新智元报道
科幻片黑客帝国、超验骇客里的桥段，正被马斯克一点点搬进现实。
今天，Neuralink甩出了一项最新进展，属实有点震撼——
这一次，马斯克这次直接把SpaceXSI的Grok，直接接进了人类大脑！
患者只要在脑海里默念一句「Hey Neuralink」，紧接着想一下要干的活，电脑上的Grok Agent就会瞬间被唤醒。
接下来，Grok便会全自动接管屏幕，直接登上了YouTube。
还有最戳人的一幕，是一通特别的电话。
重度瘫痪失语患者大脑里的芯片直接把「想说的话」转化成纯净的音频流，实时注入iPhone的麦克风通话底层。
时隔七年，他终于用意念给妻子拨通了电话。
人类距离真正的「机械飞升」，又迈出了极其疯狂的一步。
这一切能成，全靠Neuralink团队几天前干的一场「硬仗」。
几位核心工程师直接打包行李，住进了受试者家里，同吃同住，搞了场整整72小时的极限闭关开发。
整整三天三夜，他们把平时几个星期的活儿硬是在几个小时里给啃了下来，一口气爆肝出8个全新功能。
过去，我们对脑机接口的认知还停留在「用意念挪动鼠标光标」。
这72小时之后，大脑正式成为了AI的原生输入端。
这位患者名叫Kenneth，确诊渐冻症（ALS）整整七年，是Neuralink P21受试者。
一路走来他特别乐观，就像他常说的那样，「ALS很糟糕，但我的人生不是」。
这次最有科幻感的功能，当然是Grok。
脑机患者只需在脑子里想象说出「Hey Neuralink」，再加上希望完成的任务，就能触发一个Grok智能体，由它在电脑上自主执行。
手不必敲键盘，嘴也不必真的发声，指令从神经活动这一端发出。
把这条链路拆开，就容易理解了：脑机接口负责把表达意图转成可用输入，Grok智能体接住任务，再去操作电脑。
所谓的「接入」，本质上是一次交互维度的彻底重构：
- 意念控光标：大脑输出的是「往左一点、点一下」这样的底层动作，完成一个任务要几十上百个步骤，人得全程盯着；
- 意念驱动Agent：大脑只输出一句「意图」，拆解步骤、操作界面，全部交给AI。
脑机接口最大的瓶颈，一直是带宽：电极数量和解码速度都有物理上限。
Agent的价值在于，它把一句话「放大」成一连串操作——不加一个电极，就把每一次意念能完成的事情成倍放大。
就像Neuralink自己说的，这只是一次早期摸索，但它实打实地证明了：
只要大脑和电脑之间的沟通变顺畅了，「想到」和「做到」之间那道坎，就能被抹平。
这72小时里诞生的8项技术原型，从底层的信号捕捉，一路打通了系统级交互，最终接入了顶层的AI大模型。
- 神经语音活动检测（VAD）：彻底扔掉物理开关
在这次黑客松之前，Kenneth想要靠意念说话，还得面临一个很尴尬的悖论——你必须先手动按一下按钮，告诉软件「我要开始想了」，说完了还得再按一下停止录音。
对于渐冻症（ALS）或高位截瘫患者来说，按这个按钮本身就是巨大的负担。
为此，工程师直接从底层重写了一套「神经语音活动检测（Neural VAD）」系统。
这套算法就像一个24小时待命的神经哨兵，实时监听植入物传回的脑电信号流。
它能极其敏锐地分辨出，大脑此时此刻到底是处于「放空发呆」的静息状态，还是已经调动运动皮层「准备开口说话」。
一旦捕捉到说话的意图脉冲，APP会在毫秒级内自动开启解码；话音刚落，系统自动休眠。
全程零按键、零动手，真正实现了纯意念的「张口即来」。
- 实时神经信号可视化：患者和模型「一起训练」
脑机接口领域一直有个让人头疼的黑盒难题：连受试者自己都不知道，到底在脑子里「怎么想」，才能激发出最强烈、最清晰的神经元放电？
要知道，输入信号的信噪比越高，后端机器学习解码器的准确率就越逆天。
团队把植入物检测到的信号强度和空间分布，做成实时画面反馈给Kenneth。他可以换着花样「在脑中发音」，当场看到哪种方式调动的神经活动最多。
这在神经科学里属于闭环神经反馈：不只是模型在适应大脑，大脑也在反过来适应模型。
原本看不见、摸不着的调参，变成了一个有即时反馈的「闯关游戏」。
过去几个月，Neuralink内部一直在悄悄打磨一款名为Bloom的专属应用，目标只有一个：把庞大的言语脑机解码系统，装进一部手机里。
借着这次黑客松，工程师干脆推着Kenneth走出家门，带上Bloom，直奔嘈杂的公共场所。
在一家真实的咖啡馆里，Kenneth第一次没有靠旁人传话，而是通过手机上的Bloom，用意念给自己点了一杯咖啡。
而这，还只是开始。
对ALS患者来说，苹果自带的软键盘和Siri语音听写几乎用不了。于是团队直接调用iOS底层接口，手搓了一套系统级脑机输入法：BCI Keyboard Dictation。
只要调出键盘，Kenneth脑海中想说的话，就会经过神经解码，像水流一样源源不断地输入iPhone上的任意一个App。
写备忘录、回消息、刷社交网络，全程无需在不同模式间来回切换，指哪打哪。
当然，解码难免出错。
为此，工程师在黑客松期间连续迭代了好几版纠错界面，最终把一套预测式修改流程直接嵌进了键盘：一旦识别出现偏差，键盘上方会立刻弹出几个高概率的候选词，Kenneth用意念一扫，就能在任意App里完成「光速替换」。
还有一个细节，把日常沟通的启动门槛降到了几乎为零。
工程师巧妙借用了iPhone的「实时活动」和灵动岛机制，让Bloom常驻在锁屏界面上。
Kenneth不用解锁手机，屏幕一亮，就能看到当前的脑机连接状态；念头一动，就能直接在锁屏状态下开口「说话」，真正做到「秒接话」。
在所有功能里，「用意念拨通电话」的技术难度不算最高，却成了全场的催泪弹。
技术上，团队把Bloom和iPhone通话打通，让脑机接口生成的语音，直接接入通话的麦克风音频流。
但这件事的分量，不在技术，而在故事。
Kenneth和妻子的爱情，就是从电话里开始的。两人刚认识时，感情是靠一通又一通电话慢慢聊出来的。那根电话线，曾是他们之间最重要的一根线。
后来，疾病让他再难开口。那根线，也渐渐沉默了。
这一次，工程师们刚把通话功能做出来，他就拿起手机，拨给了妻子。
电话接通。他没有张嘴。脑海里想说的话，被植入体捕捉，被解码器翻译，再合成为声音，直接送进通话线路。
听筒那头，传来了妻子的笑声。
两人开始互相打趣、斗嘴，就像当年在电话里谈恋爱时一样。
工程师们就站在一旁看着。Neuralink后来说，这是整整一周里最难忘的时刻之一。
在工程领域，很少有什么比这更有成就感。你看着一个人，用你当天早上才写出来的代码，去和他爱的人说话。
过去，大脑带宽太窄，人只能一个字母、一次点击地向机器「挤牙膏」。
现在，大脑只需要输出一句意图，剩下的拆解、执行、操作，交给AI。当Grok接入脑机接口，大脑第一次有望成为AI的直接输入端。
这将成为人机交互的一次重大跨越。
未来，随着电极更多、解码更快、Agent更聪明，「想到即做到」或许会从实验室走进更多人的日常。
「黑客帝国」里，尼奥接上一根线，几秒就学会了功夫；「超验骇客」里，人的意识与网络相连，可以直接调动整个数字世界。
