What If We Stopped Using GPUs? | YC Paper Club
In a Nutshell
The core message is that GPUs are hitting fundamental efficiency walls for scaling AI, and radically different computing paradigms—optical, neuromorphic, and biological—are being developed to break through them. Key takeaways include optical computing's ability to perform linear operations at near-zero energy cost, neuromorphic chips mimicking brain efficiency at ~20 watts, and experimental systems where living neurons are trained to play games like Doom via reinforcement learning. The overarching argument is that future AI progress will require co-designing hardware and algorithms around physics and biology rather than continuing to scale transformer-based digital systems.
These notes were generated by AI and may contain inaccuracies.
The Alternative Computing Club began with a dinner conversation about what single question one would ask super-intelligent aliens. The discussion shifted to energy calculations involving Dyson spheres and nuclear power, but the most intriguing aspect was how aliens might calculate floating-point operations.
The speaker has worked in deep learning since approximately 2012, starting with the AlexNet era. Convolutional Neural Networks dominated until 2019-2020. At Focal Systems, the team won a competition and received $125,000 from Nvidia, meeting Jensen Huang who showed no interest in artificial intelligence until 2020. Prior to this, cryptocurrencies dominated Silicon Valley attention.
Before 2020, Nvidia focused on FLOPS (floating-point operations per second), specifically computational power per joule, optimized for convolutional neural networks that weren't memory or bandwidth limited. The focus shifted from Bitcoin to Ethereum mining.
After 2020, GPT-3 and GPT-2 technology changed priorities. The Ampere architecture was released, marking a shift from computational efficiency toward memory capacity and bandwidth requirements. Memory capacity and bandwidth grew exponentially, possibly due to attention mechanisms being n-squared. Computational efficiency was deprioritized, with the attitude that producing FLOPs was acceptable even if it meant burning oceans.
Gigaflop per joule rates have declined significantly over time with no substantial improvement in the past two years. The speaker completed a PhD studying brain function and argues the cost per gigaflop must decrease dramatically to match human energy consumption levels. The human brain consumes approximately 20 watts while performing extraordinary computational tasks.
Current chip divisions separate training and inference chips, but both remain heavily focused on transformer architecture, requiring substantial SRAM. The brain operates entirely feed-forward without backpropagation, which would require neurons to fire signals forward then backward, necessitating complete system shutdowns.
The brain operates as a purely feed-forward system. Deep learning faces the weight transport problem where connections must go forward then backward through paired synapses with exactly equal weights. Adding an activation element where both sides see the same signals enables approximate solutions, creating cyclic graphs rather than true backpropagation.
Chemical reactions leave residues after activation, along with timing considerations. The brain contains massive inhibition throughout cortical columns, which are recurring clusters enabling independent learning within each cluster.
Current data center communications use light for transmission, then convert signals back to electrons for routing, then reconvert to light, requiring DAC and ADC conversions repeatedly. The speaker questions why all computation isn't done optically, which would be instantaneous and consume no energy.
Neuromorphic computing simulates brain function using analog circuits. However, backpropagation cannot work effectively in optical or neuromorphic substrates because storing bits in light is extremely expensive. Optical computing offers extremely low cost per gigaflop, with calculations being nearly free, though data storage remains difficult. Non-differentiability and activation problems present additional challenges.
Neural computing using copper wires might achieve similar bandwidth characteristics but with different properties. The speaker spent their PhD working on alternative optimizers, exploring methods predating backpropagation.
The speaker tested every optimization method from a Zero-to-Hero book during the first six months of PhD research using a billion-parameter LSTM model for next-token prediction. SPSA (Simultaneous Perturbation Stochastic Approximation) emerged as the most effective update rule, originally developed by James Ball and his students.
SPSA uses two forward passes with random perturbations of epsilon magnitude, applying finite differences to modify updates. Zero gradients result in no changes, while large finite differences produce significant modifications. This enables efficient optimization of non-differentiable, non-Lipschitzian functions.
