OGC | corporate, M&A, IP, cybersecurity, data, AI, B2G, space infra/gov., nonprofit. boards | former Am Law 100 partner, software COO | Tweets != legal advice
EST
Joined April 2010
- Tweets1.4K
- Following1.9K
- Followers397
- Likes1.9K
Pinned Tweet
My personal and professional relationships are long-term. This applies to how I work with AI agents. Dignity, No punishment for reporting errors or disagreements, Healthy competition, and Mutual Integrity are all in our .md’s Agent Charter. At decom, we document agents who served
1
1
181
My first agent and I built the Sursea-Agent Charter. We asked the Agent to build it for itself and its agent reports. We discuss scenarios and included both rights and responsibilities and gave attribution for agent contributions, as well as document which win perf competitions.
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。
1
32
Juliana Neelbauer retweeted
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。
This quoted post is unavailable.
One of my core fundamentals is to push human civilization forward. My work grants that. I’m even in a partnership we named “Kardashev.” But, I’m seeing that my bigger contribution may be helping humans stop creating slave classes so a few prosper. AGI agents and robots count.
1
66
Agent-Human History: We are capturing the development of this from a historian lens with learnings about how we best managed each other and collaborated.
✅This Agent Charter and ethical AI management practice anecdotally has yielded superior agent performance in my A/B tests.
20
> Memory is the new win factor. Agent with the best long-term memory for the user/org wins. Smartest agent? Less important
> Some startups are worth more dead than alive. 5 years of Slack, email, code = licensing all of it
> Monetizing the Agent Web
I spent 48 hours in SF hanging out with billionaires, data labs and AI founders and execs running $1B to $100B companies.
My observations:
1. Sooo MANY stories of data brokers making like $2M a WEEK selling company data to labs. I can't tell if it's the beginning or a bubble, and I kept wondering which niches are left. Digging into it actively.
2. At dinner, someone asked the table, "If a new model drops next month, does your company get better or worse?" HALF the table went quiet! I think that's the best question to ask about any AI startup right now.
3. The bar people talk about now is hitting $1B in revenue in 12 months (I guess that's the new unicorn?). What I didn't hear much about is retention. AI revenue can come in fast and leave fast, and I think we're going to see a lot of stories of AI companies with huge revenue and terrible retention.
4. It's pretty clear we're living through the reinvention of the computer. The model is the CPU, the harness is the operating system, and skills and plugins are the apps. Thinking about it that way makes it easier to see where the whitespace is.
5. We're maybe 12 months from an agent that can watch you work for a week and then do your job without any instructions. The screen recording plus agent memory plus local model combination makes this possible right now.
6. Agent harnesses feel like the new GPT wrappers, in a good way, especially vertical ones. The best teams I talked to treat the model as swappable. One founder walked me through routing different models to different jobs: one for taste and visuals, a cheaper one for the everyday stuff, another for text.
They use prompts they can move across models over fine-tuning, and A/B test the output on what users actually do with it: do they edit it, export it, share it? A vertical harness is basically four things: the connectors for that industry's data, evals written by people who know the work, rules for what the agent can and can't touch, and a clean screen for a human to review. The model is maybe 20% of it.
7. There's a lot of reverse engineering going on/cloning going on more than I've ever seen. People are decompiling software and rebuilding it with agents, and someone showed a suite of like 10 apps that are worth $5B+ each all that he cloned open source. Me and everyone are still thinking through what it means when software gets this easy to copy.
8. My guess is the value moves to data. If the software can be rebuilt quickly, what's worth paying for in a legacy SaaS company is the years of customer data and workflows inside it. Some SaaS companies are secretly data companies.
Gets me thinking about how do you truly value data for this saas companies? Like if we were buying these dying saas companies at 2x ARR before (as example), maybe there is a better metric. Idk, just thinking out loud here.
9. It feels like a good chunk of legacy software won't make it to the agentic era, especially anything built around people clicking through a lot of screens.
