我为何离开 OpenAI，加入 Conduit 开发心灵感应技术

在7月23日星期四，我辞去了在OpenAI的职位。24日，我作为Conduit的创始研究员开始工作：https://condu.it/。我们正在构建心灵感应：从非侵入性神经数据训练的思维转文本模型。

我的预测是：几年后，我们与AI交流的主要方式将是通过思维。他们不会成为一个超级聪明的自动化实习生或同事：https://x.com/sama/status/1983584366547829073?lang=en，你拼命想要跟上他们。他们将是你自然、愉快的延伸。

我在下方写下了这个未来的片段。我认为它们是乐观的，但非常可能发生。

我戴上头环，打开Conduit的笔记本电脑。现在是上午9点。我的设备通过蓝牙与笔记本配对。我打开了Codex，不失一般性。

我在看代码、审查昨天的笔记并准备我们的9:30早会时，思考当天的工作。我的GPT-7代理已经在我探索的新编码器上工作了一天。我看着它的成果。看到图表标签，我感到困惑；看到第一段的AI语言，我感到恼火；看到公式中的符号，我感到好奇。我模糊的思维被发送到Conduit的模型，该模型利用其对语言及我可能说的话的先验知识输出：

图表、段落和符号我各花了10秒浏览。我开启了自动发送，因此每10秒，Conduit模型就会解码我的思维，并发送给GPT-7，GPT-7将每个接收到的任务分配给子代理。我看着新的图表、段落和公式。我想知道球谐是否可能有用；Codex派生出一个子代理。被图表说服后，我决定制作一个快速幻灯片，只显示基线和前两条线，美化后给队友看；Codex又派生出一个子代理。

我给自己倒咖啡。我的思绪大多空白，但我短暂回忆起明天与候选人共进午餐时想说的话。在背景中，Codex被提示思考我午餐谈话可能需要的信息，并决定在展示之前更新合成数据缩放图是明智的；Codex又派生出一个子代理。

我喝咖啡时不会在脑海里大声说话。我只是像平时一样读剧情，照常泡咖啡。感觉就像魔法一样。

人工智能公司现在训练模型直接与Conduit的潜在表征接口。我编码的思想直接传送到Codex，而不是先经过Conduit的解码器模型。这意味着我能轻松传达那些难以用文字描述的想法，比如心理图像。

现在我很少在和AI聊天时不戴乐队。说实话，没有它真的很烦人。戴上神经头带，我感觉自己对笔记本电脑拥有超能力！没有它，我感觉自己像是在和一个超能力外星人对话，他尽力帮忙，但不确定自己想要什么，害怕如果它做了我不想要的事我会生气。呃。

Conduit仍在不断改进非侵入式读取，但我们推出了两个新项目：侵入式通用阅读和通用写入。

大多数人愿意继续使用神经带，但也有不少人期待通过侵入性技术获得更高保真度的读数。

更有趣的是最近写作的热潮。我感觉自己对笔记本电脑的掌控感超乎常人。但我的感官依然只是人类。就像我能控制手臂，但却失去了所有知觉。是的，我还能看到手臂，这比没有手臂好多了，但还是很奇怪。比如伊恩·沃特曼：https：//mcneilllab.uchicago.edu/pdfs/IW_lost_body.pdf。我想感受我的法典所带来的感觉。现在Conduit做通用读取，学会通用写入对许多OOM的数据效率更高。

更重要的是，我想解锁写作技术的其他应用。我想让我的大脑像18岁时那样高效且神经可塑性强。

我的人工智能是我自然的延伸。感觉就像第六感和另一条肢体。我会思考一个问题，感觉自己真的在上网浏览，偶尔看到网站的片段，获得对问题的直觉闪现，最终得出答案。感觉很有趣！我的大脑就像闪电侠。因为写作技术，即使关闭AI，我的思维速度也是最快的。

2035年的超级人类样子，直接取决于我们作为社会决定让我们的人工智能变得多么超级。也许我们会选择循序渐进：https://www.pacingthefrontier.com/。但我喜欢这是一个真正“人类在环”的未来愿景，其中人工智能直接增强人类，而不是取代我们。

比我更优雅的作家和演讲者在过去十年中对这种思维-人工智能接口进行了阐述。我会链接给他们，并附上简短的摘录。

理论上，这很简单。我们的输入是大脑活动，我们的目标输出是当时人正在做的事情——例如，人写的文本。给定大脑活动，我们希望预测出语义上与人写的内容相似的输出。

为了训练可以根据大脑信号预测文本的模型，我们必须应用那些根据语音音频预测文本或根据前文预测文本时所学到的同样教训：痛苦的教训：http://www.incompleteideas.net/IncIdeas/BitterLesson.html。该教训大致表述为，你应该把更多有用的计算资源投入到你的模型中，你的模型将比任何你手工设计的巧妙算法都更好。这意味着我们必须将数据收集规模提高到远超以往学术界的数量级。