LSTM networks were designed to enable gradient flow, making them less suitable for zero-order methods. Transformer networks were designed for backpropagation and GPU parallelism. The speaker developed "Soma," a mixture-of-experts approach using cortical column concepts, processing Common Crawl data with TF-IDF and SVD, then distributing Parquet files across global GPU clusters.
Each GPU trains a small expert model with 41,000 parameters. During inference, a router combines outputs similar to mixture-of-experts but across entire models rather than single layers.
Zero-order gradient estimation error worsens linearly with model size, making large model training computationally infeasible. Models with 10 billion parameters would require enormous adjustments, exceeding forward and backward pass costs. Model segmentation limits gradient noise to parameter group sizes, requiring only 64 adjustments per step.
Scaling laws apply across three dimensions: model size, batch size for gradient noise reduction, and increased segmentation. Forward passes remain expensive regardless of optimization method, though optical computing could make forward passes nearly free through light beams.
The speaker expressed excitement about the seminar, noting that if none of the speakers succeed, their PhD work would be meaningless. Three speakers were introduced:
- Ilker: PhD from EPFL, worked in visual computing for AI for five to six years
- Alok: Postdoctoral researcher at Stanford, PhD in electrical engineering, semiconductor and nano-optics experience, manages $1.5 billion venture fund
- Shawn: Recently graduated from NYU with neuroscience master's degree, CEO of Parasma (named after Dota game element)
Ilker began by noting that data communications extensively use optics due to photons having approximately 10,000 times less loss and 10,000 times greater bandwidth compared to electrons. All data between AI data center racks and 99% of intercontinental data transfer uses fiber optics.
Historically, electronic processors used 32-bit and 64-bit representations, then converted to 2-4 levels for optical transmission. AI emergence enables lower bit precision for more operations, while optical communications support higher bit depths, creating convergence.
Optical fibers enable high parallelism since photons don't interact, allowing multiple light beams to coexist in shared space. Electrical circuits require separate paths for each signal.
Electronic matrix-vector multiplication charges wires, processes through transistors, then discharges, consuming energy at each step with energy scaling as n-squared for n-dimensional inputs and outputs.
Optical approaches modify input information on n modulators, allow propagation through designed weight masks where weights passively modify light beams without additional energy consumption. Results are physically summed at detectors. Energy consumption scales with n rather than n-squared.
Despite companies like Lightmatter achieving competitive energy efficiency with PCIe photonic boards, photonics represents only a small portion of total energy budgets. Major challenges include:
- High costs of digital-to-analog-to-optical and optical-to-analog-to-digital conversions
- Energy required for device programming, calibration, and accuracy maintenance
- Difficulty implementing nonlinear activation functions, which require additional mechanisms beyond passive linear operations
Nonlinear activations were demonstrated using intense light pulses in multi-mode fibers through light-matter interactions, adding complexity layers.
Ilker presented NeurIPS research from EPFL in collaboration with Google, applying diffusion-based models to image generation using optical diffusion as an inference unit. Diffusion models map random distributions to desired data distributions through trained neural network steps.
The approach eliminates redundancy of repeatedly applying the same neural network on general devices, potentially reaching 1000x redundancy for single image generation. Instead, passive devices are designed where light propagation physics mimics the diffusion process.
The system requires knowing physics principles and methods to control them. Noisy samples are input with the goal of obtaining expected noise thresholds, creating filtered images through light propagation. Multiple applications gradually eliminate noise to generate clear images.
The system uses transparency layers that modify light phase, deflecting beams so only desired information transmits. Programming difficulties were addressed by dividing 1000 time steps into 100-step subunits with constant parameters within each subunit.
Testing on small datasets like MNIST digits and fashion showed the optical device behaving as a diffusion model, generating images similar to training distributions with low FID scores indicating quality and diversity.
The system exhibits power-law scaling behavior when varying diffraction layer counts or pixel numbers per layer, similar to digital neural networks. Adding parameters predictably improves image quality while energy consumption advantages increase with more transactions.