这些曾经只属于银幕的想象，正被一点点写进现实。
参考资料：https://neuralink.com/updates/72-hour-voice-hackathon/
编辑：桃子
不喜欢
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
New Ji-won.
Science fiction hacking empires, super hacking bridges, are being moved to reality by Mask.
Today, Neuralink has made an update, and it's quite a shock.
This time, Mask directly connected SpaceXSI's Grok to the human brain!
All the patients have to do is say "hey Neuralink" in their head, and then think about what they're doing.
Then Grok will take over the screen and go straight to YouTube.
And the most dramatic scene, a special call.
The chip in the brain of the severe paralysing loss-of-speaker directly transforms "whatever you want" into a pure audio stream that injects iPhone's microphones in real time.
Seven years later, he finally telephoned his wife.
The human race has taken another crazy step from the real "mechanical rise".
This is all possible thanks to a "hard fight" by the Neuralink team a few days ago.
Several of the core engineers packed their bags and stayed at the testee ' s house, where they ate together and engaged in a 72-hour maximum closure development.
For three days and three nights, they gnawed their normal weeks of work into hours and gave eight completely new functions.
In the past, our perception of brain interfaces has remained in the "methical move mouse cursor."
After these 72 hours, the brain officially became the primary input end of AI.
The patient, Kenneth, was diagnosed with thaw (ALS) for seven years and was tested by Neuralink P21.
He was particularly optimistic about coming all the way, as he used to say, "Als is bad, but my life isn't."
The most science fictional function this time, of course, is Grok.
A brain-brained person who imagines "hey Neuralink" in his head, and the tasks that he wants to do, can trigger a Grok AI agent, which he owns on a computer.
Hands don't have to knock on the keyboard, mouths don't really sound, instructions come from the end of neuroactivity.
Dismantling the link is easy to understand: brain interfaces are responsible for translating expressions of intent into usable inputs, Grok AI agent takes the task and then operates the computer.
The so-called "access" is essentially a complete re-engineering of the interactive dimensions:
- A conscious calibration: the brain produces a low-level action, like a little to the left, that takes hundreds of steps to complete a mission, and that one has to keep an eye on it;
- Conscious drive Agent: the brain only outputs one word of intent, decomposition steps, interfaces, all to AI.
The biggest bottleneck in brain interfaces has been bandwidth: the number of electrodes and the speed of decodes have physical limits.
The value of Agent is that it "expands" one sentence into a series of operations — that without an electrode it multiplies every thought that can be accomplished.