使用侵入式方法很难收集足够的数据。很少有人愿意在脑子里植入芯片！但是非侵入式方法正在变得越来越好。硬件在改进且变得更便宜，尽管我对我们硬件的具体细节描述得比较模糊，表示抱歉。

随着我们在更多数据上进行训练，模型预测的文本与被试写的文本在语义上更加相似。是的，由于噪音存在一些不可约的误差，但对于大多数模态，我们还没有达到受此限制的阶段。具体来说，尺度规律看起来很好：我们潜在空间预测与目标潜在空间的余弦相似度，随着数据小时数的对数增加呈直线增长。我们处于GPT-2时代。

我们不需要完美的解码才能有用。你的思维就像城市里的GPS：噪声信号不足以确定你确切的位置。但结合地图和导航路线——也就是LLM和上下文——它会变得异常准确。

想了解我们如何进行数据收集，可以查看Conduit的博客文章：https://condu.it/thought/10k-hours。如果你更喜欢主动学习类型，可以来成为我们的研究参与者：https://condu.it/booking!

总之：我很开心！我正在研究一个令我着迷、极具挑战性的问题，并且和一小群我喜欢的人一起工作。

对于不认识我的人：你好！我两周前离开了OpenAI，在那里我作为研究员工作了1.5年。

我在华盛顿州长大，从5岁到15岁一直参加国际象棋比赛，并在成为WIM后停止：https://en.wikipedia.org/wiki/Naomi_Bashkansky。我第一次听说超越人类智能的AI潜力是在12岁，但多年里我以为这只是互联网上人们喜欢写的一种奇怪但有趣的想法。

18岁时，我进入哈佛大学学习计算机科学。在一门课上，我了解了GPT-3，并最终读了Bostrom的《超级智能》。哇，那真是一个惊醒的时刻。我开始投身于AI安全研究，短暂从事AI政策工作，最终加入了OpenAI的对齐团队。在OpenAI，除了我的研究工作，我还参与了一些副项目，包括制作OpenAI的AGI入职演示和我们的对齐博客：http://alignment.openai.com，并帮助指导OpenAI基金会的AI恢复力部门。

如果你想聊聊Conduit，可以通过naomi@condu.it联系我：mailto:naomi@condu.it。我们可以喝咖啡，或者我带你参观我们位于旧金山不寻常又美丽的办公室。我们一直在招聘研究员、基础设施人员和运营人员。

鸣谢：感谢Aidan Smith、Devansh Pandey、Julia Shephard、Lev Chizhov、Ryan Kaufman以及Conduit团队的反馈。我也很感激其他正在开发雄心勃勃BCI技术的公司，包括但不限于那些致力于改善睡眠、提升超声技术、解决失明问题和上传技术的公司——我很期待使用你们的技术！所有错误均为我本人所为。

On Thursday, July 23rd, I resigned from OpenAI. On the 24th, I started as a Founding Researcher at Conduit：https://condu.it/ . We're building telepathy: thought-to-text models, trained on non-invasive neural data.

My prediction: In a couple years, the main way we'll talk with our AIs is with our thoughts. They will not be ：https://x.com/sama/status/1983584366547829073?lang=en a super-smart automated intern or coworker：https://x.com/sama/status/1983584366547829073?lang=en that you desperately try to keep up with. They will be a natural, joyful extension of you.

I wrote down vignettes of this future below. I view them as optimistic but highly plausible.

I put a band around my head, and open my laptop at Conduit. It’s 9 am. My device pairs with my laptop over Bluetooth. I open, without loss of generality, Codex.

I think about the day's work as I look at my code, review yesterday's notes, and prepare for our 9:30 am standup. My GPT-7 agent has been chugging away at the new encoder I’ve been exploring for over a day now. I look at its work. I become confused by the plot labels, annoyed at the AI-speak in its first paragraph, and curious about the symbols in the equations. My vague thoughts get sent to Conduit's model, which uses its priors over language and the kinds of things I might say to output:

The plot, paragraphs, and notation each took me 10 seconds to glance over. I have auto-send on, so after each 10-second chunk, the Conduit model deciphers my thoughts and sends it to GPT-7, which assigns each incoming task to a subagent. I look at the new plot, paragraph, and equations. I wonder whether spherical harmonics might be useful after all; Codex spins off a subagent. Convinced by the plot, I decide to make a quick slideshow with a prettified version showing only the baseline and top two lines, to show my teammates; Codex spins off a subagent.

I make myself coffee. My thoughts are mostly empty, but I briefly recall what I want to say to the candidate I’m getting lunch with tomorrow. In the background, Codex is prompted to think about info I might want for the lunch chat, and decides that updating the synthetic data scaling plot before showing it off would be prudent; Codex spins off a subagent.

I'm not saying words really loudly in my head while getting coffee. I just read the plots as I normally do, and make coffee as I normally do. It feels like magic.

The AI companies now train their models to directly interface with Conduit’s latent representations. My encoded thoughts get sent directly to Codex, rather than having to pass through Conduit's decoder model first. That means I easily communicate thoughts that are hard to describe in text, like mental images.

