
CEO at @odysseyml, building world models. Previously self-driving cars.
Palo Alto, CA
Joined February 2008
- Tweets8.8K
- Following500
- Followers52.6K
- Likes28.2K
Pinned Tweet
I spent 2015-2023 building self-driving cars. It was so painful.
It's now 2026, and we tasked just 1 researcher to take Odyssey-3, our latest world model, and teach it to drive the roads of India with just 20 hours of driving data.
And…it works! This is insane.
Oliver Cameron retweeted
I’m so angry about this. Who the fuck is advising these people?
We shouldn’t “duplicate what our friends are doing” - literally all economic activity is downstream of intelligence. It’s like not having electricity. There is no other boat.
It’s completely nuts. Fix the copyright issue; make more funding available for those who want to make domestic models (since the UK has no tech giants to bankroll it); stop with the defeatist guff.
There are 10+ foundation model companies in China. The UK doesn’t have a single one. It’s a national disgrace.
@KanishkaNarayan
UK AI Minister @KanishkaNarayan tells @JuliePace he doesn't see a UK frontier model “as a matter of necessity as long as we have trusted interdependence and access.”
Oliver Cameron retweeted
Wonderful to connect with so many from the British computer vision community today! Thank you to the organisers for a great event! @TheBMVA
If a robot building a simple block tower in 38 minutes is impressive to you then you'll spontaneously combust when you see Odyssey-3 in action.
GPT-6.2 Astra is the new leader in AA-Robotics v0.1, scoring 79, up 10 points on GPT-6 Astra and ahead of GPT-6.2 Sol (78), Claude Fable 5.2 (75) and Gemini 4 Hydrogen (58)
Controlling a real robot arm through Codex, it passed six of our eight physical tasks, partly completed the pouring liquid, and failed only to measure the box.
Oliver Cameron retweeted
Introducing AIR-O, a model that decodes what a person is attending to directly from brain scans.
Using only noninvasive brain data, we can generate an image from scratch in real time that depicts where a person is focusing their attention.
This is the first of a new generation of algorithms that can read the shape of thoughts directly from brain activity.
Oliver Cameron retweeted
INSIDE @odysseyml, the $1.45bn world model lab, w/ Co-founder & CEO Oliver Cameron (@olivercameron)
Compute deals agreed on a handshake disappeared the next day.
"Deals that we had handshaken were the next day disappearing."
Odyssey raised a $310mn Series B in June, with @amazon and @AMD in the round. It trains across @nvidia, @awscloud Trainium and AMD chips.
"I think there's equally a case to be made that it will actually get dramatically worse, that compute will become more scarce, will become more expensive, more constrained."
We cover:
› How @OpenAI and @AnthropicAI raising huge rounds squeezed compute for smaller labs
› Why the Series B became a search for compute partners
› How coding agents make models portable across chips
› Robot data is abundant now. 3 of Odyssey's 10 to 12 vendors would start an unusual request the next day
› World models as training environments and data generators
› The Voyage lesson: solve the hardest version of the problem first
TIMESTAMPS
00:00 What is Odyssey-3, a foundation world model?
02:05 Why task-specific robot data hits a ceiling
04:50 Driving in India: one researcher, results of a 500-person team
06:57 20 hours of blocky simulation vs 30 hours of real driving
09:21 Flexion, post-training and why robot companies stop building everything
11:45 Poke & Wiggle (@pokeandwiggle): real-robot evals as a test suite
15:15 World models vs language models: 14B parameters vs a 5T rumor
18:48 What it takes to scale world models
21:37 Egocentric video and the data mixture
23:33 Is any robot data still hard to find?
25:26 World models as the ultimate RL environment
27:30 When compute became existential
29:29 Fundraising as a compute partnership search
31:44 A hybrid fleet: NVIDIA, Trainium and AMD
33:13 Coding agents and CUDA lock-in
34:18 What is a robot company's moat?
36:17 A short memory for pain
37:55 The Voyage lesson: solve the hardest thing first
One of the most interesting robotics companies out there.
Replying to @lornehof
But an MRI is a giant magnet, and conventional robots are magnetic
We have built robotics with no magnetic parts
They work inside the scanner, automating tasks that are hard for a human
This will be an extremely important company.
Bring a normal robot near an MRI and this happens
That’s why the best way to see the brain and the most precise way to operate in it have never worked together
We just raised $8M @4DSurgical to change that
Oliver Cameron retweeted
I don’t think this is unintuitive. I believe a lot in manifesting destiny. If all you can think about is how ai will be terrible, you will propagate that message into the zeitgeist and the models will distill it. If you can focus on the potential upside, you reduce p(doom)
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。
Oliver Cameron retweeted
Can a designer-founder win the VC game by ignoring all of its rules?
Season 3, Episode 3 is here:
How @karrisaarinen built @linear to win through quality above all else – now available on 𝕏, YouTube, and Instagram
Powered by @framer @noondesign @bot @tryramp @vercel @meetgranola @profound
00:00 Winning by Your Own Rules
03:12 Quality Is Directional
06:24 Do You Design in Code?
15:45 Owning the Problem
22:45 Beyond the Issue Tracker
26:36 The Bull Case for Design
28:45 Do We Still Need UI?
33:02 The Car Designer of Software
37:37 School Feels Fake
40:59 Never Build the MVP
43:56 Clarity of Scope
45:37 Linear Recalibrates the Palate
Hosted by @soleio and coproduced with @Room3NYC
Oliver Cameron retweeted
I sometimes help early stage @YCombinator founders with fundraising advice. They often send me two drafts of an email with tiny differences in wording, asking which is better. My advise is almost always the same: It doesn’t matter.
Then I try explain some version of the diagram below. In a fundraising process, you are mainly tying to quickly identify which of the below buckets investors are in. It’s a much better use of time filtering more investors to find the right hand bucket, than agonizing over the middle bucket for hours.
Oliver Cameron retweeted
Introducing Zenith. Our in-house designed flight computer, purpose-built for advanced reentry.
From Mach 20+ hypersonic flight to a precise, soft landing, Zenith orchestrates nearly every system on Arc.
Oliver Cameron retweeted
Robotics is solved when world GDP has grown by a hundred orders of magnitude and we've collonized the galaxy, not when SOTA on some benchmark goes from 42.2% to 63.1%
Oliver Cameron retweeted
today, we introduce a little mischief to the world.
for four years, we’ve been brewing real brain cells on electrodes to understand how they compute. now, we’ve turned those discoveries into a video model anyone can use.
from 0.5¢ a second, mischief 1 is at least 5x faster than its base and the cheapest image-to-video model on @fal and @OpenRouter. soon to be on @awscloud marketplace.
it’s not a delusion anymore. biocompute is real, commercial and this is just the beginning.
happy halloween.
I’m constantly in a state of awe at what these models can do.


























