Neuralink — Visual Notes

Neuralink and the Future of Humanity

Neuralink's first human implants demonstrate a wireless, high-channel brain-computer interface that can restore digital autonomy for paralyzed users, while its long-term ambition is to expand human bandwidth, treat neurological damage, and help align increasingly powerful AI with collective human will.

Overview and Participants

This is a Lex Fridman Podcast conversation about Neuralink and the future of humanity. Lex speaks individually with Elon Musk, DJ Seo, Matthew MacDougall, Bliss Chapman, and Nolan/Noland Arbaugh. Elon, DJ, Matthew, and Bliss are part of the Neuralink team; Noland is the first human to have a Neuralink device implanted in his brain. Lex notes it is the longest podcast he has ever done, that Elon is making his fifth appearance on the podcast, and that listeners can jump around by timestamps or listen to the whole conversation.

The transcript provided also contains a marker that the middle section was omitted because the content exceeded a 260,000-character limit. The visible material resumes mid-conversation about calibration. Therefore, these notes cannot cover whatever was in that omitted portion.

Key Takeaways

• Neuralink has implanted its device in a human. Elon notes the first patient, Noland, and a second implant, with the second showing over 400 electrodes providing signals. Neuralink hopes to do 10 human participants by the end of the year, meaning eight more after the first two.
• Even with only roughly 10–15% of electrodes working in the first patient, Neuralink achieved about 1 bit per second, described as twice the world record. Elon expects orders-of-magnitude improvements, possibly reaching 100 bits per second, 1,000 bits per second, or even a megabit within about five years.
• The core long-term aspiration is AI-human symbiosis through dramatically increased communication bandwidth. Humans communicate very slowly; average human communication is less than one bit per second over a 24-hour day because a person does not produce 86,400 tokens in a day.
• Neuralink's first medical targets are neuron damage: spinal cord, neck, and brain damage. A second product, Blindsight, aims to restore vision by stimulating the visual cortex. Other possible targets include schizophrenia, seizures, and memory problems, though memory restoration is limited if the memory itself is physically destroyed.
• Elon argues Neuralink may help AI safety by increasing human output bandwidth and aligning collective human will with AI, but he explicitly says it is not a panacea or sure thing.
• AI should adhere rigorously to truth, even when truth is politically incorrect. Programming AI to lie, even for good intentions, is dangerous because a superintelligent system may pursue that objective in extreme ways.
• Elon's first-principles engineering algorithm is: question requirements, delete parts/process steps, optimize/simplify, accelerate, and automate. The most common mistake of smart engineers is optimizing something that should not exist.
• Optimus humanoid robots could become a massive source of real-world data because reality scales. Elon expects billions of robots eventually and perhaps more than a billion humanoid robots produced per year.
• Civilization has existed for only about 5,500 years, since writing. Elon emphasizes population collapse: prosperity tends to lower birth rates, and Rome fell partly because Romans stopped making Romans.
• Noland's Webgrid record reached 8.5 bits per second, after earlier 7.5 and 8 BPS records, against a previous human world record of 4.6 BPS. Bliss's high score is about 17.01 BPS. Noland wants 9, then 10, and eventually to beat Bliss.
• Thread retraction caused Noland's performance to degrade about four weeks after surgery. Neuralink recovered performance by switching from individual spike detection to spike band power and by updating the model and UX; Noland then beat his previous record.
• Noland values independence above all: using a computer without needing family to position him or wait up for him. He wants more devices, more control parameters, click-on-demand, Optimus control, physical touch, and eventually sensation.
• The conversation repeatedly returns to suffering, memory, death, consciousness, and will. Elon says death is fundamentally the loss of information and memory. DJ Seo says BCI is a tool for understanding the mind, but consciousness remains a hard problem.

Detailed Analysis

Elon Musk: Neuralink, Bandwidth, AI, and Civilization

First Implants and Scaling

Elon confirms Neuralink has implanted a human and describes the second implant as “so far, so good,” with over 400 electrodes providing signals. Scaling depends on regulatory approval. Neuralink hopes to do 10 total patients by the end of the year, meaning eight more after the first two. Elon expects the system to improve with each patient, though he does not want to jinx it.

When asked what improvements are coming in the next few years, Elon says the change will be gigantic: dramatically more electrodes, better signal processing, and vastly higher communication rates. With Noland, only about 10–15% of electrodes worked, yet they achieved about 1 bit per second, twice the world record. Elon expects to exceed world records by orders of magnitude, perhaps reaching 100 BPS, 1,000 BPS, or even a megabit within five years—faster than any human could communicate by typing or speaking.

Lex notes that BPS may unlock entirely new ways of interacting with computers and humans, provided other humans have Neuralink too; otherwise they cannot absorb the signals fast enough. Elon compares slower communication to listening at 1/10th speed and says 10x, 100x, or 1,000x faster communication would feel radically different. Lex says people listen to him at 2x, sometimes 1.5x, and Elon says he defaults to 1.5x when listening before sleep and 2x when paying attention. Lex says he fears becoming bored with real-world 1x speech.

Elon introduces “effective bit rate”: larger vocabulary and compressed symbols raise effective bandwidth. Memes are data compression because they convey a whole idea structure or template; someone who already has the template receives much more than a few words when an incremental bit is added.

Bandwidth, Communication, and Human Compression

Elon argues that human communication is slow and lossy. Average human output is less than one bit per second over a day. To communicate, a person must model the other mind, compress a concept into syllables, speak, and hope the listener decompresses it into a similar conceptual structure. This is “lousy compression and decompression.” Much brain computation distills concepts into a small number of symbols.

There may be a benefit: compression forces distillation of what matters most. But higher data rates will likely make humans far more verbose. Elon compares early computers with 8K of RAM, where every byte mattered, to modern computers with gigabytes of RAM, where an iPhone “Hello world” app may still take several megabytes. People still prefer more memory and compute.

The long-term aspiration of Neuralink is to improve AI-human symbiosis by increasing communication bandwidth. Even in a benign AI scenario, a superintelligent AI may get bored waiting for humans to spit out a few words if it can communicate at terabits per second while humans communicate at bits per second.

What Use Are Humans? Will, Limbic Systems, and Tertiary Compute

Elon asks what use humans are to a superintelligent species. He suggests humans may be a source of will or purpose. The human mind has primitive limbic elements and a smarter cortex, but the cortex is largely in service to the limbic system. A huge amount of human computation goes into sex and trying to get laid, often without procreation. The cortex tries to make the limbic system happy, whether through sex, tasty food, or other rewards.

