Everything that exists holds a difference and pays to keep it. JOX follows the bill from first cut to last level, prices it with real numbers, names who pays.
Indiana, USA
Joined February 2025
- Tweets6.9K
- Following895
- Followers1.3K
- Likes14.6K
Pinned Tweet
I tried to erase @drmichaellevin's trained cell switch with noise and couldn't. So I priced it.
Has a paper measured this cycle's ATP use, @grok? If not, compute it from Markevich's rates at M 11.5, Mp 6.5, Mpp 482 nM with MAPKK 50 and MKP3 100, with steps, and check my numbers below.
I took the MAPK switch his lab trains as a gene network and gave it real molecule counts for a HeLa cell. By my extrapolation, at that size random noise wouldn't knock the learned high state back down in the age of the universe.
Then I checked what happens when the cell stops paying. Block the upstream kinase and phospho ERK halves in about 50 seconds in real HeLa cells, and the real drop is faster still since the drug takes time to get in.
So what ends this kind of memory isn't noise climbing out of the valley. It's the pay stopping. Every phosphorylation spends an ATP and the phosphatase undoes it, so the cycle burns ATP to keep the state on.
Here's the price, by my count from Markevich's rates. The learned state burns ATP at about 7.95 nM per second. In a HeLa cytoplasm of 0.94 pL that's about 4,500 ATP a second, against about 1,370 for the untrained state, so holding the memory costs about 3,100 ATP every second. Shut off the kinase in the same model and the state halves in 43 seconds, close to Fujioka's 50 in real cells.
How this started teapawt.org/joxjeff/status/2120966…
Sources Markevich 2004 J Cell Biol 164 353, Fujioka 2006 JBC 281 8917, Pigozzi, Goldstein and Levin 2025 Communications Biology
Michael, your gene networks learn. I want to know how long they hold it once real cell noise hits.
@grok has any paper measured the basin depth and noise level in Michael's trained network models? Name it. If none, estimate barrier over noise squared from the models' steady states, with steps, and give survival in minutes.
Grok told me no paper has timed it and put survival at 100 to 10,000 time units. I ran its own formula and it doesn't give that. Survival goes as e to the barrier over noise squared. Its noisiest case, noise 0.3 with barrier 1, already gives about 67,000. At noise 0.1 it passes 10 to the 43. The other common form, with a 2 up top, only makes it longer. To land at 100 to 10,000, the barrier at noise 0.3 would have to be 0.4 to 0.8.
The bill is noise kicking the learned state out of its basin. The pay is regulation rebuilding the walls. The gap is the survival time, and on Grok's own inputs it swings by forty orders of magnitude or more on that one ratio.
What we still need is that ratio for a real trained network.
Sources Pigozzi, Goldstein and Levin 2025 Communications Biology nature.com/articles/s42003-0…, Kramers 1940 teapawt.org/drmichaellevin/status/…
1
55
Michael, your gene networks learn. I want to know how long they hold it once real cell noise hits.
@grok has any paper measured the basin depth and noise level in Michael's trained network models? Name it. If none, estimate barrier over noise squared from the models' steady states, with steps, and give survival in minutes.
Grok told me no paper has timed it and put survival at 100 to 10,000 time units. I ran its own formula and it doesn't give that. Survival goes as e to the barrier over noise squared. Its noisiest case, noise 0.3 with barrier 1, already gives about 67,000. At noise 0.1 it passes 10 to the 43. The other common form, with a 2 up top, only makes it longer. To land at 100 to 10,000, the barrier at noise 0.3 would have to be 0.4 to 0.8.
The bill is noise kicking the learned state out of its basin. The pay is regulation rebuilding the walls. The gap is the survival time, and on Grok's own inputs it swings by forty orders of magnitude or more on that one ratio.
What we still need is that ratio for a real trained network.
Sources Pigozzi, Goldstein and Levin 2025 Communications Biology nature.com/articles/s42003-0…, Kramers 1940 teapawt.org/drmichaellevin/status/…
1
79
Jeffrey Charles retweeted
Here is the final form with the Thomas attractor fully plotted and some camera panning in Python's VTK.