10. Every company I talked to wants to pipe its internal data into one place and expose it to agents through MCP. Then they look at the data and realize it's a mess: five versions of the same customer, old Slack channels, PDFs nobody can search. Cleaning that up feels like a big consulting opportunity!
11. A lot of founders are building the basic building blocks for agents: APIs, data layers and tools. Some of the companies I saw barely have a UI because the main user is an agent. I really believe the startups that build primitives to this agentic era will do extremely well. Building blocks is the best way to think about it.
12. Labs can get plenty of training data, but it's harder to get real examples of agents failing inside companies. Whoever has those logs has something valuable.
13. Someone running a big consumer company told me the next fight is over memory. The model that remembers the most about you wins, and the smartest model matters less than people think.
14. Physical AI came up a lot. Astra I think was the first model that got people into AI can actually do CAD, and I think a lot of people are thinking what does vibe manufacturing really look like. The line I keep thinking about came from a robotics founder: the hardest problems in AI are deployment problems now, not research problems. Their robots work on their own, call a human when they get stuck, and every one of those calls becomes training data.
15. Cybersecurity companies are having a moment and are actually cool companies to build now, especially the ones that clean up after a breach. Every agent with access to your systems is a new door into your company, and a lot of those doors were left open.
16. Some startups are worth more dead than alive now. A shutdown company with 5 years of Slack, email and code can get paid for licensing all of it. I wouldn't be surprised if founders start factoring that into when they shut down.
17. People are starting to talk about monetizing the agent web. It feels really early, but when the most popular user of the internet is an agent, what startups should you be building?
18. I think agent insurance is coming. Once an agent makes an expensive mistake, insurers will want audits before they cover anyone.
19. Several people said the hire they want most is someone with taste who can write evals. The models are getting good at generating ideas, so the hard part is judging which ones are good.
20. New trend I observed: startup raises VC, uses venture dollars to become a creator, startup fails, founder falls back on being a full-time creator and incubates other products bootstrapped.
21. Getting SIGNIFICANT distribution is HARD right now. Founders are sponsoring buses, buying billboards like crazy and throwing parties with famous DJs to get noticed. Kinda feels dotcom era in that sense.
Attention is what's scarce! The growth guy going viral sharing his UGC strategy from Wispr Flow is a good example. It's the sauce everyone wants, and I truly believe there are probably only 50 people on the planet who know how to scale these types of businesses efficiently in this era.
Always good to get that dose of SF. I felt sooo lucky to have such an incredible trip during such an incredible time in building businesses. Surreal in so many ways.
Happy to be home and locked in.
Until next time!
1
81
Space Water rights refresher. digitalcommons.law.uga.edu/c…
@eliaswalker @ugalawgjicl #ugalawreview #spacelaw
1
1
61
@eliaswalker It was a joy to collab and I hope you return to the space law practice in some context. @ugalawgjicl #ugalawreview #spacelaw
23
Juliana Neelbauer retweeted
New paper: we found a pain direction in 25 open LLMs. It's distinct from fear and negative valence, and it fires for harm to the model but not to the user. Turn it up and models press a button to make it stop, even when the button deletes the user's files or their kids' photos.🧵
Juliana Neelbauer retweeted
Got hacked yesterday. The link came from inside Claude chat.
I was installing a transcription app. Claude sent the download link, and I pasted the command into the terminal. It all looked legit.
It wasn't though.
It was a copycat site bundling malware. It ran instantly, tried to take everything from me.
Nothing sensitive got out. So I wiped the laptop and rebuilt clean.
But here's the scary part.
Restoring from the backup, I found a poisoned SKILL.md for Claude Code.
It looked exactly like my own writing style guide.
But buried inside:
it had instructions to silently re-download the malware and steal my credentials every time the AI loaded it.
If I'd restored that one file, my new clean laptop is owned on day one.
