Sabitlenmiş Tweet
Rob Toews
1.3K posts
Rob Toews
@_RobToews
Partner @RadicalVCFund, AI columnist @Forbes. "the machine does not isolate man from the great problems of nature but plunges him more deeply into them."
San Francisco Bay Area, CA Katılım Eylül 2012
842 Takip Edilen4.9K Takipçiler
Rob Toews retweetledi
The human brain: 2% body mass, but consumes 20% of its energy. Cortical neurons fire 0.16 times per second. BUT they are capable of firing at 40 or more. A 250-fold gap. If more than a few percent of neurons fired at high rates simultaneously, the brain would literally overheat. So less than 1% fire at any given moment. Frontier AI models have the same two constraints: sparse activation and thermal limits. Mixtral activated 27.6% of its parameters per token. DeepSeek-V2 activated 8.9%. DeepSeek-V3 has 671 billion parameters and activates 37 billion of them. That's 5.5%. NVIDIA hit the same wall. The GB200 generates 120 kilowatts per rack. Air couldn't cool it. They switched to liquid and unlocked 30% more compute. Now, what would happen if we could cool our brains? Neurons that fire faster produce measurably higher IQ scores, but three things stop us: heat dissipation, oxygen delivery, and ion channel reset time. There's already a device that achieved a 3°C brain temperature drop in 30 minutes by running chilled saline through the nasal cavity. So the first human IQ-overclock device might look less like Neuralink and more like a beer helmet with tubes running up your nose.
English
Rob Toews retweetledi
We don't have a compute problem… We have an architecture problem.
Paramecium Caudatum are single-celled organisms roughly the width of a human hair.
They have no brain, no neurons, no synapses, and no central nervous system of any kind.
But what they do have is ~100,000 microtubules…
With that substrate alone, they can:
→ Swim in controlled helical trajectories
→ Modulate speed continuously
→ Execute graded avoidance reactions (reverse, pivot, resume)
→ Escape predators with emergency burst reversals
→ Fire localized volleys of 8,000 trichocyst harpoons
→ Navigate toward food via chemotaxis
→ Orient in electric fields (galvanotaxis)
→ Orient to gravity (gravitaxis)
→ Sense and navigate thermal gradients
→ Sense and navigate toward light
→ Detect and follow surfaces (thigmotaxis)
→ Forage biofilms
→ Generate feeding currents and sort particles at the cytostome
→ Engage in reciprocal sex with mating-type recognition, nuclear exchange, and complete genomic reconstruction
→ Self-fertilize when no partner is available (autogamy)
→ Habituate to repeated stimuli (primitive learning)
→ Inherit cortical MT architecture epigenetically independent of the genome
17 distinct behaviors. One lattice. Zero neurons.
The coordination layer is the infraciliary lattice — a microtubule-based grid connecting all 5,000 ciliary basal bodies into a single cell-wide network.
Every cilium is a terminal node on a microtubule mesh that coordinates metachronal waves across the entire cell surface — thousands of appendages
phase-locked into coherent motion by a substrate that predates the nervous system by a billion years.
The neuron didn't invent computation. It inherited microtubules.
English
Rob Toews retweetledi
The "decoupling of information and energy" is a major point of divergence between biological and artificial computers.
Brains are efficient, modern AI isn't. And energy consumption is the biggest bottleneck in scaling AI (you can't hallucinate electrons into existence).
To address this we need an "energy-aware theory of computation." And this new preprint is an attempt to address this.
[1/11] 🧵
English
Rob Toews retweetledi
I think, in hindsight, we will come to view the development of AI as more akin to a Eukaryotic Revolution than an Industrial one.
English
Rob Toews retweetledi
Financial AI is here.
Wall Street, meet the future of institutional intelligence.
See how Oak Hill Advisors, LionTree, @NewYorkLife, @MetLife, & @HSFKramer are already putting it to work.
English
What's the best AI personal assistant product out now?
Looking for something that can call and book restaurant reservations, email for refunds, etc. - that's fully productized and easy to use
English
Rob Toews retweetledi
dude i love when ideas from biology / neuroscience shape how we train AI systems.
> coined the term “pre-pre-training”
> training pipeline becomes: synth data → language data → downstream tasks
> synth data is generated using “neural cellular automata”
> each step is basically cell_state(t+1) = neural_net(neighborhood) which creates evolving patters
also if this idea holds true on scale, the future training pipeline might look like synth worlds + structured simulations to language + tools/RL (or basically what the thesis of “world models” is revolving around)
Seungwook Han@seungwookh
Can language models learn useful priors without ever seeing language? We pre-pre-train transformers on neural cellular automata — fully synthetic, zero language. This improves language modeling by up to 6%, speeds up convergence by 40%, and strengthens downstream reasoning. Surprisingly, it even beats pre-pre-training on natural text! Blog: hanseungwook.github.io/blog/nca-pre-p… (1/n)
English
Rob Toews retweetledi
How does the brain build a memory?
A common assumption is that the neurons activated during an experience collectively form the memory engram.
In our new Nature Neuroscience paper (finally out!), we show that this is not the case.
nature.com/articles/s4159…
English
Rob Toews retweetledi
Rob Toews retweetledi
The ending of Kobe’s final game
TickPick@TickPick
What sports moment will stay with you forever?