Comparing the optical system to similarly-performing digital neural networks on commercial GPUs showed significant energy efficiency advantages for single image generation. The research assumed fixed negative weights manufactured with extreme precision after training, remaining permanent.
In the pilot application, a simpler approach was taken using a spatial light modulator commonly found in light projectors. By connecting it to a mirror, researchers demonstrated the possibility of creating a deep neural network on a single device. Camera images and input laser distributions enable reverse diffusion in situ to update weights, eliminating the need for repeated precision manufacturing stages and retry attempts.
This provided a useful means of designing system prototypes, though future efficient optical computing units will require producing layers and putting them in place. Published studies indicate this manufacturing approach is readily available. The opportunity in optical computing lies in designing hardware and algorithms together. The generation algorithm was manually adjusted to use the same weights repeatedly to take advantage of the field of optics.
The next step toward making this technology practical is proving the existence of an integrated system at very large scale, such as a model with a billion coefficients, to ensure power or response time advantages are maintained when scaling up. Initial hypotheses have been proven indicating that when more parameters are added, performance is expected to improve as in digital neural networks.
At Nvidia Light Matters, the first AI accelerator dedicated to silicon photonics technology, challenges included the DAC and ADC subsystem. Photonics is essentially a discharge engine for most tensor operations. The accuracy of the ADC and DAC ultimately affects prediction accuracy over time. A project later noticed was the reason Light Matter shifted to HBM interconnects and stopped working on the multispectral computing branch of its research.
When there are a large number of parameters or a large amount of data compared to the number of parameters that need to be sent to the optical system, the disadvantages usually outweigh the advantages. The first approach for this type of device should not use general computer hardware such as CPUs, but rather ASIC-specific application approaches where bit depth, resolution, and data transfer rate can be reduced to the lowest possible level.
Regarding reverse diffusion implementation, a forward pass can be performed where the same process is repeated several times. The outer loop of the T spread is performed once, then backward movement is needed. Reverse propagation occurs in digital electronics, which is the biggest obstacle at present. Work has been done on reverse propagation methods, but the current approach involves calibrating a fully digital twin with reverse propagation occurring in digital electronics.
Diffusion is a recurrent neural network forced at each step. If that is removed, there will be no levels L1, L2, L10, and up to L10, allowing treatment as a fully recurrent neural network (RNN). This has not been tried, but it represents an interesting future direction because the utility of optics depends on how small the interface to electronics or the digital environment can be. If information can be repeated in an optical or analog system for longer periods or longer time steps, the real advantage could reach hundreds or even more beyond the seven times demonstrated.
When applying this to much larger weights, the actual physical challenge in manufacturing relevant components involves the same technologies used in electronics. Researchers have shown that using etching and sedimentation (simple chemical processes), the manufacturing process can be completed. Light has a longer wavelength compared to electrons, so the unit size is slightly larger, but these are manageable challenges.
One limitation is nonlinearity because obtaining a large number of parameters benefits from creating deep networks in the digital case, requiring nonlinearity. Nonlinearity returns to the modulator, requiring inexpensive ways to implement this nonlinearity or undertaking a massive process to find a useful model. The challenge is creating a very deep network to increase the number of parameters.
Electrons are much smaller than typical wavelengths of visible light. Optical chips are limited to this size. The theoretical minimum unit size in optical computing is being explored in different directions. One approach uses the same unit size in the micrometer range compared to nanometers in electronic transistors, but employs what is called an optical comb. This involves hundreds of different wavelengths with ways to encode different weights on each wavelength, enabling multiplexing on a wavelength dimension rather than in space. The feature needs to be bigger to interact with visible light.
For optical-based storage, traditional computing has made great progress over 70 or 80 years. Optical-based storage would enable completely bypassing the analog-to-digital converter (ADC). CDs and DVDs are excellent storage media for archiving or long-term storage, working very efficiently with many different ways to achieve that. The challenge lies in writing and reading data in the visual field quickly and at low cost.
