Appendix
Predictions, acknowledgments, resources, and the modality map.
Appendix cartouche: a tree-trunk cross-section of concentric growth rings

Appendix A: Predictions to Revisit

I made a number of predictions throughout this manifesto. This appendix collects the main ones in one place, each with a target year and a confidence, grouped by theme. I will update them as reality unfolds.

Predictions overview
Each prediction at its deadline, grouped by theme.
Confidence against deadline.

Appendix B: Friends and Thought Partners

This manifesto did not come to life in a vacuum. The people below have shaped how I see this field over years of conversation. Each, at some point, moved the thinking forward. The smart claims are theirs; the errors and overclaims are mine.

One person holds a special position here: my partner Olya, who read every draft, caught what did not work, and refused to let this sit in my backlog for another two years. It exists in finished form thanks to her.

Friends and Thought PartnersFriends and Thought Partners

On the hardware side: Sumner Norman and the Forest Neurotech / Merge crew, Mikhail Shapiro at Caltech, Ryan Field at Kernel, Nishita Deka at Sonera, Christina Maher, Lev Chizhov, Jay Sanguinetti at Sanmai, Mehmet Günal at Brill Neurotech, Hon Weng Chong at Cortical Labs.

On data and decoding: Trent McConaghy at Psyche, Jonathan Xu and the AllJoined team, Andreas Tolias and Sophia Sanborn at Enigma, Sharena Rice at Sanmai, Milan Cvitkovic at Integral Neuroscience, Avery Krieger at Constellation, Naomi Bashkansky at OpenAI/Conduit.

On consumer-facing devices: Ramses Alcaide at Neurable, Andreas Forsland at Cognixion, Meron Gribetz at Inner Cosmos, Akshat Sharma at Orbit, Peter Schlecht at Braingrade.

On the invasive and endovascular frontier: Matt Angle and Vikash Gilja at Paradromics, Tom Oxley at Synchron, Takufumi Yanagisawa at ivec, Marc-Joseph Antonini at Neurobionics.

On whole brain emulation: Isaak Freeman at MIT/Capable, Catalin Mitelut and Christian Larsen at Netholabs, Daniel Burger at Eightsix Science, and Adam Marblestone at Convergent Research.

On AI alignment and NeuroAI: Seth Bannon at 50y Ventures, Mike McCormick at Halcyon Futures, Judd Rosenblatt and Diogo de Lucena at AE Studio, James Fickel, Patrick Mineault at the Amaranth Foundation, Matthieu Kratz at Coefficient Giving, Frank Barbosa Nakasako at Palaestra Research, Evan Miyazono at Atlas Computing, Sean Escola at Inductive, and Ryota Kanai at Araya.

On funding and infrastructure: Michael Andregg and Maximilian Schons at Eon, Raphael Certain at Clarity, David Langer at Lionheart, Peter Zhegin at e184, Olha Dolinska at Oxford, Hilmar Eggertsson at 6 degrees capital, and Ryan Zurrer at Dialectic.

On conveners and community: Diana Saville at BrainMind and Allison Duettmann at the Foresight Institute, who bring much of this field together.

On the outside view: Karim Beguir at Instadeep and Raphael Volpert at Alexandria, Yash Shevde at Ursa Bio, Ansh Chopra at ZeroBillion, Viktor Bowallius at ThinkIF Technology, William Knottenbelt at Imperial College London, Adrian Johnston at Elyos AI, Mete Polat at Varto AI, Atlal Boudir at Louise, Daniel Faggella at Emerj, Rodolfo Rosini at Vaire Computing, Nikolas Ioannou at Hyper, Dan Girshovich at Tools for Humanity, and Auguste Pugnet, who do not build in neurotech but have been valuable thought partners nonetheless.

Appendix C: Further Resources

I mentioned a lot of companies and labs throughout the essay. I am consolidating the ones worth keeping an eye on here, for anyone wanting a bird’s-eye view of the field.

Companies and Labs Worth Watching

Non-invasive hardware: Cerca Magnetics (OP-MEG), Kernel (fNIRS), Nudge (ultrasound), Butterfly Network (ultrasound chip), Aleph Neuro (ultrasound), Sonera (biomagnetic neural sensing).

Non-invasive data and decoding: AllJoined, AE Studio, Mind Co, Neurable, Brill Neurotech, Wisear, Orbit, Conduit, Piramidal, Constellation, Sabi.

