Musk Unveils Neuralink Visual Implant Blindsight
Musk has unveiled Neuralink's visual implant Blindsight, which bypasses the eyeballs and writes images directly into the brain's visual cortex; the first-generation product offers low-resolution black-and-white pixels. The author believes its endpoint is to give humans multispectral vision including infrared, ultraviolet, and even radar waves, but clinical approval progress, whether the brain can decode heterogeneous data streams, and the electrode scarring and rejection associated with long-term implantation all await data from the first batch of subjects.
Why read this
AI score 55
The article places Blindsight alongside traditional sensory restoration approaches, allowing readers to understand the difference in thinking between brain-computer interfaces moving from treating disease to expanding the senses.
Topics
View original postEnglish text is being prepared.
View original
卧槽兄弟们，这个真的牛逼，感觉大模型诞生至今最憋屈的七年失忆史可能要翻篇了！！ 甚至我想说个暴论，这大概是今年大模型架构演进里最性感、但也最击中行业要害的一次底层突围。 刚刚开源的 MemT-1.0 第一次在 Transformer 底层焊入了一颗真正的生物级海马体， 让 AI 从聊完即忘的失忆金鱼，真正进化成了能把经验沉淀进血液里的生命体。 靠暴力堆显卡、拿几个 G 的 KV Cache 硬撑长文本的旧石器时代，今天被这个开源原生记忆大模型正式砸穿了。 他们放出的 66 页技术报告和最高 288B 的开源权重直接甩出了一组降维打击的工程账单： 把记忆变成前向计算流的原生状态后，256K 超长历史下的显存开销从 5.06 个 G 暴降到恒定的 40 兆， 直接把 KV 缓存砍掉了整整一百二十倍以上，而且首字延迟更是不管历史拉得多长，始终稳稳压在一百毫秒出头。 最先击穿传统路径的，是它彻底重构了模型的前向计算流。 以前大家为了让 AI 记住上下文，只能硬着头皮堆长文本或者外挂向量库，代价就是显存撑爆且丢失多步推理因果，它在标准网络旁并联了一套原生记忆模块，在输入进入的瞬间，擦除门、写入门和遗忘门就像生物海马体一样自动运转，不再保留所有的死文字，而是把信息实时消化成动态的内部记忆矩阵。 更绝的是它在真实智能体场景下带来的降维打击，团队把历史纠错中沉淀的一千多条代码经验，直接写入内部的原生记忆状态，完全不占用任何系统提示词的 Token 额度， 结果在标准代码测试集上，连续三次均通过的硬核指标直接提升了百分之八，稳定性甚至反超了直接塞明文提示词的做法。 这意味着经验不再是一段需要模型每次重新阅读理解的生硬规则，而是直接内化成了它的本能直觉，真是牛逼! 我觉得这才是通往真正的递归自我改进最底层的物理逻辑啊, 参数只定义了模型的出厂下限，但唯有持久内化的原生记忆，才能为智能打开复利增长的空间。 以前我们买的模型是聊完即焚的一次性计算工具，从今天起模型正在变成越用越懂业务、越用越值钱的自进化资产。 大家在做智能体或者长对话应用时，最头疼的是长上下文带来的昂贵显存和卡顿延迟，还是模型转头就忘业务规范的失忆症？ 对我来说，光是把 256K 显存压缩一百二十倍且延迟恒定这一项，就已经能让无数原本算不过账的长程业务直接原地复活了hh~
Body · AI translation
Musk has just unveiled Neuralink's next ace: the visual implant Blindsight.
Even if a person is completely blind, it can bypass the eyeball and write images directly into the brain's visual cortex.
The first-generation product only offers low-resolution black-and-white pixels, but its endpoint is not at all about restoring blind people to ordinary 1.0 vision: it is about letting humans grow multispectral vision that surpasses physiological limits—seeing infrared, ultraviolet, and even radar waves directly with the naked eye.
Traditional medicine has always followed the logic of "repairing damaged organs": if the eye is broken, fix the eye; if a nerve is dead, stimulate the nerve—the ceiling is forever locked within the physiological limits of the human body.
But Musk has overturned this repair logic entirely: the human eye is nothing more than a "biological camera" that can only receive narrow-band visible light. As long as you bypass the eyeball and write signals directly into the occipital visual cortex, it doesn't matter to the body what sensor is connected at the front end.
It's like an old TV whose antenna has been struck by lightning: traditional medicine is climbing onto the roof desperately welding that broken antenna back together, at best picking up a few blurry local channels; whereas Blindsight simply rips off the antenna and plugs a digital fiber optic cable straight into the TV's motherboard—connect a night-vision scope on the other end, and the brain sees infrared thermal imaging; connect a millimeter-wave radar, and the brain sees through-wall mapping.
The purpose of medical devices has, for the first time, shifted from "repairing the patient back to normal" to "opening an external hardware interface directly to the brain." The earliest beneficiaries are blind people, but one step beyond that, and it becomes a "sensory add-on" that even able-bodied people would envy.
Of course, the flip side of the coin remains stark: Musk is best at using endgame science fiction to pave the way for present-day engineering. In the video he only proclaimed that it is coming soon, but clinical approval progress, whether the brain can decode heterogeneous data streams, and electrode scarring and rejection after long-term implantation all have to wait for data from the first batch of trial participants to speak.
What traditional medicine pursues is restoring the incomplete body back to factory settings, whereas what brain-computer interfaces scheme to do is to pry open the factory lock on human senses directly.