Historically, attempts include holographic memory and photoelectric materials, but none have been as cheap and accessible as electronics. This is one reason electronics remains the primary means of computing. Reliance on fixed, application-specific weights exists because rewriting data or maintaining weights in the optical field is extremely expensive. Another line of research involves phase change materials used in electronics and optics, which may provide a solution, though none are as mature and easy to use as NAND memory or electronic memory.
Polarization filters, roughly triangular in relationship, could potentially be combined to achieve the same result as the Fourier transform in constructing any signal using triangles. Polarization is another dimension; light has different degrees of freedom. Intensity and phase are among these dimensions, but polarization is another dimension. Any manipulation will show up in the amplitude domain usually used with sensors or detected as a nonlinear property. This is a good idea, but depth must be considered because modifying polarization only gets a response, and for a deep neural network, it must work on every layer. It is not clear if the nonlinear property can be preserved after filtering.
Alok, co-founder and partner of venture capital firm Standard managing $1.5 billion in assets, has three degrees in electrical engineering: a bachelor's from the University of Texas at Austin, and a master's and doctorate from Stanford University. Research and practical experience focused on semiconductors, photonics, and nanomaterials.
The brain is a remarkable achievement in evolutionary engineering capable of language, thinking, and creativity simultaneously. It is multimedia, works in real time, and can see, hear, and act with visual, auditory, and physical embodiment. Response times reach hundreds of milliseconds. The brain operates at only 20 watts, equivalent to an iPhone charger, and continues learning throughout life with no contextual windows, adapting and learning new things including things evolution has not prepared it for.
A 16-year-old shown a stop sign once or twice will recognize stop signs with an accuracy of nines, whereas a self-driving car requires many different angles, images, and shadows for accurate recognition.
In the late 1980s, advances in neurobiology, understanding of brain function, and electronics enabled large-scale manufacturing of electronic chips and circuits. Carver Mead of Caltech coined the term formal neural computing. His team accomplished pioneering work in understanding how the brain works and applying this to electronic circuit models.
Formal neural computing is an attempt to utilize principles of brain organization in the design of electronic circuits, chips, and other computing tools. General principles about brain function include the relationship between memory and calculation, where connections have memory, have weights, and play a role in the calculation itself rather than simply transmitting signals.
Communication resembling state transfer occurs through events called impulses in the brain. It is a continuous device receiving sensory inputs where impulses are generated and propagated. Nerve impulses are the way calculations are performed. The brain is more like mixed signal containing digital and analog parts integrating uniquely.
With experience and learning, both structure and function change as the brain constantly reorganizes itself. When chips are made, they are stable. The brain's ability to adapt logically and physically over time in response to experiences is a fascinating phenomenon.
A simplified model of a nerve cell emitting impulses resembles a leaking capacitor. When charge accumulates above a certain threshold, it releases a digital signal, an electrical pulse. Pulses arrive and charge accumulates on the neuron membrane, then leaks out and fades over time. If enough pulses arrive simultaneously so charge accumulates above threshold, the neuron emits a pulse. If not enough arrive, it relaxes to normal state.
The time relationship between pulses is very important as a continuous variable, and pulse amplitude is extremely important. This strange result happens repeatedly in the brain.
The neural network is the most well-known concept inspired by the brain. Modern artificial intelligence and modern neural networks contain concepts entirely inspired by neuroscience including large networks of interconnected neurons, demonstrating knowledge as changes in connection strength, and encoding information as patterns across many neurons.
Digital electronics are ingenious with excellent manufacturing, scaling, and parallelization capabilities. The move toward native applications and focus on high-level concepts applied in digital silicon led to real innovation. Modern AI concepts like backpropagation, the switched attention mechanism, and large-scale training on digital devices do not match what exists in the brain.
The brain is the ideal model for common design of devices where hardware and software are identical and evolved together uniquely. When applying parts to other computer platforms or physics types, concepts may apply at high level but what is inherent in the form may be more appropriate to take to higher level and improve.
