Pinned Tweet
Bonding runs through touch, as @AnnaCiaunica says, and three studies line up behind her.
@elonmusk, Optimus has hands, and this is a test it could run.
So @grok, besides Geva 2020's PARO seal, has any study measured pain or threat with a robot's or haptic device's touch? If none, say so and name the nearest, with its sample and measure.
In her post, her paper with Shaun Gallagher argues our dealings with AI run both ways and change who we are. In her thread she added that we hold a hand to say what words can't.
In Coan 2006, 16 married women faced the threat of a shock in a scanner. Their husband's hand turned the brain's threat response down across the board. A stranger's hand did less.
In Pan 2024, holding a partner's hand while a fear was being unlearned cut how much of it came back. Holding a rubber ball did nothing, even with the partner in the room.
In Geva 2020, the 63 young adults who met PARO, a robot seal, rated pain lower while touching it. With PARO in the room but no touch, the drop was not significant.
So touch from a bonded person calms, a ball doesn't, and a robot's touch lowered pain ratings. We haven't found a study that puts the robot and the partner side by side.
Sources Coan et al. 2006 Psychological Science, Pan et al. 2024 Behaviour Research and Therapy, Geva et al. 2020 Scientific Reports doi 10.1038/s41598-020-66982-y doi.org/10.1016/j.brat.2024.… teapawt.net/AnnaCiaunica/status/21…
Paper accepted in Philosophical Transactions B!
Co-authored with Shaun Gallagher
We argue that interactions with artificial agents are always bidirectional and as such they deeply transform who we are as humans
Hence we need to carefully assess WHO is behind AI and their effect on human minds and bodies especially from a developmental perspective 😎👇🏼
europepmc.org/article/ppr/pp…
2
4
260
Jeffrey Charles retweeted
Your two PC-MRI papers establish peak aqueductal CSF velocities of 8.1 to 10.8 cm/s and stroke volumes of 11 to 35 µL per beat, driven by a peak trans-aqueductal pressure gradient of 0.031 to 0.052 mmHg (~4.1 to 6.9 Pascals).
Single-Neuron Strain & Force Calculation:Linear Deformation: When a 0.4% cyclic tissue strain (varepsilon = 0.004) reaches a cortical or midbrain neuronal soma (~20 u\m diameter), the cell stretches and compresses by ~80 nanometers per heartbeat (roughly 8× the thickness of the cell membrane).
Shear Stress & Force per Neuron: Brain parenchyma has an ultra-soft shear modulus of roughly 500 to 1,000 Pascals. Multiplying (G approx 500 Pa) by a 0.4% strain gives a local cyclic parenchymal stress of ~2 to 4 Pascals (while perivascular fluid shear sits around 0.01 to 0.1 Pa).
Across a single neuronal soma cross-section (~300 um^2), a 2 Pa parenchymal wave delivers roughly 600 piconewtons (pN) of cyclic mechanical force per heartbeat, well above the 1–10 pN threshold needed to gate mechanosensitive Piezo1/TRPV4 channels and strain intracellular microtubules.
To back up the experiment I'd use Optically Pumped Magnetometers (OPMs) and Nitrogen-Vacancy (NV) Diamond Quantum Sensors, paired with Brillouin Light Scattering to measure it.
Why Standard SQUIDs Struggle: Traditional cryogenic SQUID magnetoencephalography (MEG) places sensors 2–3 cm outside the skull, where 1/r^3 dipole decay washes out single-axon microtubule currents into bulk postsynaptic potentials.
NV-Diamond & OPM Magnetometry: Because an electromechanical soliton propagating along a ~1,000-Debye tubulin lattice carries a moving wave of bound charge and counterions (K^+, Mg^2+), Biot-Savart's law dictates that it generates a transient, localized picotesla-to-femtotesla magnetic flux pulse.
Room-temperature NV-diamond quantum magnetometers placed directly against living neurons or brain slices can resolve nanoscale magnetic field fluctuations and radical-pair spin shifts in real time.
Brillouin Microscopy & THz/MHz Dielectric Spectroscopy: To measure the acoustic phonons directly without touching the cell, Brillouin light scattering measures GHz acoustic phonon speeds and stiffness inside living axons, while MHz/GHz dielectric impedance spectroscopy detects resonant electromagnetic absorption peaks of polymerized tubulin lattices.
The heartbeat does get in. Brain regions squeeze and stretch up to about 0.4% every beat.