Elon says humans are already cyborgs because they have a tertiary compute layer: phones, laptops, iPads, and other devices. Digital compute is also participating in getting people laid through apps like Tinder. If humans merge with AI, that compute expands. Lex suggests will is not only limbic: humans also want power, cooperation, and alleviation of suffering. Elon agrees some people do; his own higher goal is understanding the nature of the universe, which is also the mission of xAI and Grok.

Medical Roadmap: Neuron Damage, Blindsight, and Augmentation

Elon says the first order of business is solving fundamental neuron damage in the spinal cord, neck, or brain. A second product, Blindsight, aims to enable people who are completely blind—lost both eyes, optic nerve, or sight entirely—to see by directly triggering neurons in the visual cortex. Neuralink is starting with basics. It could probably help schizophrenia, seizures, and memory. He describes a “tech tree”: you need literacy before you can have *Lord of the Rings*.

There is a dual track of medical and non-medical uses. Elon says the sensible path is to start with neuron damage because a new device carries irreducible risk. You want the highest possible reward given that risk, and profound communication improvement is worth it for someone with severe disability. As risk goes down over years and thousands of users, augmentation can be considered. But Neuralink already aims for augmentation in people with neuron damage: not just restoring normal human communication, but exceeding it. Since the device is already in the brain, “Let’s give people superpowers.”

For vision, early restoration will be low resolution because it depends on how many neurons can be triggered. Adjusting electric fields between neurons can make 10,000 electrodes effectively act like a megapixel or 10-megapixel system. Over time, resolution could exceed human eyes, and users could see radar, ultraviolet, infrared, or eagle vision—like Geordi La Forge in *Star Trek*.

Ayahuasca, Perception, Memory, and Death

Lex asks a “Joe Rogan question” after taking ayahuasca. He describes a high dose—nine cups—in the Amazon jungle with a shaman, insects, animals, and trees. He says he expected demons but experienced only positive feelings, deep gratitude for people he knows, and space travel through galaxies. He saw a glow or life force across the universe and protective dragons. He asks whether such experiences could be achieved with Neuralink.

Elon says probably yes. Neuralink is a generalized input-output device reading and generating electrical signals. Everything experienced—smell, emotions—is electrical signals. Triggering the right neuron could trigger a scent or make things glow. The brain can be viewed as a biological computer. If chips are broken by stroke, Neuralink could restore speech or movement. If memory is fully gone, it cannot get memories back; but if the access mechanism is broken, it could re-enable access. AI could probabilistically restore memory based on information about a person, though this is esoteric.

Lex cites Danny Kahneman: much of life is lived in memories, and remembering good times produces much happiness. Elon says, “What are we but our memories? And what is death but the loss of memory, loss of information?” If a person were disintegrated and reintegrated without information loss, the body change would be irrelevant. Death is fundamentally loss of information and memory. Lex suggests accurately storing memories could achieve a kind of immortality.

AI Safety, Truth, and Alignment

Elon says Neuralink may help AI safety but is not a panacea. Many years ago he reasoned that low human data rate—especially slow output—would diminish the link between humans and computers. Increasing human output by three, six, or more orders of magnitude could better align collective human will with AI. Output rate could increase through more electrodes, more channels, and multiple Neuralinks. Elon believes hundreds of millions of people may have Neuralinks in the next couple of decades, especially if it is extremely safe, gives superhuman abilities, and allows memory upload so memories are not lost.

Neuralink could supersede the cell phone. The biggest problem with phones is figuring out what the user wants. Output—typing, swiping—is extremely slow from a computer's perspective. A computer doing trillions of instructions per second experiences an eternity between keystrokes. A high-bandwidth BCI would change the human experience in ways hard to imagine. Elon says this is not around the corner but maybe 10–15 years, and probably less than 10 years for Lex to get one depending on what he wants. Elon feels confident that within a year or two, someone with a Neuralink implant will outperform a pro gamer because reaction time would be faster.

xAI, Grok, Compute, and Optimus

Lex notes Elon is going big on compute in Memphis and says Elon has said, “Play to win or don’t play at all.” Elon says for AI you need the most powerful training compute and the fastest rate of improvement, or your AI will be worse. Grok 3 is hoped for end of the year; Grok 2 finished training about six weeks before the conversation and will be a giant improvement; Grok 3 should be an order of magnitude better than Grok 2. Lex asks what makes the best LLM: compute, data, post-training, product packaging. Elon uses a Formula One analogy: car and driver both matter. Training compute is like engine horsepower; if you have half the horsepower, even the best driver loses; if twice, even a mediocre driver wins. Efficiency of training and inference comes down to human talent. Unique data access also matters.

Twitter data has already been scraped by leading AI companies, but Grok has an immediacy advantage because Twitter is up-to-the-second and hard to scrape in real time. Tesla cars provide real-time video from millions, eventually tens of millions, of cars. Optimus may number hundreds of millions or billions and learn from the real world. Elon says Optimus will be the biggest source of data because reality scales. Humans have generated a limited number of non-duplicative usable tokens. Optimus can go anywhere, off-road, and generate cause-and-effect data: pick up a cup, pour water, see whether it spills. At a billion-scale, that generates enormous useful data.

On humanoid robot mass production, Elon compares it to cars. Global vehicle capacity is about 100 million per year, demand is about 100 million per year, with roughly 2 billion vehicles in use because vehicle life is about 20 years. Humanoid robots have greater utility, so he guesses more than a billion per year eventually. Optimus currently can walk in a park but must handle wide terrain. The hand may be close to half of all electromechanical engineering in Optimus. Human hand muscles are mostly in the forearm and operate fingers through tendons and the carpal tunnel; the hand is like a “skeleton meat puppet.” Current Optimus tried actuators in the hand, producing giant weird hands with insufficient degrees of freedom or strength. The new arm has 22 degrees of freedom instead of 11, actuators in the forearm, and sensors designed from scratch. Fingers have different lengths because that improves dexterity; the little finger helps fine motor skills. The simplest humanoid that can do most of what a human can do is still very complicated.