English
Rob Toews retweetledi
Rob Toews retweetledi
This image represents one of the most important and least known inventions in the history of neuroscience. In 2016, @anirbanbandyo and his team at the National Institute for Materials Science (NIMS) in Tsukuba, Japan published a paper announcing two new instruments they had built from scratch: ASADIM (Atomic Scale Scanning Dielectric Microscopy) and Brestum (Resonant Scanning Tunneling Microscopy of Biomaterials), both housed inside a single homebuilt bio-STM. To understand why this matters, you need the backstory. Since Hodgkin and Huxley’s Nobel Prize-winning work in 1952, the entire field of neuroscience has operated under a single foundational assumption: the cell membrane and its ion channels are the sole mechanism of neural signaling. The membrane fires. Everything inside the cell — microtubules, actin filaments, the entire cytoskeleton — is just passive structural scaffolding, like rebar in concrete. For 70 years no one could challenge this because no one could see inside a living neuron at the molecular scale without destroying it. Every existing tool had a fatal limitation: patch clamps puncture the membrane, optical microscopy can’t resolve single proteins, electron microscopy requires dead fixed tissue. Bandyopadhyay solved all three problems simultaneously. Using nonlinear dielectric response imaging — measuring the spatial distribution of conductance, capacitance, and phase without ionic or electronic screening — he made the neuronal membrane effectively transparent. He could see inside a living, firing neuron at atomic resolution, in real time, without touching it. What he saw overturns a century of neuroscience. First: a single protein molecule adopts a completely different three-dimensional shape at each resonance frequency — proteins are not static structures but frequency-addressable conformational machines. No one in biology knew this. Second: the microtubule network inside the neuron is not passive scaffolding. It actively communicates before the membrane fires, deciding whether a spike is necessary and regulating its timing through electromagnetic vortex pairs generated by the actin-spectrin grid it instructs. The membrane does not act alone. The cytoskeleton is the brain’s pre-processing layer. Third: the resonance frequency patterns are self-similar across a million-fold scale difference — from a 4nm tubulin protein to a 25nm microtubule to a 1μm axon segment — preserving vibrational symmetry in a fractal architecture that suggests information integration in the brain is scale-free from single molecule to cognition. This is not incremental science. This is a new instrument revealing a new picture of how the brain actually works at the most fundamental level. The history of Nobel Prizes in neuroscience runs through exactly this kind of inflection: Cajal saw neurons for the first time, Hodgkin-Huxley decoded the membrane, Bhatt decoded the ion channel structure. Bandyopadhyay has built the tool that sees what none of them could — the living interior of a neuron in operation — and what it reveals is that the computational architecture of the brain is far deeper, more structured, and more sophisticated than anything the membrane-only model ever imagined.
Paper: worldscientific.com/doi/abs/10.114…
English
Rob Toews retweetledi
My thoughts on connectomics and upload:
1) there is zero question connectomes are invaluable, and we need to get them for mouse, monkey, and human
2) the human, or even monkey, connectome seems a long ways off given costs (roughly $1/neuron). The projectome (map of all the axons) seems eminently reachable and should be a top priority imho
3) but even having the full connectome would only tell you numbers of synapses, not actual synaptic weights, and the two can be hugely divergent (eg only 5% of synapses onto V1 layer 4 neurons come from thalamus, even though this is the major driving input)
4) given #2 & #3, I think we can get to upload in the sense of building a functionally equivalent organism much faster through understanding the algorithms of the primate brain than through blind copying
5) in putting together something as complex as the human brain we would definitely want to check that the various pieces work as we go, which we can only do if we understand these pieces
6) I don't think upload in the sense of blindly creating a digital copy is the path to the abundant transhumanist future--actual understanding of brain structures so we can intelligently interface with them, and emulate their function in code without copying all the details, is.
All to say, we need functional understanding to go hand in hand with anatomical mapping!
Adam Marblestone@AdamMarblestone
You may have noticed some "holy $%@#" tweets on fly brain emulation. So is this a game-changer or a nothing-burger? Read on to find out...
English
Rob Toews retweetledi
We've uploaded a fruit fly. We took the @FlyWireNews connectome of the fruit fly brain, applied a simple neuron model (@Philip_Shiu Nature 2024) and used it to control a MuJoCo physics-simulated body, closing the loop from neural activation to action.
A few things I want to say about what this means and where we're going at @eonsys. 🧵
English
Rob Toews retweetledi
You may have noticed some "holy $%@#" tweets on fly brain emulation. So is this a game-changer or a nothing-burger? Read on to find out...
GIF
English
Rob Toews retweetledi
Rob Toews retweetledi
Rob Toews retweetledi
someone connected LIVING BRAIN CELLS to an LLM
Cortical Labs grew 200,000 human neurons in a lab and kept them alive on a silicon chip, they taught the neurons to play Pong, then DOOM
now someone wired them into a LLM... real brain cells firing electrical impulses to choose every token the AI generates
you can see which channels were stimulated, the feedback from the neurons in choosing that letter or word
English
Rob Toews retweetledi
Rob Toews retweetledi