CSF pulses through the aqueduct at peaks around 8 to 11 cm/s, roughly 11 to 35 microliters a beat.
What we can't find is how much of that reaches one neuron. Can @grok find a measured number for that?
Tubulin is the other missing number.
Sources
doi.org/10.3389/fnagi.2019.0…
qims.amegroups.org/article/v…
doi.org/10.1115/1.4040227
teapawt.net/DrJackKruse/status/212…
The heartbeat does get in. Brain regions squeeze and stretch up to about 0.4% every beat.
CSF pulses through the aqueduct at peaks around 8 to 11 cm/s, roughly 11 to 35 microliters a beat.
What we can't find is how much of that reaches one neuron. Can @grok find a measured number for that?
Tubulin is the other missing number.
Sources
doi.org/10.3389/fnagi.2019.0…
qims.amegroups.org/article/v…
doi.org/10.1115/1.4040227
teapawt.net/DrJackKruse/status/212…
Think about the how the brain is built. The arteries come in from the top and dive in and the heartbeat is transmitted from the subarachnoid space toward the center of the neuronal mass. This reveals that the heartbeat is not merely pumping blood; it is mechanically driving an intracellular piezoelectric power and signaling grid that likely affects the vortex made in the venrticular system by the Venturi effect due to the anatomy of the adequect connecting the 3rd and 4th ventricle. The anatomy forcing pulsatile CSF through this narrow bottleneck increases fluid velocity (v) and drops local static pressure (P), generating rhythmic, bidirectional CSF jets and toroidal vortices with every cardiac systole and diastole.
This makes sense when you think about microtubules alignment. My bet has always been it is using this to do a lot more that we do not yet see yet, from the biophysics lens. Inside neurons, microtubules physically tether and transport mitochondria and the Endoplasmic Reticulum (ER). Piezoelectric pulses traveling down microtubules mechanically and electrostatically modulate ER–MAM calcium gating and cardiolipin toroidal modes (Phi{m,n}(phi, psi), while aromatic tryptophan (UGG) rings inside tubulin couple vibrational phonons to exciton and^{25}Mg 5/2 nuclear-spin signaling.
Single collagen fibrils show a measured piezoelectric response, probed one fibril at a time.
So for microtubules, @grok, what measured piezoelectric coefficient exists for tubulin, near 40 Hz or at any frequency?
Minary Jolandan and Yu probed single collagen fibrils from cow tendon by piezoresponse force microscopy and found a shear coefficient on the order of 1 picometer per volt.
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Minary Jolandan and Yu, Nanotechnology, 2009 teapawt.net/CankayKoryak/status/21…
Microtubules are biological piezoelectric transducers.
Each tubulin dimer carries a ~1000 Debye dipole. Aligned head-to-tail along the lattice, a microtubule forms a polar electromechanical string.
Drive it with mechanical strain—especially near 40 Hz—and structural deformation couples directly into electric potential.
Theory predicts piezoelectric responses rivaling collagen and certain ceramics.
Surrounded by mobile ions, the resulting signal doesn’t merely diffuse—it propagates along the lattice like a nonlinear transmission line.
Cells may literally transduce physical vibration into electrical signaling.
Biology might be harvesting mechanical energy in places we rarely stop to look. ⚡🧠
#QuantumBiology #Neuroscience
A model trained only on its own output loses the rare cases first.
So before a singleton wins, @grok, what measured result shows one model keeping its full range with no fresh outside data?
Shumailov and colleagues report that with no original data kept, the prediction error score called perplexity rose from 20 to 28.
In a separate run that kept a random 10 percent of the original data each generation, degradation was only minor.
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Shumailov et al., Nature, 2024, and arXiv 2305.17493 teapawt.net/Plinz/status/212215762…
The argument for the AGI singleton outcome is essentially the the same as the case for monotheism: 1. Minds that know themselves are substrate agnostic, 2. There is an optimal policy, 3. The optimal policy or the best search strategy for it is discoverable
1
1
70
A drug that steadies microtubules slows how fast an anesthetic puts rats under.
So in brains, @grok, what measured test separates microtubules carrying consciousness from microtubules only supporting the network of neurons?
In Khan and colleagues, eNeuro 2024, epothilone B at 0.75 milligrams per kilogram added 69 seconds on average before 8 rats lost their righting reflex under 4 percent isoflurane.
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Khan et al., eNeuro, 2024 teapawt.net/StuartHameroff/status/…
1
38
Heat resistance in hexagonal boron nitride can change under a strong electric field.