Neural data for AI: Inductive.

Open hardware and community: OpenBCI.

Sub-vocal and peripheral neural interface: CTRL-labs (now Meta neural interface): the EMG sub-vocal/intent path that is the closest non-invasive cousin to BCI in the consumer interface frame.

Prior consumer non-invasive wave (mostly cautionary): Muse, Emotiv, NextMind (acquired and then wound down by Snap).

Neuromodulation and stimulation: Sanmai Technologies, META-BRAIN (EU magnetoelectric + ultrasound neuromodulation), Integral Neuroscience.

Medical-grade non-invasive: NeuroPace.

Minimally invasive and invasive: Paradromics, Synchron, Neuralink, Precision Neuroscience, Starfish Neuroscience, Neurobionics.

Adjacent and meta: Psyche, Forest Neurotech (now Merge), Inner Cosmos, Cognixion, Enigma, Lymbic AI, Clarity, Cortical Labs.

Adjacent, endovascular, and emerging: ivec, Araya Inc., Neuradaptive, Braingrade, Synlinx (China’s national BCI incubator, Shanghai).

Whole brain emulation: Eon, Eightsix Science, Netholabs.

Investors and supporting capital: Amaranth Foundation, Jumpspace Ventures, Lionheart Ventures, e184.

In addition to companies, here is a list of write-ups for anyone wanting to go deeper into neurotech:

Further Reading
Companion essays
Related to the essay
  • Leopold Aschenbrenner, Situational Awareness, 2024. An AGI Manifesto.
  • AI 2027, 2025. A scenario forecast for AI over the coming years.
  • Sam Altman, The Merge, 2017. One of the early arguments that human and machine cognition would converge.
  • Rich Sutton, The Bitter Lesson. Two pages. Referenced several times in this essay.
Onboarding to AI alignment
AI scaling laws
Foundation models for the brain
Community resources
Adjacent: brain emulation and connectomics
  • Anders Sandberg and Nick Bostrom, Whole Brain Emulation: A Roadmap, Future of Humanity Institute, 2008. The canonical reference for what WBE would actually require, from substrate scanning through compute.
  • Isaak Freeman, WBE feasibility thesis. Recent cost analysis putting the all-in figure for human emulation at roughly $5-50 billion.

Appendix D: The Modality Map

A reference breakdown of the modalities named throughout this manifesto. The interactive comparison is in the modality mapThe modality map⁠▸; below is each modality in detail:

The Full Modality Map

Non-invasive BCIs

EEG: Measures electrical activity at the scalp with millisecond precision. Cheap and scalable but spatially noisy, it is the main test bed for early brain foundation models like AllJoined’s.

fNIRS: Uses near-infrared light to track blood oxygenation, an indirect signal that lags neural activity by several seconds. It is wearable, and Kernel is one of its leading commercial platforms.

Ultrasound: Uses sound waves to image blood flow and stimulate targeted regions deep in the brain. Still early, but promising for compact read-and-write interfaces. Butterfly Network put ultrasound on a chip; Nudge and Merge Labs are applying it to the brain.

MEG: Measures magnetic fields produced by neural activity with millisecond precision. Once limited to room-sized cryogenic scanners, it is becoming wearable through OPMs.

fMRI: Uses powerful magnetic fields to track changes in blood oxygenation, an indirect measure of neural activity. It offers millimeter-scale spatial resolution and has enabled some of the best non-invasive semantic reconstructions, but it is slow, expensive, and confined to room-sized scanners.

Invasive BCIs

Endovascular: Guides an electrode through the jugular vein into a blood vessel beside the motor cortex. Synchron’s Stentrode avoids open-brain surgery, trading signal precision for a far lower surgical burden.

ECoG: Places an electrode grid directly on the brain’s surface, capturing cleaner signals than EEG without penetrating the cortex. It has enabled speech decoding at 78 words per minute.

sEEG: Inserts thin depth electrodes into cortical and subcortical regions. Widely used to localize seizures, it has produced large clinical archives of human intracranial recordings.

Intracortical: Inserts microelectrodes into the cortex to record individual neurons, providing the highest-fidelity signals. Used by Neuralink and Utah arrays, it enabled Stanford’s 90-character-per-minute handwriting decoderHigh-performance brain-to-text communication via handwritingWillett, Avansino, Hochberg, Henderson, Shenoy · Nature · 2021Implanted microelectrode arrays in a paralyzed person’s motor cortex decoded imagined handwriting at 90 characters per minute, faster than most people type on a phone. The existence proof that the brain encodes the variables we want.⁠▸ but surgery and long-term durability remain major barriers to scale.