I rarely go band-less when chatting with AIs these days. It's annoying, honestly, to go without. With the neural headband, I feel like I have superhuman powers over my laptop! Without it, I feel like I'm talking with a superpowered alien who's trying its best to be helpful but isn't sure what I want and is scared I'll get mad if it does something I don't want. Ugh.

Conduit continues to iterate on non-invasive read, but we’ve spun up two new efforts: invasive general read, and general write.

Most people are happy to stick with their neural bands, but a good number are excited to get higher fidelity reads via invasive tech.

More interesting is the recent excitement in writes. I feel superhuman in my control over my laptop. But my senses are still merely human. It's like if I could control my arms, but I'd lost all feeling in them. Yeah, I can still see my arms, and it's way better than not having arms, but it's still really odd. Like Ian Waterman：https://mcneilllab.uchicago.edu/pdfs/IW_lost_body.pdf. I want to feel what my Codex feels. Now that Conduit does general read, learning to do general write is many OOMs more data efficient.

More importantly, I want to unlock the other applications of write technology. I want to make my brain as efficient and neuroplastic as when I was 18.

My AI is a natural extension of me. It feels like a sixth sense and another limb. I wonder about a problem, feel as though I’m literally surfing the web, see glimpses of the websites, get flashes of intuition about the problem, and ultimately derive the answer. It feels fun! My brain is like the Flash. Because of write tech, my thinking is the fastest it's ever been even when I turn my AI off.

How superhuman 2035 looks depends directly on how superhuman we, as society, decide to make our AIs. Perhaps we choose to pace ourselves：https://www.pacingthefrontier.com/. But I like that this is literally a human-in-the-loop vision of the future, where AI directly empowers humans rather than replacing us.

Writers and speakers more elegant than I have elaborated on this thought-to-AI interface over the past decade. I will link to them, along with brief excerpts.

In theory, it’s simple. Our input is brain activity, and our target output is what the person was doing at the time – for example, what text the person wrote. Given the brain activity, we want to predict output that is semantically similar to what the person wrote.

To train models that can predict text given brain signals, we must apply the same lesson learned by those predicting text given speech audio, or text given preceding text: the bitter lesson：http://www.incompleteideas.net/IncIdeas/BitterLesson.html . The lesson roughly states that you should throw more useful compute at your model, and your model will become better than any ingenious algorithm you could've hand-crafted. That means we must scale up our data collection by orders of magnitude beyond what has ever been done in academia.

It’s hard to collect enough data using invasive methods. Few people want a chip in their brain! But non-invasive methods are getting much better . The hardware is improving and getting cheaper, though I apologize for being vague about the particularities of our hardware.

As we're training on more data, the model is predicting text that is more semantically similar to the subject-written text. Yes, there's some irreducible error due to noise, but for most modalities we're not yet in a regime where we're pushing against that. Concretely, the scaling laws are looking good: the cosine similarity of our latent space predictions with the target latent spaces goes up as a straight line with respect to the logarithm of the number of hours of data. We're in the GPT-2 era.

We don’t need perfect decoding to be useful. Your thoughts will be like GPS in a city: a noisy GPS signal isn’t enough to determine your exact location. But combined with a map and a navigation route — equivalently, the LLM and context — it becomes remarkably accurate.

To read up on how we do data collection, check out Conduit's blog post：https://condu.it/thought/10k-hours. If you're more of the active learning type, come be a research participant：https://condu.it/booking!

In short: I’m happy! I’m working on a problem I’m obsessed with, that is frighteningly ambitious, with a small group of people I like.

For those who don’t know me: hello! I left OpenAI two weeks ago, after spending 1.5 years there as a researcher.

I grew up in Washington State, where I played competitive chess from ages 5–15 and stopped after becoming a WIM：https://en.wikipedia.org/wiki/Naomi_Bashkansky . My first time hearing about the potential for smarter-than-human AI was when I was 12, but for years I figured it was just a weird but interesting idea that people on the internet liked to write about.

At 18, I enrolled in Harvard to study computer science. In a class, I learned about GPT-3 and finally read Bostrom’s Superintelligence . Wow, what a wakeup moment. I got invested in AI safety research, had a brief stint in AI policy, and ultimately joined the alignment team at OpenAI. At OpenAI, other than my research, I spent some time on various side projects, including making OpenAI’s AGI onboarding presentation and our alignment blog：http://alignment.openai.com , and helping advise the AI Resilience division of the OpenAI Foundation.

If you want to chat about Conduit, reach out to me at naomi@condu.it：mailto:naomi@condu.it . Let’s grab coffee, or I’ll give you a tour of our unusual and beautiful office in San Francisco. We’re always hiring researchers, infra folks, and operators.

Acknowledgements: Thank you to Aidan Smith, Devansh Pandey, Julia Shephard, Lev Chizhov, Ryan Kaufman, and the Conduit team for feedback. I'm grateful, too, for the several other companies building ambitious BCI tech, including but not limited to those building better sleep, better ultrasound, a solution for blindness, and uploads - I'm excited to use your tech! All mistakes are my own.