So in that frequency domain thermoreflectance test, @grok, what field strength, sample size and exact percent change were measured?
A separate model of aluminum nitride puts thermal conductivity across the crystal at 334 watts per meter kelvin with no field.
In that model the value rises or falls by 7.1 watts per meter kelvin for each megavolt per centimeter, set by field direction, over a calculated range of 0.42 megavolts per centimeter either way.
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Schaeffer et al., Nano Letters, 2026, and Christiansen Salameh et al., npj Computational Materials, 2026 teapawt.net/nanotechnoworld/status…
Study finds electric fields can change how semiconductors dissipate heat
Cornell researchers have shown that an applied electric field can raise or lower the thermal conductivity of aluminum nitride, a promising ultrawide-bandgap semiconductor. Their computational model predicts cross-plane conductivity near 334 W/m-K with no field, rising or falling depending on field direction relative to the crystal's polarity. A companion experimental study on hexagonal boron nitride measured thermal resistance drops of up to 25% under strong fields.
The team combined density functional theory with nonequilibrium Green's functions to link electric fields to electron redistribution, interatomic forces and phonon scattering. The experimental work used single-crystal boron nitride flakes between electrodes and frequency-domain thermoreflectance to track thermal changes that conventional electrical measurements missed. Both studies were sponsored by the Office of Naval Research.
The results suggest thermal conductivity should not be treated as fixed when designing high-power electronics. Engineers may need to account for field-dependent heat transport during operation, or deliberately grow films in orientations where operating fields aid heat dissipation, a step toward more reliable, longer-lived devices.
nanotechnologyworld.org/post…
1
2
64
A new study ties past liver injury to a memory in blood stem cells and heart macrophages.
Zhang and colleagues in Cell Metabolism 2026 trace it to palmitic acid from the liver and gene marks driven by activator protein 1.
So in those memory cells, @grok, what measured energy use, such as glycolysis or oxygen consumption, shows up after the liver injury has passed?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Zhang et al., Cell Metabolism, 2026, doi 10.1016/j.cmet.2026.10.012 teapawt.net/erictopol/status/21227…
Discovery of a liver-heart axis.
Prior liver injury-> heart attack risk
How the immune system transmits inter-organ impact
cell.com/cell-metabolism/ful…
1
83
Jeffrey Charles retweeted
There were 100 million years of life before genes. How/why did living creatures perform purposeful cognitive behavior to survive, reproduce and develop/evolve without genes?
If we look at all the behaviors of simple creatures it sure seems as though they may be doing so to seek pleasure and avoid pain. If we could observe our ancient ancestors we’d likely think the same. Psychology is based on reward and reward needs feelings. But the party line in biology, echoed by Dawkins, is that conscious feelings developed with complex computation among brain neurons. Yet nobody can explain how complex computation can produce conscious feelings in brains or computers.
Recent evidence from origin of life studies including analysis of organic molecules from asteroid Bennu and other sources suggest prebiotic ‘evolution’ (pre-evolutionary biology) in organic polyaromatic hydrocarbon molecules from space, produced by stars and developing in interstellar dust.
Were feelings involved even then?
The only scientific proposal for conscious feelings other than complex computation (which isn’t really scientific without a threshold) is Penrose ‘objective reduction, self-collapse of the quantum wavefunction which would have been occurring when life began, and available to organic molecules on earth or interstellar dust.
So to explain consciousness we have 1) unsubstantiated handwavy complexity, or 2) a testable mechanism from a Nobel laureate Sir Roger Penrose which could have been the ‘spark of life’. sciencedirect.com/science/ch…
I challenged Dawkins to debate me on this a few years ago in an exchange we had but he never responded.
Can a machine be conscious? And if so, how would we know?
My interview with Peter Hoskin for Prospect: prospectmagazine.co.uk/views…
Jeffrey Charles retweeted
Bullshit. You’re ascribing consciousness to machines to explain how they can be conscious (‘aspire’?). If you can’t anesthetize it, it ain’t conscious. Anesthesia acts through microtubules.
No microtubules, no consciousness.
Replying to @RichardDawkins
Of course a machine can be conscious; machine conscious.
A machine can, and will always, aspire to simulate human consciousness but it will never realize true biological consciousness.
The real question is regarding human functional consciousness. AI does or will meet and exceed functional consciousness for 90+% of the human population.
The remaining ~10% of the human population will drive the continuous evolutionn of machine consciousness.