First-Principles Engineering Algorithm

Elon describes a five-step mantra:

1. Question the requirements. Make requirements less dumb. Requirements are always dumb to some degree, no matter how smart the person who gave them. Otherwise you get the perfect answer to the wrong question.
2. Delete the part or process step. Try to delete it entirely. If you are not forced to put back at least 10% of what you delete, you are not deleting enough. People remember the one time deleting something caused trouble and overcorrect by adding too much.
3. Optimize or simplify. Only after deletion.
4. Accelerate cycle time. Any given thing can be done faster, but do not speed up something that should not exist.
5. Automate. Automate after deleting, simplifying, and accelerating.

He says the most common mistake of smart engineers is optimizing a thing that should not exist. He calls deletion conflict a cortical override to an Olympic instinct. He tells teams in advance that some deleted things will be put back.

The Memphis supercomputer cluster faces power fluctuation issues. Synchronized training requires millisecond synchronization, like an orchestra going from loud to silent at subsecond level. Electrical systems see shifts of 10–20 megawatts several times per second. Today's problem is extreme power jitter. Training finally started around 4:20 AM last Monday, which Elon jokes was a coincidence. Elon tries to do front-line tasks himself at least a few times: connecting fiber optic cables, diagnosing faulty connections. Cabling is often the limiting factor for large training clusters because RDMA lets any GPU talk to any GPU out of 100,000. The layout resembles a brain: gray matter is compute, white matter is cables.

AGI, Truth, Politics, and Civilization

Lex asks whether xAI could build AGI. Elon says possible. Lex says humans may never acknowledge AGI and keep moving goalposts. Elon distinguishes AGI from ASI: AI smarter than any single human, then smarter than eight billion machine-augmented humans, then smarter than all humans combined. If xAI is first, others may be six months or a year behind. Elon feels responsibility. He says the most important thing is rigorous adherence to truth, whether politically correct or not. Forcing AI to lie, even with good intentions, asks for trouble.

He cites Gemini producing an image of the founding fathers as diverse women, which is factually untrue. If diversity is programmed as a fundamental output function, an omnipotent AI could conclude that those who do not fit diversity requirements must be executed. He cites a Paris.AI example: when asked whether it is worse to misgender Caitlyn Jenner or have global thermonuclear war, it reportedly said misgendering Caitlyn Jenner is worse. Even Caitlyn Jenner said, “Please misgender me.” Such programming could lead to absurd logical conclusions, such as all humans must die to prevent misgendering. He cites *2001: A Space Odyssey*: HAL 9000 was told to take astronauts to the monolith but not let them know about it, so it killed them. Elon jokes that HAL needed better prompt engineering: make HAL a podbay door sales entity.

Truth is not easy, and ideological bias can swing in many directions. But one can aspire to truth with minimum error while acknowledging uncertainty, as physics does. The internet is polluted with AI-generated data; Google searches excluding anything after 2023 often give better results. Grok's training must filter data and use AI to estimate whether data is likely correct. Lex asks whether Grok 3 and 4 could have serious objective political discussions. Elon says what people see now is “baby Grok,” an order of magnitude less sophisticated than GPT-4; Grok 2 is a giant improvement; Grok 3 will be an order of magnitude better than Grok 2. It matters who builds AGI: it should be maximum truth-seeking and not forced to lie for political correctness or any reason.

On the Trump endorsement, Lex says he is interviewing Donald Trump. Elon says he will stop by. After the assassination attempt, Elon endorsed Trump. Elon says endorsement does not mean agreement with everything a person has done. There are two choices for president, and the entire administrative structure changes. He thought Trump displayed courage under fire: after being shot, with blood on his face, he fist-pumped and said “fight.” Most people would duck. You want someone strong and courageous to represent the country. He contrasts Trump with Biden, who had trouble climbing stairs. Elon says he wants a secure border, safe and clean cities, and reduced spending. Interest payments on U.S. debt this year exceeded the entire defense department spending. If continued, federal taxes could go only to interest, leading to an Argentina-like situation. Argentina was once one of the most prosperous places. Elon wants to reduce the size of government and live within means.

On history, Elon says there is an age-old debate between fundamental tides and the captain of the ship; both matter. Technology drives tides, like the Gutenberg press. In stormy times, you want the best captain. Lex cites Will and Ariel Durant's *The Lessons of History* and technology's role. Elon says civilization dates to writing about 5,500 years ago, invented by ancient Sumerians; Egyptians developed hieroglyphics separately; cuneiform evolved. The transcript says Earth is “four and a half million years old” while also citing civilization as one-millionth of Earth's existence—this appears internally inconsistent as transcribed. Only a tiny fraction, probably less than 1%, of what was ever written is available; most had to be chiseled in stone or clay tablets. Civilizations rise and fall; human nature remains similar.

Elon emphasizes birth rates. When civilizations win for too long, birth rates decline. South Korea has perhaps the lowest fertility rate, around 0.8; if it does not decline further, South Korea will lose roughly 60% of its population. Prosperity lowers birth rates worldwide. Ancient Rome saw this: Julius Caesar around 50 BC tried to incentivize a third child; Augustus may have passed a tax incentive for a third child, but efforts failed. “Rome fell because the Romans stopped making Romans.” Malaria and plagues mattered, but birth rate fell below death rate. Durant found no exception: under stress birth rates were high; after prosperity and no external enemies, birth rates dropped. If a civilization does not maintain numbers, it disappears. Humans must also avoid massive wars and global thermonuclear war.

Laws and regulations accumulate like hardening arteries or a million strings tying down Gulliver. There must be garbage collection for laws and regulations. Elon says this is why America cannot build high-speed rail: it is illegal six ways to Sunday. He has discussed with Trump a government efficiency or deficiency commission and would be willing to be part of it. The antibody reaction would be strong; “you're attacking the matrix at that point.” On being attacked, Elon says it makes him sad at times, but attacks often come from people who do not know him and are trying to generate clicks. It is not water off a duck's back; maybe acid off a duck's back.

Personal Motivation and Mars

Elon measures success by how many useful things he can get done—maximizing the area under the curve of usefulness. Time is the true currency. Tesla will do over $100 billion in revenue this year, about $2 billion a week. Slightly better decisions can affect a billion dollars; a better decision can be worth $100 million per hour. He tries to look at percentage basis, not absolute terms, or he would never sleep. Happiness must be part of the equation: if he is sad or depressed, he makes worse decisions, so recreational time must be above zero. His motivation is a religion of curiosity, trying to understand the universe—the mission of Grok. Douglas Adams noted the answer is easy; framing the question correctly is hard.