As Neuralink himself said, it was an early search, but it proved:
As long as the communication between the brain and the computer becomes smoother, the "think" and "do" can be wiped out.
The eight technological prototypes that were born in these 72 hours, captured from the bottom of the signal, reached the system-level interface and eventually reached the top level of AZLLM.
- Neural speech activity detection (VAD): Drop the physical switch.
Before this hacker pine, Kenneth wanted to talk on his mind, and had to face a very awkward paradox. - You have to push the button manually, tell the software I'm gonna start thinking, and then you have to stop the recording.
For persons suffering from progressive freezing (ALS) or high levels of paralysis, pressing this button is itself a huge burden.
To that end, the engineer rewritten a Neural VAD system directly from the ground floor.
The algorithm is like a 24-hour-on-a-clock nerve sentry, listening in real time to brain-telecom flows from implants.
It can tell with extreme sensitivity whether the brain is at the moment in a static state of emptiness or whether it has mobilized the motion cortex to "preparate to speak."
Once the intent pulse of the speech is captured, the decoding is automatically activated by the AP within milliseconds; the voice is just down and the system is automatically dormant.
The whole trip was zero-key, zero-action, and it really came to being pure.
- Real-time neural signal visualization: patients and models "train together"
There's always a headache about the black box in the brain interface: not even the testee himself, but what's in his head to get the strongest and clearest neurons discharged?
You know, the more the signal is entered, the more accurate the back-end machine is to learn the code decoder.
The team used the signal strength and spatial distribution detected by the implants as a real-time feed back to Kenneth. He can say it in his head, and he sees the most neuroactivity in the field.
This is in neuroscientific closed-ring neural feedback: not only is the model adapting to the brain, but also the brain adapting to the model.
The invisibility and invisibility of the move turned into an instant feedback game.
Over the past few months, within Neuralink, a unique application known as Bloom has been quietly polished, with only one goal: to decode a large speech brain into a cell phone.
With this hacker pine, the engineer simply pushed Kenneth out of the house, took Bloom and headed straight for a noisy public place.
For the first time in a real café, Kenneth had ordered himself a cup of coffee by means of Bloom on his mobile phone rather than by someone else.
And that's just the beginning.
For ALS patients, the apples' own diskettes and Siri's voice-writing were almost useless. The team then called the iOS bottom interface, hand-stamping a system-level brain input method: BCI Keyboard Notification.
When you pull out the keyboard, what Kenneth wants to say in his head, you decode through the nerve and enter any of the Apps on the iPhone like the current.
A memo is written, a message is sent, a social network is painted, and there is no need to switch between different modes.
Of course, the decoding is inevitable.
For this reason, the engineer has successively developed several versions of the error correction interface during the period of the hackers, which has resulted in a set of predictive modifications being embedded directly into the keyboard: once the deviation is identified, several high-probability candidates will pop up on the top of the keyboard, and Kenneth will be able to complete the "light-speed replacement" in any App.
There is also a detail that lowers the threshold for day-to-day communication to almost zero.
The engineers cleverly borrowed the iPhone's "real-time activities" and the Visible Island mechanism to keep Bloom on the lock interface.
Kenneth does not unlock his cell phone, and as soon as the screen is clear, he can see the current brain-to-heart connection; as soon as he moves his mind, he can say "talk" directly under the lock-in, and actually "second-to-second."
In all of these functions, the technical difficulty of "intentional phone calls" is not the highest, but it is a full tear gas.
Technically, the team communicates Bloom and iPhone, allowing voice from the brain interface to directly access the microphone stream of the call.
But the weight of the thing, not the technology, but the story.
Kenneth and his wife's love started on the phone. When they first met, their feelings came from one telephone call and another. That phone line was the most important line between them.
Later, the disease made it difficult for him to speak. That line is getting quiet.
On this occasion, as soon as the engineers made the call, he picked up his cell phone and assigned it to his wife.
Phone connection. He didn't open his mouth. What he wanted to say in his head was captured by implants, translated by decodors, then synthesized into sound and sent directly to the phone line.
There's a wife's laughter on the end of the listener.
They started to have fun and talk to each other, just like they did on the phone.
Engineers stand by and watch. Neuralink then said that this was one of the most memorable moments of the week.
In the field of engineering, there is little more than this. You look at a man and you use the code you wrote that morning to talk to the one he loves.
In the past, the brain bandwidth was too narrow, and people could only click on the machine "squeeze toothpaste" one letter at a time.
Now, the brain just needs to output an intent, and the rest of the dismantling, execution, operation will go to AI. When Grok enters the brain interface, it is expected that for the first time the brain will become the direct input end of AI.
It's going to be a big leap of human interaction.
In the future, with more electrodes, faster decoders, and Agent smarter, "getting ahead" may go into more people's daily lives from the lab.
In the Hacking Empire, Neo took a line and learned kung fu in a few seconds; in the SuperAxer, people's consciousness is connected to the network and can directly mobilize the digital world.
These images, which used to belong only to the screen, are being written into reality.
References: https://neuralink.com/updates/72-hour-voice-hackathon/
Edit: Peaches
I don't like it.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.