AsianDadEnergy's Podcast

Speaker critiques biocomputing limitations and ethics while detailing Cortical Labs' neural hardware

11 min Listen to the original Episode page
BiocomputingArtificial IntelligenceEnergy EfficiencyNeuroscienceEthics in AICortical Labs

Summary

An ex-big tech engineer discusses biocomputing as an alternative to Large Language Models, citing energy inefficiencies in current AI data centers. He highlights Cortical Labs' CL1 system, which uses living human neurons to perform tasks like playing Pong, noting its superior energy efficiency compared to silicon chips. The speaker details the training methods involving electrical rewards and punishments, and the clustering of units into 'hive minds.' However, he expresses deep skepticism about scalability due to technical fragility and raises ethical concerns regarding the treatment of living tissue and potential consciousness.

Episode sections

  1. 0:00 – 1:39

    Introduction to Biocomputing and LLM Limitations

    The speaker, an early-retired ex-big tech engineer, introduces biocomputing as a potential alternative to Large Language Models (LLMs) like ChatGPT and Claude. He highlights key limitations of current AI, including the inability to learn continuously post-training, struggles with novel problems, and massive energy consumption by data centers.

  2. 1:39 – 4:49

    Biological Advantages and Historical Breakthroughs

    The speaker explains that biocomputing aims to replicate the human brain's unified memory and processing structure, which is more energy-efficient than conventional computing. He cites a 2008 University of Reading experiment where rat neurons controlled a robot and a 2013 breakthrough in creating human neurons from skin cells.

  3. 4:49 – 6:48

    Cortical Labs CL1 and Training Methods

    Cortical Labs' CL1 biocomputer uses ~800,000 human neurons maintained for six months via a life support system. The speaker describes the training process for playing Pong, where neurons are rewarded with electrical signals for correct actions and punished with noise for errors, causing them to rewire themselves.

  4. 6:48 – 7:48

    Hive Minds and Physical Robotics

    Cortical Labs has clustered up to 15 CL1 units into a 'hive mind' that outperformed traditional AI (like DQN) in guiding video game characters. They also demonstrated a robotic dog controlled by lab-grown neurons navigating a map and offer a cloud-based 'wetware as a service' API for $300/week.

  5. 7:48 – 8:48

    Technical Disadvantages and Future Skepticism

    The speaker notes biocomputers are slow, delicate, require expensive life support, have short lifespans (<1 year), and are difficult to reproduce at scale. He expresses skepticism that biocomputing will succeed, suggesting biomorphic silicon chips might achieve similar goals without living tissue.

  6. 8:48 – 11:24

    Ethical Concerns and Cosmic Horror

    The speaker expresses deep aversion to biocomputing, fearing it twists human intelligence to create artificial general intelligence without conscience. He questions whether the neurons experience pain or primitive consciousness during training and worries about the ethical implications of their 'deaths' and memory loss.

Claims worth verifying

America's data centers may consume more electricity than entire countries like the UK or Italy.

Unidentified speaker (speaker_1) ·0:59

How to verify: Check energy consumption reports from the U.S. Energy Information Administration (EIA) or the International Energy Agency (IEA) comparing US data center usage to national consumption figures for the UK and Italy. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The adult human brain runs on like 20 watts a day.

Unidentified speaker (speaker_1) ·3:11

How to verify: Verify against peer-reviewed neuroscience literature or textbooks on human metabolic rates, which typically cite the brain's power consumption in watts (e.g., ~20W total, not per day, though the speaker says '20 watts a day' which is likely a transcription error for '20 watts' or '20 watts per day' is physically impossible for a brain; standard fact is ~20W continuous). Note: The transcript says '20 watts a day', which is factually incorrect for power (Watts are J/s), but the claim is what was said. Verification: Search 'human brain power consumption watts'. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

In 2008, researchers from the University of Reading extracted ~300,000 living neurons from rat fetuses to control a wheeled robot.