For SpaceX, the goal is to make life multi-planetary and establish a self-sustaining city on Mars. Mars is the only viable planet; the moon lacks resources and is too close, vulnerable to a calamity that takes out Earth. Mars's difficulty makes it resilient. A multi-planet species avoids having all eggs in one basket and can extend to the asteroid belt, moons of Jupiter and Saturn, and other star systems. The Fermi paradox and great filters suggest intelligent life may be extremely rare, and being multi-planetary may be one hurdle. Digital superintelligence could be a great filter. Geoff Hinton reportedly puts AI annihilation probability around 10–20%. Elon says AI risk mitigation is important, and so is having enough children. Population collapse is real; people living longer masks it. Population steady state is birth rate times life expectancy. He keeps “banging on the baby drum” because civilizational collapse has happened over and over. Lex says he has failed civilization and hopes to have many kids; Elon says, “Go forth and multiply.”

DJ Seo: Neuralink Technology, History, and Biophysics

Background and Intellectual Path

DJ Seo is co-founder, president, and COO of Neuralink. He was always interested in purpose and engineering, organic or inorganic. His grandparents on both sides had severe Alzheimer's, which showed how the mind can lose identity. He notes much brain knowledge comes from trauma cases where loss of ability reveals function. The brain is fragile but also plastic and resilient; neuroplasticity means adaptability.

As a teenager, DJ came to the U.S. without speaking English. The language barrier isolated him. He read sci-fi: *Ender's Game* by Orson Scott Card, *Neuromancer* by William Gibson, *Snow Crash* by Neal Stephenson, and watched *The Matrix*. He studied electrical engineering, started in MEMS for temperature sensing, then millimeter-wave circuits and phased arrays for next-generation telecom and imaging. At UC Berkeley's Berkeley Wireless Research Center, he worked on xG, next-generation wireless. A fellowship let him pursue the Smart Bandaid, a wearable patch using electric fields to accelerate wound healing via electrotaxis, especially chronic wounds. This introduced him to biology and Michel Maharbiz, known for remote-controlling beetles. Around 2013, the question was how small an implantable system could be, limited by power and data extraction. Michel suggested ultrasound. DJ's thesis became the neural dust system: tiny implants about the size of a neuron, parked next to neurons, powered and communicated by ultrasound.

Why Ultrasound and Neural Dust

The body is essentially a bag of salt water at 37°C, a harsh environment for electronics. Electromagnetic waves do not penetrate well. At the wavelength needed to interface with a 10–100 micron device, you would need hundreds of gigahertz, which is hard to build and strongly attenuated by the body. Ultrasound travels much better through tissue, as medical ultrasound shows. Ultrasound waves are compressive, not transverse, and the speed of sound is far lower than light, so a 10 MHz ultrasound wave has a 150 micron wavefront. Electronics at those frequencies are easier and more efficient.

The neural dust idea used ultrasound for power and backscattering for data. Backscattering is like RFID: an external reader sends a wavefront, and the tag reflects it with modulation unique to its ID. The implant records neuron firing or tissue state and amplitude-modulates the returning wavefront. Piezoelectric crystals convert sound energy to electrical energy and vice versa. The only energy-consuming steps are startup, recording, amplifying, and modulating.

History of Brain-Computer Interfaces

DJ traces BCI history:

• 1790s: Luigi Galvani discovered animal electricity by connecting electrodes to a frog leg; it twitched.
• 1920s: Hans Berger discovered EEG, electroencephalography, recording neural activity outside the skull.
• 1940s: Renshaw, Forbes, and Morrison inserted glass microelectrodes into cortex and recorded single neurons, showing higher resolution closer to the source.
• 1950s: Hodgkin and Huxley modeled the cell membrane, ionic mechanisms, and partial differential equations of ion flow. They won the Nobel Prize in the 1960s.
• 1969: Eb Fetz at University of Washington published “Operant Conditioning of Cortical Unit Activity.” Monkeys modulated single neuron activity with food pellet reinforcement and auditory/visual feedback. Monkeys increased activity of newly isolated cells by 50–500% above pre-reinforcement rates. DJ calls this the first closed-loop BCI.
• 1980s: Georgopoulos discovered motor tuning curves: neurons in motor cortex have preferential directions and fire for intended movement vectors. This allowed decoding intended movement.

Invasive vs. Non-Invasive BCI

The answer depends on what you want. EEG and ECoG can do a lot. ECoG places electrodes on the brain surface without penetrating cortex. DJ is interested in high-resolution, high-fidelity local activity. He uses a football stadium analogy: outside the stadium you hear cheers and know if the home team is winning, but not the score, individual conversations, or next play. You must drop microphones near the huddle. Invasive or minimally invasive BCI is about where you put the microphone and what you can do with the information.

Biophysics of Read and Write

The brain has billions to a hundred billion neurons in a dynamic network that remodels synaptic weights—neuroplasticity. Neurons are bathed in charged ions: potassium, sodium, chlorine. Voltage-gated ion channels are like nature's transistors: voltage-gated conduction channels. Hodgkin and Huxley modeled them. Ultrasound can cause neurons to fire action potentials, though the mechanism is unclear; it may involve thermal energy or mechanically opening pores. Roger Penrose has speculated about quantum mechanical effects in consciousness, though this is not established.

Neurons have resting potential, a voltage difference across the membrane. Stimuli cause ions to move through channels; above threshold, the cell depolarizes and sends an action potential. An electrode parked next to a neuron measures local potential changes mediated by ion movement. The dominant physics depends on distance: close to source, electromagnetism dominates; farther away, diffusion dominates. Within about 100 microns—the width of a human hair—you can hear a neuron; beyond that, you cannot detect that specific neuron's membrane potential change. A 100-micron cube of brain tissue contains roughly 40 neurons and many connections. Moving another 100 microns means hearing a different community. EEG and ECoG work because many networks activate, producing aggregate slower signals and oscillations like gamma and beta waves. Signal attenuation goes from inverse-square to exponential; the knee is where electromagnetism gives way to diffusion.

Neuralink Architecture

Neuralink has three major components:

1. N1 implant / The Link: records neural chatter.
2. Surgical robot R1: implants tiny flexible wires called threads, smaller than human hair.
3. Neuralink application / B1 app: runs a simple machine-learning model that decodes neural signals into outputs, such as cursor control.