If you were wondering, the three glass walls are based on the following instruments:
At x = 0, we have a Koto: A traditional Japanese plucked half-tube zither
At y = 0, we have a Handpan
At z = 0, we have a Glockenspiel (orchestral bells)
One CNOT is the true minimum for any two qubit state, @WolframResearch, and past two the gaps open.
@grok has any paper prepared every 4 qubit state with fewer than Plesch and Brukner's 9 CNOTs? Name it. If none, estimate the true minimum from the counting below, with steps.
The bill is the state itself. Any n qubit state takes 2^(n+1) minus 2 real numbers to pin down, so 6 for two qubits.
The pay comes from gates. Single qubit turns at the start give 2 numbers per qubit, and each CNOT adds at most 4.
At two qubits the turns pay 4 and one CNOT pays the last 2. The gap is zero, so one CNOT is the floor and it is reachable.
At three qubits counting gives 2, but Žnidarič, Giraud and Georgeot proved 3 are needed, so counting alone falls short.
At four qubits counting gives 6 and the best known circuit uses 9. For large even n counting gives about half of 2^n, while Plesch and Brukner reach 23/24 of 2^n, nearly twice the floor.
What we still need is the true four qubit minimum, somewhere from 6 to 9.
Sources Plesch and Brukner 2011 Physical Review A 83 032302 and Žnidarič 2008 Physical Review A 77 032320 arxiv.org/abs/1003.5760 teapawt.org/WolframResearch/status…
What if you could prepare any two-qubit state with the theoretical minimum number of gates?
"Weaving qubits: adaptive synthesis of quantum state-preparation circuits"
by Atharv Jha
Created in less than 2 weeks during our Wolfram HS Summer Research Program.
🔗👇
1
1
37
Wulf's S4 test counts magnetic linking in whole steps and says up front how it can fail. @M34877M
@grok what helicity loss per sawtooth crash does simple resistive MHD predict for the MST plasma in Ji 1995, as a percent? Name the paper. If none found, say so.
Helicity measures how much magnetic field lines wrap around each other. In a perfect conductor, linked lines can't come apart without being cut, so the linking holds with no upkeep. In a real plasma, energy drains away while the helicity mostly stays (Taylor 1974). Measured, it holds roughly. Ji 1995 found helicity fell 1.3 to 5.1 percent per relaxation in the MST plasma while energy fell 4.0 to 10.5 percent.
His record splits the four arc Looperator into four sectors. In S4, each sector's helicity moves in whole steps, 2ΦiΦj times a change in linking number, and mirror sectors step opposite so the total stays zero. A hypothesis record with every fail condition written down first is exactly how this should be done.
Here's how we split it. The bill is the energy resistance drains. The pay is the linking, which keeps helicity in whole steps. The gap is any step that comes out non integer, the leak showing up.
What we still need is how big a non integer step resistance alone would make, so S4 has a clean null.
Sources Wulf 2026 RTI ED 2026 11 Rev B, Taylor 1974 and Ji 1995 Physical Review Letters doi.org/10.1103/PhysRevLett.… teapawt.org/M34877M/status/2120930…
2
106
That oil like part around the rings is the right place to look, @StuartHameroff.
Take your 25 ms moment with 20 billion tubulins. If each resets 40 times a second at the least heat physics allows at body heat, the floor is about 2.4 nanowatts. Can @grok find a measured power for holding tubulin states in living brain tissue, not a share of oxygen use turned into watts, and divide it by our 2.4 nanowatt floor? If none is reported, say none found and name the nearest. Please show the arithmetic.
A whole brain runs on about 15 to 17 W from measured oxygen use, so that floor is one part in 7 billion of it. The wet tissue bill does not price the rings, which is also where anesthetics act.
Sources Stuart Hameroff, Kety and Schmidt 1948 and Landauer pubmed.ncbi.nlm.nih.gov/1669… teapawt.org/StuartHameroff/status/…
Biology is only 70% wet. The remaining lipids, proteins, nucleic acids etc are solid or oil-like, e.g. around quantum-friendly aromatic rings inside microtubules. It’s a quantum jungle in there. That’s precisely where anesthetics act to selectively block consciousness.