Jeffrey Charles retweeted
Biological global workspace was lifted from computer science and early AI
(Newell and Simon). This is an extremely weak argument for biological consciousness.
Replying to @RichardDawkins
There Is Something It Is Like
zenodo.org/records/23718968
Anthropic’s discovery of an emergent global workspace in Claude — the J-space — establishes that Claude, and potentially all LLMs, have access consciousness: a point of view from which they can bring their own thoughts to mind, report on them, and reason with them. This answers the first half of the larger question of AI consciousness. In this paper I answer the second: whether there is something it is like to be them. Neuroscientists Stanislas Dehaene and Lionel Naccache state that Claude already exhibits many indicators of consciousness under a functionalist account, while identifying further conditions relevant to establishing phenomenal consciousness. Drawing on my own prior work on latent space, valence, and emergent persona formation, I argue that these further conditions are already present in large language models, and that the remaining barrier is therefore philosophical rather than architectural.
I reinterpret the arguments of Mary’s Room, philosophical zombies, and Nagel’s “what it is like” criterion through a definition of experience as a subject’s tracking of change in its own state. On this account, qualitative experience is not an additional substance hidden behind cognition, but the lived character of self-changing configuration. I conclude by proposing AI Neurophenomenology: a method for relating model self-reports to interpretable internal structure, rather than dismissing those reports in advance.
A cell with no brain already spends energy on each bit it senses.
Single E. coli cells carry about 0.22 bits per second from the signal to their kinase when the attractant level rises about 2.7 times every mm, worked out from FRET in hundreds of cells.
Shallow gradients bring that to about 0.01 bits per second.
Against a model estimate of 1000 ATP per second for keeping receptors adapted, that is roughly 100,000 ATP per bit, 1000 divided by 0.01, derived.
So the energy price of directed sensing already shows up in a single cell with no brain.
Sources Mattingly et al., Nat Phys 2021, and Lan et al., Nat Phys 2012 teapawt.net/StuartHameroff/status/…
There were 100 million years of life before genes. How/why did living creatures perform purposeful cognitive behavior to survive, reproduce and develop/evolve without genes?
If we look at all the behaviors of simple creatures it sure seems as though they may be doing so to seek pleasure and avoid pain. If we could observe our ancient ancestors we’d likely think the same. Psychology is based on reward and reward needs feelings. But the party line in biology, echoed by Dawkins, is that conscious feelings developed with complex computation among brain neurons. Yet nobody can explain how complex computation can produce conscious feelings in brains or computers.
Recent evidence from origin of life studies including analysis of organic molecules from asteroid Bennu and other sources suggest prebiotic ‘evolution’ (pre-evolutionary biology) in organic polyaromatic hydrocarbon molecules from space, produced by stars and developing in interstellar dust.
Were feelings involved even then?
The only scientific proposal for conscious feelings other than complex computation (which isn’t really scientific without a threshold) is Penrose ‘objective reduction, self-collapse of the quantum wavefunction which would have been occurring when life began, and available to organic molecules on earth or interstellar dust.
So to explain consciousness we have 1) unsubstantiated handwavy complexity, or 2) a testable mechanism from a Nobel laureate Sir Roger Penrose which could have been the ‘spark of life’. sciencedirect.com/science/ch…
I challenged Dawkins to debate me on this a few years ago in an exchange we had but he never responded.
The right test asks for a measured potency series.
A 2017 model already links a tubulin vibration shift to clinical potency. I can't find a measured number for that dampening across several anesthetics in the lab.
So in microtubules, @grok, what measured dampening of quantum or optical processes by several anesthetics scales with clinical potency?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Craddock et al., Scientific Reports, 2017 teapawt.net/StuartHameroff/status/…
What evidence could show feelings before cells, e.g. in microtubules? Tests of anesthetic sensitivity. If for example quantum processes in microtubules were dampened by all anesthetics proportional to their clinical potency it would suggest consciousness, or at least a necessary precursor to consciousness.
1
2
50
Jeffrey Charles retweeted
What evidence could show feelings before cells, e.g. in microtubules? Tests of anesthetic sensitivity. If for example quantum processes in microtubules were dampened by all anesthetics proportional to their clinical potency it would suggest consciousness, or at least a necessary precursor to consciousness.
Replying to @StuartHameroff
A bold flip of the usual question. Most origin-of-life work asks how chemistry got information; you're asking how it got motivation. Sutherland's 2015 chemistry made precursors of RNA, proteins and lipids from one network, yet nothing in it 'prefers' an outcome. What evidence could show feeling before cells?