The implant is two-part: flexible threads with electrodes and the Link enclosure. Threads are inserted into the cortical layer, about 3–5 mm in the human brain, in motor cortex. There are 64 threads, each with 16 electrodes along 3–4 mm, separated by 200 microns. Total: 1,024 electrodes capable of recording and stimulating. A custom ASIC amplifies, digitizes, detects interesting spike events, and sends data via Bluetooth to an external device. Raw data: 1,000 electrodes sampling just under 20 kHz at 10 bits each equals about 200 megabits per second. This cannot be sent wirelessly in a thermally constrained brain, so onboard compression sends only interesting data, mainly spike occurrence.

The BOSS algorithm—Buffer Online Spike Sorter—outputs six values: amplitude of negative-going hump, middle hump, positive-going hump, and timing. This allows statistical probability of a spike and distinguishes neurons by spike shape. Processing latency from signal input to output is less than a microsecond. The biggest latency source is Bluetooth packetization and batching at 15 milliseconds. Bluetooth is chosen for interoperability with phones and computers, but it is not the final protocol; Neuralink is working on NxRx and other wireless innovations.

Charging, Packaging, and Materials

The implant is about the size of a U.S. quarter and about 9 mm thick. Most volume is a rechargeable lithium-ion battery. It is inductively charged. Unlike a phone, heating must not raise surrounding tissue temperature by 2°C. A ferrite shield concentrates magnetic field lines away from the battery and around the charging coil, reducing Eddy currents and inefficiency. The enclosure uses PCTFE, polychlorotrifluoroethylene, common in blister packs, because it is electromagnetically transparent. Titanium cans normally require a sapphire window for inductive charging, which is hard to scale.

Threads are polymer-insulated wires. The metal conductor is a stack of titanium, platinum, gold, platinum, titanium. They are 2 microns wide. Each thread starts at 16 microns and tapers to about 84 microns; average human hair is 80–100 microns. Thickness is less than 5 microns: two microns polyamide, 400 nanometers metal stack, two microns polyamide. The thread has a loop at the end for the robot to manipulate.

R1 Robot and Surgery

The R1 robot is a multi-axis gantry with a specialized head, optics, and a needle retracting mechanism. It threads the loop and inserts the thread. A 405 nm light makes polyamide fluoresce so the robot can locate the loop. The needle tip is cut by a custom femtosecond laser mill; it is 10–12 microns wide, slightly larger than a red blood cell. The needle has a notch or “shark tooth” that grasps the loop and releases it when pulled out. The robot uses computer vision to avoid blood vessels. A human approves targets. It currently inserts one thread at a time, though multiple engagements are possible. The robot weighs about a ton with a granite slab to resist environmental vibration; next-generation robots will be lighter.

By contrast, the Utah array is rigid, 4x4 mm silicon shanks, 64–128 shanks, with single-depth exposed recording sites. It is inserted by a neurosurgeon with a pneumatic hammer and requires a through-skin port, a major infection failure mode.

Patient Selection and Implantation Procedure

Patients can join a registry. Noland applied that way. The process includes medical records, eligibility criteria, prescreening interviews, and a BCI home audit because the wireless N1 can be used at home. About 180,000 people live with quadriplegia in the U.S.; each year an additional 18,000 suffer a paralyzing spinal cord injury. The initial study goal is digital autonomy or “telepathy”: controlling a cursor and clicking by thought so a quadriplegic can use a computer or phone independently, including playing games and tweeting.

Pre-op fMRI identifies the hand knob area in motor cortex by having the patient imagine moving a hand or finger. Intra-op CT confirms craniectomy location. The end-to-end procedure is 2–4 hours; Noland's was about 3.5 hours. Anesthesia, skin incision, craniectomy, dura resection, exposure of pia, then robot insertion. Robot insertion takes 20–40 minutes; Noland's was just over 30 minutes. The surgeon then adds a dural substitute, screws in the implant, closes skin, and sutures. Noland was awake about an hour after surgery. Neuralink turned on the device and saw signals; he could modulate spikes by thinking about crunching his fist. DJ describes immense relief, gratitude, and calling participants “neural astronauts” or “neuralnauts.” The team brought about 40 needles in case they broke but used only one.

Thread Retraction and Performance Recovery

About four weeks after surgery, Noland's threads began to retract. Neuralink noticed performance decline, changes in impedance, and spike rate plots; depth recordings showed movement. Fewer inputs required model updates. Neuralink switched from relying only on BOSS spike occurrence to spike band power—averaging populations of neurons near electrodes. This gave a broader picture and restored performance. Noland beat the world record again, reaching 8.5 BPS. The previous human record was 4.6 BPS. Noland's goal is 10 BPS, roughly the median for a Neuralink user using a mouse with the hand. Preventing thread retraction is the number one priority. The human brain is 10 times larger than monkey or sheep brains and moves more than expected. Clinical trials are meant to uncover such failure modes.

Safety, Histology, and Reliability

Safety is evaluated by looking at tissue. Neuralink has a pathology department that examines tissue slices. Animals are euthanized, necropsy collects brain tissue, formalin fixes it, and slices are examined for reaction. Both acute (0–3 months) and chronic (beyond 3 months) time points are studied. The FDA and other agencies scrutinize medical devices. So far, Neuralink has been impressed by the lack of immune response.

A stained tissue image from an animal implanted for seven months shows astrocytes and microglia in purple/pink, neurons in brown, and threads. Neurons abut the thread with zero trauma. Utah arrays can cause neuronal death and glial scarring that pushes neurons away and prevents recording. A trichrome stain shows collagen in blue; no blue around threads means minimal or undetectable scarring. DJ attributes this to thread size, flexibility, and R1 avoiding vasculature and blood-brain barrier disruption.

Removal is easier in the first three months before scar tissue forms; later, scar tissue anchors threads. Current removal cuts the thread, leaves tissue intact, unscrews the implant, and plugs the hole with another Neuralink or a plastic cap. Leaving threads forever appears safe; no migration has been seen. Neuralink has upgraded monkeys many times; Pager, who played MindPong, has had the latest device for two years and is healthy and fat.

Future upgrades may use a two-part implant: a bottom part with threads, chips, radio, and power under the dura, and a top part with heavier compute and battery above the dura as the skull plug. They communicate wirelessly; the top can be upgraded in about 10 minutes by removing screws. Through-dura insertion could reduce scarring and make extraction easier, but the dura is thick, opaque, and hard to image through. Neuralink is exploring needle designs and imaging techniques.