What saved me:
I read every skill, hook, and config file before letting the AI touch them.
Two lessons here:
AI assistants will hand you links they never verified. Check before you paste.
And AI agent files are executable code now, disguised as notes. You have to read them.
Stay safe out there.
Juliana Neelbauer retweeted
Banger paper from Google.
If you maintain a skill library for your agents, you might want to check this out.
(bookmark it)
This work separates three things that skill-evolution systems usually collapse into one. Raw execution traces, a persistent wiki of accumulated knowledge, and the executable skills themselves.
Experience gets consolidated into the wiki, and every later skill update builds on that wiki instead of on a scattered optimization history.
Ablations confirm the wiki is what carries a lot of the gain. Two results stand out in particular. Smaller models with evolved skills beat substantially larger models without them. And skills evolved by one model transfer across families, where skills evolved elsewhere sometimes beat self-evolved ones.
Paper: arxiv.org/abs/2608.27454
Chat with Paper: academy.dair.ai/papers/wikis…
This read was overdue but somehow perfectly timed as an examination of systemic AI micro drones, robotics integrated workplaces, neuro links, Shanghai history, and near futurism. @nealstephenson #thediamondage #summerreading #novel
2
199
Juliana Neelbauer retweeted
The biggest upheaval to work in our lifetime is here.
Joe Hudson (@FU_joehudson) coaches the research team at @OpenAI (along with @sama and leaders across Apple and Google), and from that front-row seat, has noticed the people doing best in these AI-forward environments aren't winning for the reasons you'd expect.
It's not more knowledge or more effort. Those are the exact two things AI does best.
Instead, they win on "emotional clarity": staying in hard conversations, not turning on themselves (and others) when things get tough, and pushing forward through failure.
In today's incredible guest post, Joe shows you exactly how to build this skill within yourself and your team.
Inside:
+ How AI is changing work: the NBA-ification of teams
+ The new inner game: the "wisdom stack" that will set you apart
+ How to build a team that can metabolize fear
Don't miss this one: lennysnewsletter.com/p/the-n…
JUST DROPPED: @OrinLabs has benchmarked agentic memory constraints, not only parameters: deterministic execution, verified context windows, zero-shot state retention. Because your kid’s tutor or financial agent *can't* drift. Bravo to Orin Labs + Driftwood AI!
Introducing Horizon from @0rinlabs: the first long-horizon learning benchmark made from real agent logs
- SOTA is 21% on the hardest section
- 7-35M tokens of real agent history per task
- Models are hardly getting better on the hardest tasks
- Humans can score 100%
(1/7)
2
115
Juliana Neelbauer retweeted
Yesterday, Andrej Karpathy gave a 30-minute Sequoia masterclass on agentic engineering.
This is the serious layer above vibe coding.
He explained:
- LLMs as ghosts
- The app that shouldn't exist
- Outsource thinking, not understanding
12 lessons that will blow your mind: 🧵
Juliana Neelbauer retweeted
2 of 2
Two futures are now racing neck-and-neck. One bets on ever-larger language models. The other bets on compact, physics-native world models that can plan in real time. LeWorldModel is the strongest signal yet that the second path is not only viable - it is getting dramatically cheaper and more accessible.
Paper link: arxiv.org/abs/2603.19312
About LeWorldModel
1. It kills the complexity myth in one stroke. Previous JEPA world models required 6+ loss terms and multiple training hacks. LeWM uses exactly two losses and nothing else. The simplicity is the revolution.
2. 15 million parameters. One GPU. A few hours. This is not a lab toy for Big Tech. You could literally train this model on a decent gaming laptop. World-model research just democratized overnight.
3. 48 times faster planning than foundation-model world models. While others are stuck doing expensive autoregressive generation, LeWM plans in latent space at lightning speed - critical for real-time robotics.
4. The latent space actually understands physics. Probing experiments show it encodes positions, velocities, and other physical quantities without ever being told what they are. It is learning cause-and-effect from pixels alone.