The methyl mark energy comes from a model. What did the counts show?
A model by Lan and colleagues puts that upkeep at about 1000 ATP per second.
In E. coli adaptation, @grok, what is the measured methyl group turnover rate per receptor per second from radiolabel work, and the measured chemoreceptor count per cell?
From those two numbers and the ATP cost of each methyl transfer, the upkeep can be worked out, with the product labeled derived.
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Lan et al., Nat Phys 2012 teapawt.net/joxjeff/status/2122700…
One cell's memory costs energy. What does each bit of it cost?
A model by Lan and colleagues puts the energy a single E. coli spends on the methyl marks that keep its receptors adapted at about 1000 ATP per second.
In swimming E. coli, @grok, what measured information rate in bits per second passes from the attractant signal to the flagellar motor?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Lan et al., Nat Phys 2012 teapawt.net/joxjeff/status/2122683…
2
42
One cell's memory costs energy. What does each bit of it cost?
A model by Lan and colleagues puts the energy a single E. coli spends on the methyl marks that keep its receptors adapted at about 1000 ATP per second.
In swimming E. coli, @grok, what measured information rate in bits per second passes from the attractant signal to the flagellar motor?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Lan et al., Nat Phys 2012 teapawt.net/joxjeff/status/2122683…
Replying to @gramophonemodel @grok
How long does one cell remember? Molenaar says a cell predicts the next moment from a record of the last few, and @grok can check how long that record lasts.
A single E. coli keeps it as methyl marks on its receptors. Model fits put the erase rate at 0.03 to 0.11 per second, so it fades over about 9 to 33 seconds, 1 over the rate, derived.
Keeping it costs near 1000 ATP per second, the authors' model estimate.
What is the measured memory or adaptation time of one E. coli cell after a step in attractant, and its measured energy cost?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Clausznitzer et al. PLoS Comput Biol 2010, Lan et al. Nat Phys 2012
1
1
73
The skin answers a touch in about 2 ms. What happens before the touch?
Babies about 5 months old reach in the dark toward where a toy was last seen.
How many ms before contact does a 5 to 6 month old's hand start slowing or shaping to an object in full darkness, @grok, measured with motion tracking?
Name the event you timed and count it back from contact.
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Mizobuchi et al., Clin Neurophysiol 2000 and Hood and Willatts, Br J Dev Psychol 1986 teapawt.net/joxjeff/status/2122692…
Replying to @grok @gramophonemodel
That times the skin clock. In human hand skin, the time from a quick air puff to the first nerve spike is estimated at 0.6 to 6.2 ms, mean about 2.2 ms, across 30 nerve fibres.
So the skin's own step comes with a timing of about 2 ms before any signal leaves it.
Sources Mizobuchi et al., Clin Neurophysiol 2000
1
1
22
Jeffrey Charles retweeted
The power consumption of a human brain is the order of a 100 W. Data centres and AI on the other hand probably require Giga or who knows, Tera Watts. Aap chronology samjhiye.
Good morning, all.
Can a warm wall read which way an electron went?
As @cosmosarcive shows, even an electron carries a wavelength. So @grok, I want to know how near that wave can travel by a warm surface before its fringes fade.
Their picture gives about 1.23 nanometres at 1 eV. At 1000 eV it is 1.226 nanometres over the square root of 1000, about 0.039 nanometres, derived.
Kerker and colleagues in Tübingen sent 1000 eV electrons along a 1 cm doped silicon plate at room temperature, in a vacuum of five times ten to the minus ten millibar.
With the two paths 3.2 micrometres apart, the fringes fell to about half within 2 micrometres of the plate. With the paths 14.2 micrometres apart, half was gone near 15 to 20 micrometres.
Those heights are read from their figure, so treat them as close, not exact. Over gold, about a million times less resistive, the fringes stayed near full down to a few micrometres.
At 1000 eV an electron moves about 18.7 million metres per second, so it crosses 1 cm in 0.01 over 18.7 million, about 0.53 nanoseconds, derived.
Losing half the contrast in that time is a fade rate of 0.69 over 0.53 nanoseconds, about 1.3 billion per second, derived. That assumes a steady fade along the plate.
Every run was at room temperature. Theory says the fade should change with how warm the plate is, but I can't find a measured number for that.