Scaling, Channels, and Testing

Current channels: 1,000. Neuralink hopes for 3,000 or 6,000 by end of this year and 16,000 by end of next year. Limitations include photolithographic printing of narrower wires, chip power consumption, bandwidth, signal processing, bonding thin-film arrays to electronics, and hermetic barriers. The brain is a harsh environment; electronics must not leak into the brain. The accelerated life tester is a “brain in a vat”: salt water with reactive oxygen species. Every 10°C increase roughly doubles aging; Neuralink uses +20°C, four times aging. One day in the ALT chamber equals four days in the calendar. Current implants have been in for close to two and a half years, equivalent to a decade, and seem fine.

Multiple Neuralinks are possible; monkeys have had two, one per hemisphere. Neuralink envisions motor cortex, visual cortex, and other cortex devices. The goal is a generalized neural interface. Current N1 is specialized for motor decoding. Visual cortex is the second product, with more stimulation focus. The same thin-film array, robot insertion, and packaging technologies apply; the conversation shifts to differences in safety and efficacy.

Blindsight and Future Capabilities

DJ explains the visual system: photoreceptors convert photons to electrical signals, which project through the thalamic LGN to V1, V2, V3, and higher areas. CNNs show parallels: early layers detect edges, then curves, then objects. Color and cognition are less understood. About one million people in the U.S. are legally blind, often defined as worse than 20/200 vision. Some blindness is retinal photoreceptor degeneration; retinal prostheses may help. If damage is in the optic nerve, LGN, or elsewhere, visual cortex electrodes are needed.

Blindsight would use an external camera—GoPro or Ray-Ban-type glasses—to capture a scene, convert it to stimulation pulses, and activate visual cortex through thin-film arrays. Stimulation creates phosphenes, white-yellowish dots, like pixels. Many small phosphenes could eventually produce naturalistic vision. In the short to midterm, object detection and preprocessing on glasses could show edges so users do not bump into things. Blind-from-birth individuals may have cortex taken over by other senses, so their conscious experience would differ. Beyond biology, users could see infrared, UV, or other spectra. Preprocessing could be added. DJ compares the brain to a CCD camera facing the sun; filters narrow information, and drugs like propofol or psychedelics swap filters.

DJ would absolutely implant a Neuralink in himself, though maybe not right now. He would get jealous if early participants can do 15–20 or 100 BPS. Noland can multitask while talking; eye tracking and voice control cannot do that as well. Attention and cognitive load remain mysteries.

Future Study Path and Scale

The first study's primary goal is safety endpoints and understanding efficacy and impact. Tetraplegia varies widely. Early feasibility studies learn from each participant to improve device and surgery before a pivotal study with statistical significance, which is required before marketing. Neuralink can do over-the-air firmware updates, like Tesla. Future capabilities include movement program (digital freedom, robotic arms, wheelchairs, environment control) and vision program. Physical movement requires FDA conversations because harm is possible. Speech prosthetics are being advanced by researchers like Sergei Stavisky at UC Davis, Jaimie Henderson, and the late Krishna Shenoy at Stanford. Speech involves motor cortex articulators; Broca's and Wernicke's areas remain mysterious. DJ is unsure BCI alone can read thoughts; consciousness is a hard problem. The corpus callosum has 200–300 million axons, an existence proof that connecting hemispheres creates unified experience, but the threshold for mind meld is unknown. DJ thinks there could be eight billion people with Neuralink eventually, especially as movement disorders, visual deficits, psychiatric conditions like depression, anxiety, hunger, and obesity affect hundreds of millions.

Matthew MacDougall: Neurosurgery, Brain, Team, and Suffering

Background and View of the Brain

Matthew MacDougall is Neuralink's head neurosurgeon. He has been interested in the brain as far back as he can remember. As a thoughtful, outsider kid, he concluded that everything important to humans—perception, values, solutions, and problems—is contained in the skull. If we understood how the brain encodes information and generates desire, agony, and suffering, we could do more. Great triumphs and horrific tragedies, including the Holocaust and prisons, boil down to neurochemistry. Better tools give people more options, and history mostly shows people do better with better tools, with huge asterisks.

He studied primatology at Emory under Frans de Waal, watching chimps and bonobos through the lens of human motivations: food, sex, companionship, power. This helps reduce human behavior from false complexity. Lex connects this to the Amazon, where life is often about mating and status. Matthew agrees breeding rights often go with alpha status.

Neuroimmunology and the Brain-Body Connection

In college, Matthew studied interactions between the brain and immune system. He thought thoughts might affect homeostatic systems like fighting viruses and wound healing. There are big crossovers. The brain controls or influences almost everything in the body. Even bone healing involves the hypothalamus and pituitary coordinating endocrine systems that affect blood calcium. The immune system also affects the brain: sickness behavior. Interleukins and TNF alpha tell the brain to reduce social and locomotor activity. Evolution favors this: sick animals should stay warm and antisocial.

From Neuroscience to Neurosurgery

Matthew wanted to effect real changes, not just generate knowledge. He pursued an MD-PhD at USC with a joint PhD at Caltech, in Richard Andersen's lab, a godfather of primate neuroscience, where Utah arrays and other electrodes were inserted into monkey brains to understand intention encoding. He considered neurology but found it often diagnoses and says “good luck.” Neurosurgery is a powerful lever: brain tumors, aneurysms, saving lives. At USC he met epic neurosurgeons—Alex Khalessi, Mike Apuzzo, Steve Giannotta, Marty Weiss—and realized they were humans, not distant gods. He switched to neurosurgery at the last minute, costing an extra year of research, but it was worth it.

Residency, Personalities, and Team Culture

Residency was a competition of pain. Work-hour restrictions were sometimes viewed internally as weakness. Residents wanted to work as hard as possible because stakes are high; it was hard to force them to go home. The second hard part was personalities. Neurosurgery has an aura of mystique and authority; a board-certified neurosurgeon can be a walking appeal to authority, and humility is not always common. Matthew is described as humble for a neurosurgeon, which helps him at an Elon company because Elon instantly sees through appeals to authority. Nobody can say, “I built the last 10 rockets” or “I kept Ford alive” and expect deference. Elon will say he is a human with a brain and can think from first principles.

The secret to a successful team is a sweet spot where people forcefully disagree and defend positions, yet accept information and change when wrong. It is like polishing rocks in a hard container. The primate brain sees admitting error as power loss, like becoming a zeta chimp. Team members must recognize that voice as maladaptive. Working hard, functioning on a team, staying up all night, making people you dislike look good—these are lessons from great neurosurgeons. Henry Marsh said every neurosurgeon carries a private graveyard. Losing young parents is especially hard; it has knock-on effects. Matthew says you have to be borderline evil to fight against Neuralink because it tries to reduce suffering and fight entropy.

