←News detail

AYi AI NotesUpdated Oct 24 23:31

Neuralink's Blindsight visual implant is about to arrive, bypassing the eyes to write directly to the visual cortex

AI summary

Musk has unveiled Neuralink's visual implant Blindsight, claiming it can skip the eyes and optic nerve and write images directly into the brain's visual cortex, and that even people who are completely blind can use it. The first-generation product initially provides only low-resolution basic imagery, with the long-term goal of letting people see infrared, ultraviolet, and even radar signals. The original article notes that the video did not explain the progress of human clinical approval, and that whether the brain can decode non-visible-light signals and whether long-term electrode implantation causes rejection will all have to wait for data from the first batch of subjects.

Why read this

AI score 60

The original article reviews Neuralink's shift in thinking on visual implants, from repairing the senses to connecting directly to sensors, and also points out unresolved issues such as clinical approval and cortical decoding.

Topics

Tags

Quoted post

AYi

@AYi_AInotes

English 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~

View original post

Body · AI translation

Elon is awesome! Musk has just unveiled the next ace from his brain-computer interface company Neuralink: a visual implant called Blindsight is about to arrive.

Even if a person is completely blind, this device can bypass the eyeballs and write images directly into the brain's visual cortex.

In the video, he laid it out: when the first-generation product is first implanted, it can only provide low-resolution basic images at first,

but the endpoint of this technology is not at all to help blind people recover to a normal person's 1.0 vision, but to let people grow high-definition multispectral vision that surpasses physiological limits, directly seeing infrared, ultraviolet, and even radar signals with the naked eye.

In the past, when the medical field developed artificial retinas or blind-assistance devices, the default approach was always to repair the damaged organ: if the eye is broken, fix the eye; if photoreceptor cells die, stimulate the remaining nerves. The ceiling was always locked within the physiological limits of a healthy human eye as it comes from the factory.

But Musk has directly overturned this repair logic. Since the human eye is just a biological camera that can only receive an extremely narrow band of visible light, once a brain-computer interface bypasses the eyeball and optic nerve and writes digital signals directly into the visual cortex in the brain's occipital lobe, what sensors are connected at the front end is completely free from the limits of the biological body.

This is like an old television whose antenna has been struck by lightning. Traditional medicine climbs onto the roof and tries every way to weld that broken antenna back together, at best receiving a few blurry local channels,

whereas Blindsight simply pulls out the antenna and plugs a high-definition digital fiber optic cable into the TV's motherboard. If you connect a night-vision device to the other end of the fiber, what you see in your brain is infrared thermal imaging; connect a millimeter-wave radar, and what emerges in your brain is a radar wave map that sees through walls. For everyone paying attention to brain-computer interfaces and hardware evolution, this is a huge watershed. When the purpose of medical devices shifts from repairing patients back to normal people to directly opening a hardware interface to external sensors for the brain, the earliest beneficiaries are indeed blind patients, but one step across, and it becomes a sensory add-on that even healthy people would envy.

But stripping away the glamour on stage, the other side of the coin is actually quite grim. Musk himself is the leader best at using ultimate sci-fi visions to pave the way for current engineering,

and in the video he only shouted that it is coming soon, without explaining how far human clinical approval has progressed,

moreover, the brain's visual cortex has been accustomed to visible-light signals for millions of years. Suddenly forcing streams of infrared and radar wave data into it—whether the human brain can decode meaningful images, and whether long-term electrode implantation will trigger cortical rejection and scar aging—all must wait for the real data from the first batch of low-resolution clinical subjects to speak.

What traditional medicine pursues is to restore the incomplete body to factory settings, while what brain-computer interfaces scheme to do is to directly pry open the factory lock on human senses. https://t.co/KHVP0unJhF

Source

AYi AI Notes