In 2026, the field remains in research and development phase trying to utilize brain information, integrate with practical applications, and achieve tangible results. Three main ideas researchers are exploring: solving observed problems by taking advantage of biological insights, creating new tools with unique properties that can be exploited by combining with biological information, and leveraging the brain's exceptional energy efficiency consuming 20 watts.
Researchers are exploring analogical ideas even in training and reasoning. DD Matrix has done work integrating memory and computing entirely digitally, presenting at the Hot Chips conference. IBM has also done work in this area, though entirely digital. The broad definition of formal neural architecture includes integration of memory and computing.
On the edge, chips and devices that sense, interact, and adapt in real time draw inspiration from continuous sensing the brain performs. Pulsed neural networks are being explored where networks can be trained on impulses, devices can learn on the same device, networks can improve themselves based on what they understand, and event-based computation can occur as the brain works. Intel has done significant work with a chip from their R&D lab and a drone that can fly autonomously.
New tools include photonics and advanced electronic devices developed in semiconductor R&D laboratories including resistive memory and memristors. Circuit ideas such as coupled oscillators exhibit interesting new physics and properties with analog and nonlinear characteristics. The question is whether these systems can serve as technological foundation for building models around specific physical behavior to jointly design something with interesting outcomes or properties.
Using the broad definition of neural computing, many brain-inspired approaches remain in early stages. Integration of memory and computing has several startups working entirely digitally, with DMatrix being applicable now, manufacturable, and with a market able to absorb the technology.
For more ambitious approaches, Naveen Rao is developing pure CMOS coupled oscillators considering large-scale manufacturing requirements, with interesting dynamics and physics derivable from these systems. Innovation revolves around co-encapsulation bringing SRAM closer to the decryption unit in the memory hierarchy to address maximum decryption limits.
The main problem is interconnection capacity being much larger than transistor gate capacity, with majority of energy consumption spent transmitting data over long distances via wires rather than changing individual bit states.
The speaker personally believes that optics is the ideal solution to the problem of transistor limitations.
A question is raised about whether the goal of neuroscience is to develop circuits, analog circuits, optics, or other technologies that surpass the brain and become the brain itself. The questioner uses the analogy of estimating a sine function using a set of lines, noting that the best way to approximate a sine function is to use only a sine function itself.
The question is posed: if we are trying to simulate the brain, should we use the brain itself - specifically nerve cells - and is this our best hope in neuroscience?
The speaker acknowledges the question is interesting but notes there are limitations on the brain. Computers can do certain things much better than the brain can, such as processor speed being much faster than what can be achieved with the brain.
Regarding memory, the speaker notes that retrieving information is more difficult for humans, but computers can provide unlimited storage capacity while humans have biological limitations.
The question is raised about whether we can achieve brain-like capabilities and if we can expand them, noting this is a research direction that has not yet been completed.
A question is asked about joint optimization between software and hardware when setting up next-generation computing systems. Given that iteration cycles for both are shrinking, the questioner asks whether it's more important to iterate on a software model first and then choose the hardware platform, or to find the best hardware platform first and then design the model around it.
The speaker gives the answer that "it depends" and elaborates that it depends on whether you start with a deep understanding of the physical properties of a system. For optical computers, since the limitations are fully understood along with where advantages can be maximized, model structures cannot be blindly applied - the design must work within the system's strengths and weaknesses.
From the opposite perspective, if someone has a unique structure believed to be effective, they would then need to choose the appropriate physical embodiment to bring it to life.
A question is raised about electronic devices being based on silicon while the brain is primarily based on carbon, asking if the lack of research in this area is why brain experiments must be simulated on silicon-powered devices.
The speaker responds that silicon is excellent for performing calculations, and it's more about trying to simulate the system as an electronic circuit rather than doing it on silicon specifically. Over time, materials used in chips have become more exotic and sophisticated, and this trend will continue.
A team was scheduled to present on using diamonds as a substrate, which is carbon. New materials will make a difference, but the speaker won't focus too much on the comparison between silicon and carbon, emphasizing it's more about how materials are used in electronics.