The Future

NV-Center: Defects in diamonds that detect tiny magnetic fields at room temperature. Not yet practical for BCI, but huge potential if we make it work.

Appendix E: Glossary

ALS
Amyotrophic lateral sclerosis. Progressive motor-neuron disease; one of the conditions BCI work has historically focused on.
Bandwidth
How much information you can send and receive per unit of time.
Bandwidth gap
The vast mismatch between how fast machines can think and communicate, and how slowly humans can process information.
BIDS
Brain Imaging Data Structure. A community standard for organizing neuroimaging data (MRI, EEG, MEG) in a uniform directory layout. Adoption is still uneven.
Connectome
The full wiring diagram of a nervous system: every neuron and every connection between them.
Diffusion models
Image-generation models that learn to reverse a noise process: start with pure noise, denoise step by step into a coherent image.
Dura mater
The tough outer membrane enveloping the brain and spinal cord. Semi-invasive devices sit beneath the skull but rest on top of this layer.
EEG
Electroencephalography. Reads voltages off the scalp. Cheap, ubiquitous, noisy.
Endovascular
Electrodes delivered through blood vessels rather than open craniotomy.
Exocortex
An external digital layer of your mind that extends your memory, attention, and mental computation beyond your biological brain while functioning as part of it.
FAIR
Meta’s Fundamental AI Research lab. The open-research arm of Meta AI.
fMRI
Functional magnetic resonance imaging. Room-sized scanner with millimeter spatial resolution; the highest-quality non-invasive signal we have, at the lowest portability.
fNIRS
Functional near-infrared spectroscopy. Measures blood oxygenation through the skull using infrared. Cheap, wearable, slow.
Glial signaling
Communication between glial cells, the non-neuronal majority of brain cells. Increasingly suspected to modulate cognition, not just support neurons.
Hemodynamic
Relating to blood flow. Sensors such as fNIRS read the blood-oxygenation shadow of neural activity through the skull.
Interpretability
The effort to understand what happens inside a neural network: which internal features and circuits produce its behavior.
Limbic system
A set of deep brain structures that drive emotion, motivation, and reward. It is the circuitry that advertising has always tried to reach from the outside.
MEG
Magnetoencephalography. It reads the tiny magnetic fields cast by neural activity. Millisecond timing, but classically needs a room-sized cryogenic scanner; OPMs are a new wearable path.
Microelectrode arrays
Tiny silicon chips with dozens to hundreds of hair-thin needles, surgically inserted into brain tissue to record the firing of individual neurons.
Microtubules
Tube-shaped protein structures inside neurons. Some theories (Penrose-Hameroff) propose they support quantum computation; mainstream neuroscience is still skeptical.
Motor cortex
The strip of cortex that plans and executes voluntary movement. Easiest brain region to decode from, because intentions there map cleanly to body actions.
NWB
Neurodata Without Borders. A standardized data format and ecosystem for neurophysiology recordings. The closest thing to a common schema across labs.
OPM
Optically-pumped magnetometer. A chip-scale magnetic-field sensor that works at room temperature; the path to wearable MEG.
Phase Lock Window
The brief interval between recognizing that AI will surpass us and the moment it does; the span in which humanity still has the agency to shape the terms of the relationship. Roughly now to the mid-2030s.
Quantum effects
Hypothesis that quantum-mechanical phenomena (coherence, entanglement) play a functional role in brain computation. Live but speculative, no consensus evidence.
Ransomware
Malicious software that encrypts a victim’s data or device and demands payment to restore access.
RLHF
Reinforcement learning from human feedback. The training step where humans rank a model’s outputs to steer it toward preferred behavior.
sEMG
Surface electromyography. Reads electrical activity of muscles through skin electrodes. Well-suited to subtle motor intent and silent speech.
Signal-to-noise
How strong the wanted signal is against background noise; the core obstacle to reading the brain through skull and skin.
Whole Brain Emulation
A hypothetical process of scanning, mapping, and simulating a biological brain on a computer so that the digital model functions identically to the original mind.
wpm
Words per minute. A standard measure of communication speed.