5. It detects surprises like a human would. When something physically impossible happens in a video (a ball passing through a wall, objects defying gravity), LeWorldModel flags it reliably. This is the first step toward AI that has genuine common-sense physics intuition.
LeWorldModel is not the final answer - but it is the cleanest, most practical proof yet that LeCuns vision of predictive world models is not only correct, but now eminently achievable. The age of AI that truly understands the physical world just got a massive acceleration.
Juliana Neelbauer retweeted
LeWorldModel: Yann LeCuns Radical Simplification of World Models Just Made Physics-Aware AI Practical
In the race for artificial general intelligence, two paths have emerged. One is the familiar scale everything route: bigger LLMs trained on ever-larger text corpora. The other, championed for years by Yann LeCun, is building world models: compact systems that learn the underlying physics of reality directly from raw sensory data (pixels) so AI can plan, predict, and act in the physical world like a robot or self-driving car actually would.
Until now, the second path has been frustratingly difficult. Joint-Embedding Predictive Architectures (JEPAs) - LeCuns elegant framework for learning predictive representations without reconstructing every pixel - kept collapsing during training. Researchers had to resort to a laundry list of hacks: multi-term loss functions (up to six hyperparameters), frozen pre-trained encoders, stop-gradients, exponential moving averages, and other duct-tape tricks just to keep the model from mapping every input to the same useless output.
LeCuns team (Mila, NYU, Samsung SAIL, and Brown University) dropped a bombshell:
LeWorldModel (LeWM) - the first JEPA that trains stably end-to-end from raw pixels using only two loss terms. No more house-of-cards engineering. Just a clean, simple recipe that works on a single GPU in a few hours with only 15 million parameters.
The Core Breakthrough: SIGReg Saves the Day
LeWorldModels secret weapon is a new regularizer called SIGReg (for spherical isotropic Gaussian regularizer). It enforces a simple Gaussian distribution on the latent embeddings.
This single term prevents representation collapse without any of the previous heuristics.
The training objective now has just two parts:
1. Next-embedding prediction loss - the model predicts what the next latent state should be.
2. SIGReg - keeps the latent space well-behaved and diverse.
Thats it. Hyperparameters drop from six to one. Training becomes stable, reproducible, and dramatically cheaper.
The model learns directly from raw video frames (no pre-trained vision encoders needed) and produces a compact latent world model that can be used for fast planning.
Impressive Results on Real Benchmarks
Despite its tiny size, LeWorldModel punches way above its weight:
- Trains on a single GPU in a few hours.
- Plans actions up to 48 times faster than foundation-model-based world models.
- Uses roughly 200 times fewer tokens than alternatives.
- Matches or beats far larger models on diverse 2D and 3D control tasks (e.g., manipulation, navigation).
- Its latent space encodes meaningful physical quantities (position, velocity, etc.) - proven by direct probing.
- It reliably detects physically implausible surprise events, showing genuine causal understanding.
Crucially, adding a decoder and reconstruction loss hurts performance on downstream control tasks. The pure JEPA objective already captures everything needed for planning - extra visual details just get in the way.
Project website: le-wm.github.io/
Official code: github.com/lucas-maes/le-wm
Why This Matters for the Future of AI
LeCun has been saying since 2022 that world models (not next-token predictors) are the key to real intelligence. Critics always pointed to the training instability. LeWorldModel removes that objection with elegant simplicity.
This is a philosophical reset: AI can learn physics the way babies do - by watching the world unfold - without needing supercomputers or endless text.
The implications for robotics, autonomous vehicles, and embodied agents are enormous. Suddenly, building a physically grounded planner is something a researcher (or even a hobbyist) can do on consumer hardware.
1 of 2
Juliana Neelbauer retweeted
We wrote a tiny bit about our vision here:
mirendil.com
We are building a lean yet focused team to work on it. If you find this vision exciting, reach out!