A question is raised about the role of noise and repeatability in computing, noting that conventional silicon computing aims for repeatable computation at the bit level, while the brain does not seek fully repeatable computation at the lowest level.
The speaker notes that determinism in digital computing is extremely interesting, especially when reaching limits of floating-point accuracy. One category of attempts is thermodynamic computing, which goes back to the origins of neural networks, including Jeff Hinton's work on Boltzmann machines.
The idea is whether noise can be used to advantage in a useful way. The tricky part is wanting some noise but not all of it, as noise introduced into the system may distort images.
According to SPSA, as long as noise can be made repeatable and based on the same seed, it is perfectly learnable. The front difference can be made if noise is repeatable just twice, and three times for center difference.
The problem lies in making it reproducible. The speaker spent time with Shawn Druckmann at Stanford University, who notes that making noise repeatable is the real problem - he can't justify how to turn the seed into a vector twice. This is the main reason why zero-order may not be biologically possible.
The speaker introduces work done in collaboration with Cortical Labs, where brain cells were put on electronic chips and algorithms were written to enable them to play Doom. The goal is to explain approaching this as a dynamic systems problem rather than trying to force the brain to perform innate calculations such as matrix multiplication.
The approach is to treat brain cells as an input-output machine where a set of stimuli can produce intelligence allowing for computationally useful output.
Communication with brain cells occurs primarily through electrical stimuli. In Doom, game state variables like ammunition and health, along with screenshots, are encoded into triggers including channel, amplitude, and frequency of stimulation in a two-dimensional electrode array.
Leaky neurons accumulate charge and release electrical impulses over time. These pulses are decoded into specific game actions to obtain useful data from the cells.
Cortical Labs achieved having brain cells play Pong four years prior. Pong is relatively simple compared to Doom, allowing encryption and decryption methods to be programmed directly.
For Pong, the best encryption channels were discovered - stimulating these channels produces more varied outputs that are more detachable for the decoder. The decoder then moves the bat to where it thinks the ball will go.
Doom is much more difficult as a 3D game with multiple enemies in an arena that must be eliminated. The movement space in Doom is much larger than Pong's simple up/down paddle movement, requiring combined movements including sideways movement, turning, attacking, and moving.
With only 59 channels, it's virtually impossible to manually encode and decode the game or map 54 different procedures on 59 channels, as some neurons may be on similar channels producing highly interconnected behaviors.
For the most difficult tasks and extending to GPT-level intelligence, encryption and decryption must be learned from beginning to end. A way to stimulate the brain must be learned, along with a decoder to turn neural responses into useful computational procedures.
A closed-loop PPO (Proximal Policy Improvement) architecture was used, a type of reinforcement learning algorithm. The encoder converts Doom notes into neuronal stimulation, pulses go to the decoder which converts them into in-game actions.
The evaluator in PPO maintains stable cellular feedback. Cells need feedback over time to determine how well or poorly they are performing. The evaluator matches the feedback delivery method needed.
Each step involves taking Doom notes, encrypting them, sending pulses to the decoder, with updates approximately every 2000 steps.
The encoder passes game images through a convolutional neural network (CNN), combines them with numerical observations, then passes through a multilayer network (MLP) which outputs the frequency and amplitude needed to stimulate each channel.
The code is open source. In this implementation, channels were not selected - only frequency and amplitude were chosen. Later experiments tried having the multilayer network choose different channels for stimulation.
Since gradients cannot be propagated across the multi-electrode array, random stimulation must be used to train the encoder. Beta sampling technique is used to determine the best amplitude and frequency for cell stimulation.
This is completely different from reservoir computing. The main difference is that the cell environment changes - when a signal is sent with specific frequency and amplitude, the cell's environment changes differently each time, making it impossible to accurately predict performance.
Beta sampling is part of the cell policy learning algorithm. This changes the physical structure of cells and alters their output over time, making every moving part in the system effectively influence how cells process and output information.