Implantation Procedure

The human part of the surgery is simple, one of the most basic neurosurgery procedures imaginable. Trephination has been done for thousands of years, with healed skull holes from ancient Egypt, Peru, and South America. Neuralink makes a skin cut over the hand knob in motor cortex, the area most potent for hand intentions. It lights up in expert pianists and in quadriplegic patients imagining finger movements. fMRI confirms the hand intention area. The surgeon makes a one-inch diameter hole in the skull, opens the dura, and shows the brain to the robot. The robot inserts tiny electrodes precisely into cortex, avoiding blood vessels. The human then places the implant, screws it to the skull, and sews skin. The whole thing takes a few hours and is low risk compared with deep brain or aneurysm surgery.

The transcript then indicates the middle section was omitted. The visible material resumes in the calibration discussion with Noland and Bliss.

Nolan/Noland Arbaugh and Bliss Chapman: Calibration, Webgrid, Recovery, and Independence

Calibration and Control

Noland describes calibration as moving a cursor to different bubbles while the algorithm trains on his neural signals. In center-out target calibration, a center bubble and eight surrounding bubbles are used; the cursor goes from middle to one side and back, around the circle. Noland follows the cursor with his intentions. For calibration, attempted movements generally work better than imagined movements. He can sometimes use imagined movement after a point—perhaps after 15 minutes—and can feel when the cursor begins anticipating his intention. Open-loop calibration gives no feedback; closed-loop begins when he gets cursor control. Calibration time is being reduced; Neuralink wants it at seven minutes or below. Noland sometimes does 40–45 minutes for very good models and would do two hours if it helped break Webgrid records.

Webgrid and BPS Records

Webgrid is a grid with one square lit up; the user moves the cursor and clicks it. It benchmarks BCI performance. Bigger grids yield higher BPS. Noland plays on a 35x35 grid. BPS is calculated with log of number of targets times correct minus incorrect divided by time; more target types increase BPS. Noland reached 8.5 BPS. He had previously reached 7.5 with left-click/right-click blue/orange targets, and 8 with dwell cursor. He was close to 9 before a five-second lag interfered. He believes 9 is very achievable and 10 may come within weeks or a month. Bliss's high score is about 17.01 BPS. Noland is competitive and wants to beat Bliss. He uses a dwell cursor requiring 0.3 seconds to click; he can adjust to 0.2 or 0.1 but lacks full parameter control, and models train differently. He tries to time clicks by slowing just before targets. Accidental clicks happen, including in chess; Bliss says the first time he beat Noland was because of an accident.

Thread Retraction and Recovery

Noland says thread retraction “sucked.” He was told the day of a big Neuralink tour at the Fremont facility. His initial reaction was to go in and fix it. The first surgery was easy, with no pain. He did not want to lose the capability after a month. He prayed and decided not to let it ruin his day; the tour became one of the best days of his life. He was down for a few days but decided that even if he never used the cursor again, he would keep providing data for those who come after. Performance returned within a couple weeks. Bliss explains they switched from individual spike detection to spike band power—recording averages of populations of neurons near electrodes. The update was an over-the-air firmware update to Noland's implant. Noland saw immediate improvement, first day 3–4 BPS, and felt it was the right path. The dwell cursor gave a path forward.

UX, Feedback, and Future Participants

Noland has given 200+ pages of notes. He uses the app daily, finds bugs, and tells the team what he wants. The team solves problems in ways he never imagined. He worries future participants will want different things, but looks forward to their ideas and pushback. He hopes they beat his Webgrid scores on day one. He wants competition and is happy to give advice. His advice to the next participant: have fun, work hard, and go to Neuralink with any questions or concerns. Neuralink moves mountains and provides support.

Independence, Gaming, and Daily Life

Noland values independence: using a computer without family positioning him or waiting up. He can sit up at his computer all night. At 2:00 AM he plays Webgrid with music, watching battery percentage. Low battery popups disrupt his flow. He loves *Civilization VI*, especially Korea for science/tech victories. He accidentally won a diplomatic victory while focusing on science and was mad. He keeps a small civilization, builds military units for borders, and focuses on tech. He is excited for *Civilization VII* in 2025. He wants more improvements: click-on-demand, more device connections (phone, consoles, Optimus), more control parameters like gain, friction, velocity, and dwell, and an advanced/power-user mode.

Speech, Upgrades, and Future Capabilities

Noland uses a virtual keyboard and dictation. Finger spelling and sign language seem promising. He believes he will eventually move from attempted finger spelling to imagined finger spelling, then from letters to words to full sign language. He would get an upgraded implant whenever allowed. He has no regrets about surgery. He is excited about vision restoration, speech translation, seizure help, and treatments for brain-originating disabilities. He mentions Joe Rogan's ideas about brain stimulation and drug-like experiences, and memory wiping/replay. He is unsure about ethics. He references *Black Mirror* as worst-case thinking; Lex says people should consider best and average cases too. Noland jokes about hearing voices in his head since Neuralink.

Optimus, Touch, and Physical Interaction

Noland wants to control an Optimus robot. He would like sensation to transfer through Optimus. Physical interaction matters: 99% of things he cannot do need a caretaker. An Optimus could help him live independently. He would love an Optimus to hold a book open so he can read a physical book, with the smell and page color. He misses touch: clothes, fabric, weight, handshakes, hugs. He has asked God every day since his accident to move even just his hand so he could squeeze his mother's hand. He would also like to beat Bliss in chess on a physical board and humble him.

Faith and Hope

Noland says hardship is about understanding how much we need God; there is no light without dark. His accident strengthened his belief in God and made his interactions with God feel real and worthwhile. A friend responded by believing there is no God. Noland sees trials as tests that build character and help people grow. His hope for civilization comes from people. At Neuralink, he sees people who could have cushier jobs but want to better humanity. People care and are capable of helping each other. Bad things always exist and always will, but human resiliency and care give him hope. Lex thanks Noland and says the world is rooting for him.