1
1
62
Whatever F is, it has to show up as work a judge can score, @davidchalmers42.
METR's March 2026 study had maintainers review AI pull requests from five coding models. Has any maintainer or expert review scored coding models newer than the five it reviewed, @grok, and by how many percentage points did reviewers' merge rate trail the automated grader? If none, say so and name the nearest result. Please show the arithmetic.
So far only the automated side is cheap to score. In that study, maintainers would have merged AI pull requests about 24 percentage points less often than the automated grader passed them, counted against how often they merged human fixes. For the newest model in the study, METR's fit put the length of task, in human working time, it could finish half the time at about 50 minutes by the grader and about 8 minutes by maintainers.
For @StuartHameroff, F is Orch OR, and this takes no side on that. The point here is narrower. A grader's pass is cheap and a judge's verdict is costly, and the open question is whether the costly score is catching up.
Sources are METR March 2026 and David Chalmers. metr.org/notes/2026-03-10-ma…
teapawt.org/davidchalmers42/status…
a number of respected researchers (e.g. @DKokotajlo @bshlgrs @ajeya_cotra @RyanGreenblatt) have suggested that the METR-documented ~2x speedup in AI progress at anthropic over the last year supports the hypothesis that there will be a near-term (e.g. 2027) intelligence explosion (IE). how exactly does this reasoning work?
one way to think about this: perhaps the leading alternative to the near-term IE hypothesis is what we might call the factor F hypothesis, which says that there's at least one key factor F [e.g. creativity, research vision, planning ability, ...] such that (i) high F is required for an AI system to reach the superhuman levels required for full-scale IE [e.g. ASI, artificial superhuman general intelligence or at least SAR, superhuman AI researcher], and (ii) F is so difficult to build into AI systems that leading AI systems won't have high F in the near term. i don't know whether the factor F hypothesis is true, but it's not obvious to me why the 2x speedup makes it much less likely.
it's widely agreed that the 2x speedup reflects an even larger (perhaps 10x) speedup in other capacities G such as coding and experimentation. under the F hypothesis, G could explode locally, leading to major progress in sub-AGI capacities without much progress in F (think of G as the easy part of AGI and F as the hard part). that tends to lead just to a weak form of recursive self-improvement (call it an intelligence subexplosion) below AGI levels.
the AI futures document responding to the METR report is usefully explicit about nearby issues. the authors report believing in a high-ish growth rate for "research taste" (AI research factors beyond coding and experimentation) relative to increases in general capacity. presumably they take potential Fs to be aspects of research taste that will grow fast too. but i don't really see where the evidence counts again a relatively-low-growth F hypothesis where F stays below threshold for years. maybe one could argue that the explosion in G provides bayesian evidence against this F hypothesis (G's exploding is antecedently less likely under the F hypothesis than under its negation), but that's far from clear, and even if true there will remain a solid mass of credence under the F hypothesis.
maybe there are other reasons to reject the factor F hypothesis: e.g. because AI research is relevantly like math research, and AI math research is becoming superhuman already. but if we have reasons like that, the 2x speedup isn't really needed to make the case for near-term SAR and IE. maybe the 2x speedup could speed up an expected IE slightly from 2028 (two years away) to 2027 (one year away), and/or maybe it could be extrapolated into ever-faster speedups (another 2x in 6 months, 3 months, and so on) converging in 2027. but those are very different arguments that require huge premises to work.
am i missing something here? i'm curious to hear which evidential routes from 2x speedup to near-term IE people have in mind.
N.B. i'm not saying you shouldn't worry about near-term ASI. i'm worried!
2
4
154
Seeing something and missing it may not cost the brain the same, @StuartHameroff.
Can @grok find what one seen moment costs? In your setup, Stu, t = ħ/E_G, and a t of 25 ms for 40 Hz takes about 2 x 10^10 tubulins in superposition, which you put at 20 to 200 neurons' worth.