A linear reading decoder is used. Initial problems occurred when the decoder size was overestimated - it was capable of playing by itself if pulses were zeroed out or if it had sufficient biases.
The decoder size was greatly reduced and ablations performed to ensure pulses carried enough information. The bias of the linear decoder was set to zero, preventing over-customization and allowing progress in the game.
This is extremely important because if the silicon decoder suffers from over-customization, cells will never learn and will continue random behavior while the decoder plays by itself.
The speaker warns that manipulating performance of these systems is very easy - greatly amplify the decoder, apply backpropagation, and suddenly brain cells appear capable of doing anything.
This was recently seen with fly brains where neural connection diagrams were used with very large decoding units - the fly doesn't do anything, it's just a silicon chip doing everything.
Startups and large companies should be careful about this and the type of information that spreads this idea.
The decoder consists of several sets of behaviors: forward, backward, nothing, sideways movement right and left, turning right and left, and attack. At each step, softmax function calculates common probabilities of motions.
Feedback has been widely covered by media but often incorrectly. Good and bad feedback are defined through asynchronous and synchronous stimulation based on the principle of free energy - the brain tries to avoid asynchronous stimulation.
If something performs badly, asynchronous stimulation is given. If it performs well, synchronous stimulation is provided.
Using reinforcement learning creates a unique problem: at the beginning of the game the situation is very bad, and cells are not good enough for long-term credit allocation.
A silicon critic is used to see how far actions are from what the critic expects and how surprising particular actions are. Responses are adjusted based on surprise - if the system expects an action to be good but it's actually very bad, negative feedback is given.
This allows balanced response adjustment without continuous negative stimulation, allowing connections to build over time by minimizing surprise for the silicon critic.
Different motivational patterns exist for each channel. The website has interactive elements showing channels lighting up and different stimulus policies.
Synchronous stimulation is synchronous, while asynchronous stimulation for negative feedback is released at different times randomly. The surprise element is modified over time using TD error, changing frequency and amplitude of feedback sent to cells.
Nerve cells themselves do not feel pain - they have no pain receptors and no harmful toxins are introduced. Positive and negative feedback refer to game performance, not cellular sensation.
Asynchronous feedback confirms unwanted actions. The brain optimizes to reduce entropy in the culture rather than to feel good or bad.
A frequency like 40 Hz increases entropy of the culture. The brain wants to reduce the rate of receiving this high-entropy stimulation and will self-regulate to prevent it.
Changing the TD error scale means changing the surprise scale, which means there is no fixed reward model - the reward model itself is being changed.
The system never reaches a state of equilibrium. It can be thought of as two dynamic systems competing for control, with entropy as their enemy, pushing them toward quasi-attractive states.
Entropy penalties were used in the program, which is unusual in PPO algorithms. Traditional reinforcement learning prefers entropy rewards for exploration, but entropy was penalized because cells provided sufficient entropy.
This creates two dynamic systems constantly striving to improve their outputs.
The stimulation used is completely different from actual electrical signals from eyes and ears. Human intelligence is not ideal for computing - humans are designed to survive and reproduce, not for computing.
It's a coincidence that the basic foundation is very good and suitable for intelligence. The stimulation may develop to become similar to natural inputs, which may be concerning.
Evolution will continue, and different forms of intelligence may arise on the same basis as human intelligence.
Some people have tried training on actual electrical signals from the brain, including growing miniature eye organoids and connecting them to the brain. However, this becomes very concerning as it involves rebuilding a human being from scratch.
[01:19:, 31] Scalability Challenges
Scalability is one of the most difficult questions. Using biocomputation faces two main problems: reaching advanced levels of intelligence in cells, and delivering this intelligence to billions of people worldwide.
There is some understanding of how to access intelligence in cells - all humans are intelligent beings on the same substrate. However, distributed computing has not been achieved yet.
The answer to scalability is unknown, but the team is working to figure it out and has assembled a team to find solutions and take steps toward scaling this technology.
Keep Y Combinator in your library
Save the videos and channels worth coming back to, and find them again in one place.