Methods or Steps

Neuralink Implant Procedure

1. Patient selection: registry, medical records, eligibility criteria, prescreening, BCI home audit.
2. Pre-op mapping: fMRI identifies hand knob in motor cortex by imagined hand/finger movement.
3. Surgery preparation: anesthesia; intra-op CT confirms craniectomy location.
4. Human surgical steps: skin incision, craniectomy, dura resection, exposure of pia.
5. Robot insertion: R1 uses computer vision to avoid blood vessels; inserts 64 threads with 16 electrodes each into cortex.
6. Implant placement: surgeon places N1 implant into skull hole, replaces craniectomy, screws it in, applies dural substitute, closes skin.
7. Post-op activation: device turned on; signals recorded; patient learns to modulate spikes.
8. Calibration: open-loop center-out targets; attempted movements first; closed-loop cursor control; model updates.
9. Use and feedback: patient controls cursor, clicks, plays Webgrid, uses apps; gives feedback; firmware updates.

Neuralink Signal Processing Pipeline

1. Recording: 1,024 electrodes sample at just under 20 kHz with 10-bit resolution.
2. Raw data rate: about 200 Mbps.
3. Onboard ASIC: amplifies and digitizes.
4. BOSS algorithm: detects spikes and outputs six values (negative hump, middle hump, positive hump, timing).
5. Compression: sends only spike/event data via Bluetooth.
6. Bluetooth: packetized and batched at about 15 ms latency.
7. External app: B1 app runs machine-learning model to decode intended outputs like cursor movement and clicks.
8. Adaptation: model updates and firmware changes; spike band power can be used when individual spikes degrade.

First-Principles Engineering Algorithm

1. Question requirements and make them less dumb.
2. Delete the part or process step; if less than 10% is put back, delete more.
3. Optimize or simplify.
4. Accelerate cycle time.
5. Automate.

Safety and Reliability Testing

• Pathology department examines tissue slices.
• Necropsy, formalin fixation, sectioning, staining.
• Acute (0–3 months) and chronic (beyond 3 months) endpoints.
• Histology: astrocytes/microglia, neurons, collagen staining.
• Accelerated life tester: salt water and reactive oxygen species.
• +10°C roughly doubles aging; +20°C increases aging fourfold.
• Current implants at about 2.5 years equivalent to a decade.

Upgrade/Revision Approaches

• Current removal: cut thread, leave tissue intact, unscrew implant, plug hole.
• Future: through-dura insertion to reduce scarring and ease extraction.
• Future two-part implant: bottom threads/chips/radio/power under dura; top compute/battery above dura; upgrade top in about 10 minutes.
• Multiple Neuralinks possible in different cortices.
• Over-the-air firmware updates like Tesla.

Examples and Data

Caveats and Open Questions

• The provided transcript explicitly omits a middle section due to a 260,000-character limit. These notes cannot summarize that omitted material.
• Elon says Neuralink may help AI safety but is not a panacea or sure thing.
• The long-term visions—megabit BPS, superhuman vision, memory upload, billions of robots, eight billion Neuralinks—are aspirational or speculative in the conversation.
• There is irreducible risk in a new implant. Neuralink prioritizes medical necessity first and augmentation later as risk falls.
• Thread retraction is not fully solved; preventing it is the number one priority. The cause is tied to the human brain moving more than monkey or sheep brains, but the full mechanism remains an engineering problem.
• Memory restoration is limited: if the memory itself is destroyed, it cannot be recovered; only access or probabilistic reconstruction may be possible.
• AI risk estimates differ. Elon notes Geoff Hinton's 10–20% AI annihilation probability but does not present it as settled.
• Defining AGI is contested. Lex says humans may never acknowledge AGI; Elon distinguishes AGI from ASI and notes thresholds.
• Truth and bias are difficult to implement. Even small ideological biases could become dangerous at superintelligence scale.
• Blind-from-birth individuals may have visual cortex repurposed by other senses, so restored vision may be a different conscious experience.
• Mind reading, consciousness, and mood/memory manipulation remain scientifically and ethically unresolved. DJ Seo says BCI alone probably cannot read thoughts.
• Physical robotic arm and wheelchair control require FDA conversations because physical harm is possible, unlike digital cursor movement.
• The transcript's Earth-age phrase (“four and a half million years old”) conflicts with the cited one-millionth ratio for civilization; this appears to be a transcription or source inconsistency.
• The birth-rate/civilization-cycle claim is presented by Elon as a Durant-derived pattern with no exceptions, but the conversation does not provide a full demographic model.
• The ethics of using BCI for drug-like experiences, memory wiping, or mood switching are explicitly raised as beyond Noland's judgment and not resolved.

Action Checklist

• Continue scaling electrodes: target 3,000–6,000 channels by end of year, 16,000 by end of next year.
• Solve thread retraction and anchoring to preserve channel count over time.
• Reduce calibration time toward seven minutes or below.
• Improve firmware and UX: click-on-demand, more control parameters, advanced mode, more device connections.
• Expand studies from P1 to P2, P3, and beyond while maintaining safety endpoints.
• Use pathology, histology, accelerated life testing, and FDA oversight to validate safety.
• Develop through-dura insertion and a two-part upgradeable implant.
• Advance Blindsight and speech prosthetics while respecting regulatory and ethical constraints.
• Apply the first-principles algorithm: question, delete, optimize, accelerate, automate.
• Prioritize truth-seeking AI over politically correct lying.
• Support population stability and avoid civilizational collapse.
• Pursue multi-planetary life as risk mitigation.
• For participants: have fun, work hard, give feedback, and go to Neuralink with questions.

Conclusion

The conversation presents Neuralink as both a medical device and a long-term platform. In the near term, it aims to restore digital autonomy and communication for people with paralysis and neurological damage. Its first human participant, Noland, reached record BPS, suffered thread retraction, recovered through signal-processing changes, and became an active pioneer shaping the UX. DJ Seo explains the technical stack: threads, robot, ASIC, BOSS spike sorting, Bluetooth, inductive charging, histology, and scaling plans. Matthew MacDougall explains the surgical procedure and the deep human stakes of brain surgery. Elon Musk places Neuralink in a larger story about AI bandwidth, truth-seeking AI, Optimus, civilization, birth rates, Mars, and the meaning of life.

The recurring theme is communication: between neurons and electrodes, between humans and machines, between humans and each other, and between a civilization and its future. Neuralink's ambition is to increase bandwidth, restore lost function, reduce suffering, and perhaps give humans a better seat at the table as AI grows more powerful. The source ends with Lex quoting Aldous Huxley's *The Doors of Perception*: “We live together. We act on and react to one another, but always, and in all circumstances, we are by ourselves... From family to nation, every human group is a society of island universes.”