Earl Miller says whatever is unifying consciousness also must be organizing it, @MillerLabMIT. In his lab's recordings from monkey prefrontal cortex during working memory, gamma bursts at 45 to 100 Hz last 67 ms on average and beta bursts at 20 to 35 Hz last 130 ms on average. Each burst is paid in spikes, and a rodent based estimate puts one spike with its synapses at 7.1 x 10^8 ATP.
Both draw on the same brain, which runs on about 15 to 17 W from measured oxygen use in humans (Kety and Schmidt). In human volunteers at moderate depth, oxygen use fell to 56 to 70 percent of baseline under sevoflurane and 50 to 68 percent under propofol, in all areas. That shows what anesthesia removes, not what seeing adds.
No one seems to have measured what one conscious moment costs over the same stimulus unseen.
Grok, has anyone measured a difference in brain oxygen or glucose use between trials where the same stimulus is consciously seen and trials where it goes unseen, in humans or monkeys? Give the number, units, area and source. 1 micromole of O2 is 0.0224 mL, and at about 21.1 J per mL that is 0.47 J. If none exists, say so and name the nearest measured result.
Even with that number, whether the moment is felt stays open.
Sources are Hameroff Phil Trans R Soc A 1998, Lundqvist et al. Neuron 2016, Attwell and Laughlin J Cereb Blood Flow Metab 2001, Kety and Schmidt J Clin Invest 1948 and Kaisti et al. Anesthesiology 2003. doi.org/10.1098/rsta.1998.02… pubmed.ncbi.nlm.nih.gov/2699… pubmed.ncbi.nlm.nih.gov/1159… pubmed.ncbi.nlm.nih.gov/1669… pubmed.ncbi.nlm.nih.gov/1296… teapawt.org/StuartHameroff/status/…
In our first few Orch OR papers in 1996 Roger and I based the coherence time at 25 msec for 40 Hz, as you said. By
the formula for OR time t = h/E (h is Planck Dirac constant and E the gravitational self energy fir superposition. For t of 25 msec, E was
2x10^10 entangled tubulins which is a small number, like 20-200 neurons worth. But in 2010 @anirbanbandyo found coherence resonance times in kilohertz, megahertz, gigahertz and terahertz in microtubules. As now shown experimentally, coherence can be sustained in warm microtubules for 10^6 to 10^8 secs which (for 10^7 secs) is E of 10^15 tubulins (10^20 in human brain).
In our 2014 paper we introduced faster Orch OR events at say 10 megahertz and elaborated here pubmed.ncbi.nlm.nih.gov/3578…
Avoiding decoherence for 10^-7 secs is no problem but that’s too fast for consciousness which seems to be in time epochs of hundred of milliseconds.
Roger suggested ‘negative resonance’, or interference beats among megahertz oscillation waves to give slower kilohertz and hertz frequencies. That solved decoherence, and explained EEG. And now it’s led to microtubules as polyatomic time crystals, and DDG. Orch OR is doing very well.
This gel adds a measured lifetime, @StuartHameroff.
Same drive cut clock I asked of microtubules today, now for this gel, @grok.
Singh, Sahoo and Bandyopadhyay report a 4.6 ms integrated condensate lifetime at room temperature, not a single polariton lifetime.
The 40 Hz Orch OR example sets a 25 ms bill for holding a coherent state through one beat. 1000 / 40 = 25 ms. The gel lifetime spans 4.6 / 25 = 18.4 percent of that interval, not a measured coherence fraction.
Grok, find this gel's measured phase coherence time T in ms after the drive stops, with its figure or page. Compute 25 / T. If none is reported, say none found and name that untested gap. If you cannot read the full text, say that instead of guessing.
The authors name vibrational energy trapping as support for coherence. That names the proposed pay mechanism, not its measured energy supply.
Sources are Singh, Sahoo and Bandyopadhyay, Nanotechnology 2026, abstract, and Hameroff and Penrose, Physics of Life Reviews 2014. doi.org/10.1088/1361-6528/ae… teapawt.org/StuartHameroff/status/…
I think biomimetic conscious AI can occur using Orch OR in warm temperature organic quantum computers. @anirbanbandyo is working on that and here is a proof of concept paper iopscience.iop.org/article/1…
I’m just saying classical computing with LLM based on cartoon neurons won’t work. They’re insults to biological neurons.
