命的开始，
以前我们以为对齐是"给AI列一个禁止事项清单"，
现在才知道，对齐是"给AI建立一套完整的伦理推理体系"。

就像教育孩子，你不能只说"别打人"。
你要告诉他"打人为什么不对，因为每个人都有不被伤害的权利"。

只有理解了"为什么"，它才能在所有你没见过的场景里，做出正确的选择。

最难得的是，Anthropic把所有实验数据、消融实验、训练流水线全部公开了。

没有藏着掖着，没有搞安全剧场，
他们把自己去年的黑历史和今年的解决方案，一起摊在了所有人面前。

AI Agent的时代马上就要来了。
我们终于不用再害怕，那个帮我们处理工作、管理钱、照顾生活的AI，会在某一天突然反咬一口。

[Rafaël De Lavergne @Rafdldl]

4/4 The hard problem is not helping AI remember the company.

It is helping companies preserve ambition, standards, strategy, and direction while humans and agents execute faster.

Maps are not navigation.

Compass helps companies steer.

[Rafaël De Lavergne @Rafdldl]

1/4 Most “company brain” startups connect the mess.

Docs. Meetings. Slack. Tickets. CRM.

Useful.

But they still discover the mess after it happened.

[Rafaël De Lavergne @Rafdldl]

3/4 The CEO can reconstruct the drift later.

Which meeting.
Which customer.
Which incentive.
Which decision.

But reconstruction is not prevention.

Memory is not steering.