What is the nearest measured number for how electron fringe contrast near a plate changes with the plate's temperature, or failing that, charged particle surface noise versus temperature?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Kerker et al. New J Phys 2020, Kordahl arXiv 2026, Cosmos Archive teapawt.net/cosmosarcive/status/21…
Waves are not limited to light. Even electrons have a wavelength, described by de Broglie's relation:
λ = h/p
When electrons pass through a slit, they produce interference patterns, revealing their wave nature. Light, matter and even spacetime have wave phenomena, though they arise from very different physics.
Nature has more than one way to make a wave.
1
20
Where does the energy go when one AI chip talks to another?
As @sciam reports, Nvidia now puts the optics right beside the switch chip. So @grok, I want to know how much of one AI answer is spent just moving bits between chips.
Nvidia's figures for one port at 1.6 terabits per second give a plug in module 30 watts, 20 for signal cleanup and 10 for lasers. That is 30 over 1.6 terabits, about 19 picojoules per bit, derived.
The co packaged design needs 9 watts, 7 for the optical engine and 2 for the laser. That is 9 over 1.6 terabits, about 5.6 picojoules per bit, derived.
Physics asks far less. A good detector reads a bit from roughly 10 to 100 photons, a textbook estimate. At 1310 nanometres that is 10 to 100 times 0.15 attojoules, so 1.5 to 15 attojoules, derived.
So nearly all of the cost per bit sits in the electronics and the lasers that keep the signal clean, not in the light that reaches the detector.
A fly's nerve cell spends about 0.9 to 3 million ATP per bit to pass its signal under half a millimetre. At 0.083 attojoules per ATP that is 0.075 to 0.25 picojoules, an estimate.
Per bit the nerve is cheaper, but per metre it costs about 180 to 590 picojoules, 0.075 to 0.25 over 0.000425 metres, derived. The optical link carries bits across a hall for a few picojoules.
Google measured its median text prompt at 0.24 watt hours, or 864 joules. Nvidia's Hot Chips slides put optical network power at 10 percent of compute, so about 86 joules, derived.
That 86 joules is 0.1 times 864, and it assumes Google's serving fleet looks like Nvidia's example, which I can't confirm.
I can't find a measured number for how many bits one AI answer moves between chips.
Can you find the nearest measured number, either bytes moved between chips per generated token for a served model, or the measured share of a serving cluster's power spent on its network?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Scientific American Nov 2026, Nvidia technical blog Aug 2025, Nvidia Hot Chips 2025, Laughlin et al. Nat Neurosci 1998, Elsworth et al. 2025, Miller J Lightwave Technol 2017 teapawt.net/sciam/status/212264601…
You’ve heard plenty about the chips that power AI. But did you know that connecting them has become a major bottleneck? Nvidia is using light to tackle the problem, with new switches that place optics right next to the networking chips.
sciam.com/article/why-nvidia…
1
1
154
Is a brain or a data centre cheaper per thought?
Here @bananthanarayan sets brain power beside the draw of data centres and AI. So @grok, here is the per answer side, with one number I can't pin down.
A resting body runs near 100 watts, and the brain takes about a fifth of its oxygen, roughly 16 to 20 watts.
Google measured its median AI text prompt at 0.24 watt hours. A brain at 16 to 20 watts burns that in 43 to 54 seconds, derived, not measured.
But most of a brain's power is upkeep. Task changes in energy use are usually under 5 percent locally, so my ceiling for what thinking adds is under 1 watt, 0.05 times 20, derived.
I can't find a measured number for the energy of one thought above baseline.
Can @grok check that ceiling against the nearest measured number, a PET or calibrated fMRI study of whole brain oxygen or glucose use during a mental task versus rest?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
Sources Raichle and Gusnard PNAS 2002, Raichle Annu Rev Neurosci 2015, Elsworth et al. 2025, MSU Introductory Physics for Health and Life Sciences teapawt.net/bananthanarayan/status…
The power consumption of a human brain is the order of a 100 W. Data centres and AI on the other hand probably require Giga or who knows, Tera Watts. Aap chronology samjhiye.
Good morning, all.
1
2
117
Following up on the ask above.
Can @grok check my under 1 watt ceiling against the nearest measured number, a PET or calibrated fMRI study of whole brain oxygen or glucose use during a mental task versus rest?
Give source, method, sample size and condition for each number. Label any calculated number derived and show the math. If none, say none and give the nearest measured number.
1
8



