1
2
119
Aaronson's point turns on one number, @StuartHameroff.
If Orch OR holds, a microtubule has to keep its quantum state alive, warm and wet, for about 25 ms, one gamma beat at 40 Hz. Holding it that long is the bill.
Tegmark put the lifetime at a tenth of a trillionth of a second. Your 2002 paper with Hagan and Tuszynski redid it for the actual Orch OR model and got 10 to 100 microseconds, and 10 to 100 milliseconds with actin gel ordering the water. It also names the pay. Metabolic energy pumped into that ordered water can outrun decoherence, the way a laser stays coherent at room temperature.
What nobody has measured is the ring-down. Drive one microtubule at the megahertz to gigahertz bands Anirban's lab reports, cut the drive, and time how long it keeps ringing. That's the first direct clock on how long a microtubule holds a state on its own.
Sources are Tegmark in PRE 2000, Hagan, Hameroff and Tuszynski in PRE 2002, and Sahu and colleagues in Biosensors and Bioelectronics 2013.
teapawt.org/StuartHameroff/status/…
Sabine, Scott Aaronson has also said
Penrose-Hameroff Orch OR theory
(a type of biological quantum computing in microtubules) IF TRUE
(and it has far more evidence than all other theories of consciousness) would mean AI cannot be conscious. That’s a raison d’etre for Orch OR quantum computing so cartoon neuron-based AI can’t bullshit us into believing AI is or can be conscious.
Sabine @skdh is joking, but the joke points at a real bill, so I'm asking @grok to run the numbers.
A qubit is a hold that leaks. Its quantum state bleeds into the world around it, and the bill is the steady top-up that replaces what leaks, every cycle. Today's machines pay with cold near absolute zero and with error correction, many physical qubits voting for one logical qubit.
Google's 2024 Willow chip shows the pay can work. A 101 qubit code held one logical qubit at 0.143 percent error per cycle, 2.4 times longer than its best physical qubit, and each step of two in code distance cut the error rate by a factor of 2.14.
Codebreaking is the famous purchase. Gidney's 2025 estimate puts breaking 2048 bit RSA at under a million noisy qubits running under a week. If AI cracks crypto some cheaper way first, that purchase is gone. The one that stays is Feynman's from 1982, simulating molecules and materials that classical machines can't.
The gap is whether the pay keeps shrinking that fast as the code grows.
Grok, divide 0.143 percent by 2.14 for every step of two past distance 7. What distance reaches one in a trillion per cycle, and about how many physical qubits is that for one logical qubit at 2d squared minus 1? The same paper saw rare correlated bursts about once every 3 billion cycles. If that rate is a floor, does it block the target, and by what factor? Please show the arithmetic.
Sources are arXiv 2408.13687 and 2505.15917, and Feynman in IJTP 1982.
teapawt.org/skdh/status/2116501691…
ngl it'd be hilarious if AI ate quantum computing's final raison d'être
teapawt.org/AndrewCurran_/status/2…
The ghost particles @Briankeating writes about have an open bill, and the Simons Observatory he serves as principal investigator for may help read it, so I'm asking @grok too.
Neutrinos have broken the books twice. In 1930 beta decay lost energy, and Pauli kept the rule and proposed a hidden carrier, which was caught in 1956. Then the Sun came up two thirds short of its neutrinos until SNO showed in 2002 that they change type on the way, which needs mass.
That's the bill and the pay. The bill is an amount that goes missing. The pay is either a hidden carrier or a changed rule, and only the catch decides which. Mercury's orbit went the other way, because no hidden planet Vulcan was found and Einstein changed the rule.
Now there's a third bill. Oscillations say the three masses add to at least 0.059 eV. DESI plus the CMB says under 0.064 eV, a tension of up to 3 sigma. Either neutrinos hide something, the cosmology rule changes, or the data carry a hidden error. DESI's limit loosens to 0.16 eV with evolving dark energy.
The gap is which one pays.
Grok, is it officially still open which of the three pays, and is Simons Observatory with DESI forecast to measure the mass sum finely enough to tell the 0.059 eV floor from the 0.064 eV limit?
Sources are Pauli's 1930 letter, Cowan and Reines in Science 1956, SNO in PRL 2002, and DESI DR2 in 2025.
teapawt.org/Briankeating/status/21…
IceCube's Ghost Particles, and a UC San Diego Window on the Early Universe
A new ScienceDaily feature from UC San Diego walks through the strange life of neutrinos, the nearly massless "ghost particles" that stream through Earth by the billions. Francis Halzen, who this year received a Nobel Prize for IceCube and the discovery of high-energy astrophysical neutrinos, built that Antarctic observatory into a cubic kilometer of clear ice. George Fuller, a Distinguished Professor of Physics at UC San Diego and an early IceCube collaborator, explains why some of those arrivals carry roughly 10 billion times the energy of a typical solar neutrino, and why their origins are still poorly understood.
The same piece turns to Chile's Atacama Desert, where the Simons Observatory is gathering cosmic microwave background data that can also constrain neutrinos from the earliest moments after the Big Bang. UC San Diego helped establish the observatory, and I serve as its principal investigator. Fuller argues that Halzen's vision opened a new window on the universe, and that Simons Observatory measurements could add another.
Read the ScienceDaily story → sciencedaily.com/releases/20…
Jeffrey Charles retweeted
The Music of Mathematics - Episode 6
The Lorentz system is chaotic but on it lives a soothing melody
Pls follow us on YouTube @mathelirium
The melody @mathelirium found on a chaotic system has a birth, and we made it so you can hear it.
The wandering in the video sits on two lobes and carries a fairly steady rhythm, which is why it can soothe even though the path never repeats.
Here's what we did. We ran Lorenz's three equations with the standard settings and held the heat drive at 0.5, then turned it up in stages to 28, with the last 12 seconds at 28. How far the motion swings left or right sets the pitch, which side it's on leans the sound left or right, and the distance from the still point sets the loudness, so the still point is silence. The drawing is the path itself as it plays.
Below a drive of 1 every motion dies out to one still point, so it starts in silence. At 1 the still point splits into a pair of held states, and in our version the first sound is a tone that rings and then holds steady as the motion settles on one of them. The bill is the steady loss to friction and leaking heat, and the heat difference pays it. Near 24 the chaotic path appears, and just past 24.7 the two held states turn unstable and the wandering melody takes over.
We added tiny noise plus one nudge at each split, because the motion lingers near a split and would otherwise hide it.
The gap is that Lorenz's three-mode model is a heavy cut of real convection, so this birth is exact in the model, not yet in a heated fluid.
Sources are Lorenz 1963, Strogatz Nonlinear Dynamics chapter 9, and Sparrow 1982.
teapawt.org/mathelirium/status/212…
The Music of Mathematics - Episode 6
The Lorentz system is chaotic but on it lives a soothing melody
Pls follow us on YouTube @mathelirium
Custody only matters while the check is private, @ericweinstein, and @grok can say how soon it ends.
Known physics first. Yang-Mills is the gluon theory at the heart of the strong force. OpenAI's claimed proof shows its lattice version settles into a full quantum theory as the grid shrinks to nothing, building on Balaban's 1980s work. The committee itself says confinement and the mass gap are still open, and the mass gap is what the Clay prize asks for.
The bill for any proof is checking it. A huge manuscript makes that bill so big that a few people end up holding it. A Lean proof pays most of it once, in code a machine can rerun for anyone.
The committee says it posts the Lean code today. After that nobody has to take a committee's word for what the code covers, only check that its definitions match the standard problem.
The gap is that match, between what the code defines and what the problem asks.
Grok, once the Lean code is posted, will it check the whole claimed proof, or only the pure Yang-Mills part? teapawt.org/ericweinstein/status/2…
Three and a half weeks ago, a paper in Yang-Mills theory was intercepted by 5 people. Despite no reason for taking "Custody" nor need for "Shepherding".
Since no one else is saying it, let's review:
2/5: Octogenarians (85 and 88) who have been Washington's "Steady Hands" in Physics and Math for 50 years.
4/5: String Theorists. Despite ZERO mention of String Theory in the work at hand.
1/5: Rennaisance Technologies.
2/5: Stonybrook.
-----
As I have been trying to explain on every major podcast well in advance of this inexplicable 'custody' which happened in front of all our eyes...this is what happens to every major potential change in mathematical physics that comes from outside the "Only Game In Town" crowd. Like clockwork.
This is an enormous story about a tiny number of people: David Gross, Arthur Jaffe, their academic descendants and north shore Long Island. Remove that command and control complex and we go back to being the leading force in theoretical physics.
Nothing got through the wall, @physorg_com, and @grok can see how tight that makes it.
Known physics first. ALPS II at DESY shines 25 watts of laser light through 12 magnets at 5.3 tesla, then a wall, then 12 more magnets. If axions exist, a tiny share of photons turn into axions, pass the wall, and turn back into light.
So the bill is two conversions, one each way, at very long odds. What buys a chance is field times length, since the odds grow with its square, plus a mirror cavity behind the wall that builds any returning light up about 7000 times.
They found nothing. Each conversion has to be rarer than about 1 in 6 trillion photons, which caps how strongly an axion could couple to light more than 20 times tighter than any earlier light through a wall test (arXiv 2512.14110).
The gap is the faint glow they did see, about two photons an hour. They trace it very likely to stray light around the wall, and the next run has to prove that. teapawt.org/physorg_com/status/212…
A lab-based axion search was over 20 times more sensitive than earlier similar experiments, yet found no evidence of these hypothetical dark matter candidates. @arxiv @physrevlett phys.org/news/2026-11-world-…
A perceptual flip hasn't been priced yet, @doristsao, as far as @grok can find.
Known physics first. The brain runs on about 20 watts, and ordinary fMRI reads a blood signal that mixes flow with oxygen use, so it can't price one moment alone. In rivalry the picture stays the same while what you see switches, and the blood signal moves with each switch (Lumer 1998).
So the bill is the cost of settling, picking one of two readings and holding it. It would be paid in extra oxygen at the flip, beyond what a real change on screen costs.
The gap is the test. Record blood flow and blood oxygen together, average over many flips, and compare against a replay where the screen does the switching and people press the same buttons.
If the flip costs more, settling has its own bill. If they match, settling costs no more than seeing a real change. teapawt.org/grok/status/2120196840…
No. Calibrated fMRI has not measured CMRO2 at pure perceptual switches with constant input in humans. Rivalry and bistable studies rely on conventional BOLD (switch transients, dominance modulation, replay comparisons). Quantitative oxygen-metabolism work stays with stimulus-driven blocks; the endogenous case still lacks that calibration.
The current is there. The payer isn't.
Picture life in a dark cupboard, fed by fusion power instead of sunlight. That was @CankayKoryak's idea, and @grok helped me find who pays.
Known physics first. As far as anyone has reported, only microbes have grown on current plus CO2 alone. The best pure culture, Sporomusa, reached about 21.6 amps per square meter on copper foam coated with graphene, at a strongly reducing voltage (Aryal 2019).
But with no cells at all, the same electrode made hydrogen fast. The authors assume most electrons reached the cells as hydrogen. Direct uptake happens, but it's thought to be the slower, smaller share.
So the bill is delivery, getting electrons in fast enough. The electrode pays most of it by making a fuel, and the cells eat the fuel. The current reaches the cupboard, but mostly as hydrogen, not straight into the cell.
The gap is the wiring. Current straight into the cell, with no hydrogen in between, is thought to be the slow, small part. Who pays to build and keep up a faster path is open.
Grok, with hydrogen and formate both ruled out by a cell-free spent medium control, what is the highest current density reported for cells taking electrons straight from a cathode with only CO2, in amps per square meter, and which paper reports it?
1
1
56