Unidentified speaker (speaker_1) ·3:31

How to verify: Search academic databases (IEEE Xplore, PubMed) for University of Reading biocomputing papers from 2008 involving rat neurons and robotic control, specifically looking for authors like Mike Levin or similar researchers in that field. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

In 2013, scientists figured out a way to convert human skin cells to pluripotent stem cells and reprogram them into human neurons.

Unidentified speaker (speaker_1) ·4:30

How to verify: Verify via scientific publications from 2013 on induced pluripotent stem cells (iPSCs) and neuronal differentiation, such as papers by Shinya Yamanaka or related groups. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Cortical Labs' CL1 biocomputer consists of about 800,000 human neurons.

Unidentified speaker (speaker_1) ·5:09

How to verify: Check Cortical Labs' official website, whitepapers, or press releases for specifications of the CL1 device. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

The CL1 life support system allows human neurons to stay alive for around six months.

Unidentified speaker (speaker_1) ·5:29

How to verify: Verify via Cortical Labs technical documentation or scientific papers detailing the lifespan of cultured human neurons in their specific bioreactor setup. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Cortical Labs trained the CL1 to play Pong using electrical reward/punishment signals.

Unidentified speaker (speaker_1) ·6:09

How to verify: Search for Cortical Labs publications or demos on 'Doom' and 'Pong' with biological neurons, specifically looking for their 'Hydra' or 'CL1' project details. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

A hive mind of up to 15 CL1 biocomputers outperformed the DQN algorithm in guiding tasks.

Unidentified speaker (speaker_1) ·7:08

How to verify: Verify via Cortical Labs research papers or blog posts comparing their biological hive mind performance against traditional reinforcement learning algorithms like DQN. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Cortical Labs released a cloud-based wetware as a service API for $300 a week.

Unidentified speaker (speaker_1) ·7:48

How to verify: Check the Cortical Labs website or API documentation for current pricing and service offerings for 'wetware as a service'. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Primary source

Biocomputers have a lifespan of less than a year.

Unidentified speaker (speaker_1) ·8:28

How to verify: Verify via Cortical Labs technical specs or scientific literature on the viability and lifespan of cultured human neuronal networks in vitro. For recent events, verify 2026-08-19 through 2026-11-19. Use another period only when the speaker explicitly names it.

Do your own research

  • Verify Cortical Labs' CL1 specifications and neuron count via their official website or whitepapers.
  • Check the 2008 University of Reading study on rat neuron robotics for methodology and outcomes.
  • Review Cortical Labs' API pricing and service terms to confirm the $300/week 'wetware as a service' cost.
  • Investigate the 2013 stem cell research breakthroughs regarding ethical sourcing of human neurons.
  • Assess the energy consumption metrics of major AI data centers against national grid data.
  • Research the technical limitations of biological hardware scalability and lifespan.
  • Look for counter-arguments from biomorphic silicon chip researchers regarding the viability of living tissue in computing.
  • Examine ethical guidelines and regulations surrounding the use of living neurons in commercial AI applications.
Full transcriptexpand

0:00Unidentified speaker (speaker_1): Hello world. I'm an unemployed ex big tech software engineer with 25 years of experience in tech. Now that I'm early retired with so much free time on my hands, I find myself sometimes exploring subjects I probably should not explore. And guys, over the last couple of days, I have fallen down this deep, dark technical rabbit hole. Let me take a step back and explain how I got here. So, AI systems, right, like Claude and ChatGPT. they're all the rage these days. But today's AI, they're systems built around the large language model architecture. And these large language models, they have some inherent limitations. For example, once a model has been trained, it can't learn anything new or update its information on its own. These models, they also struggle with completely new, unfamiliar, novel types of problems. This is especially the case when, for these problems, they don't have enough underlying training data to find these matching patterns. And then there's the enormous amount of energy needed, right? Electricity to power the data centers that run these AI models. By some estimates, America's data centers, it may consume more electricity than entire countries like the UK or Italy. Now, these limitations are big enough that a lot of people

1:19out there, myself included, we believe that true AGI may not come from just scaling up large language models. It may require a completely different approach to AI. So naturally, researchers out there, they have been exploring different alternative ways to build artificial intelligence. Some are developing more efficient machine learning algorithms and more efficient chips. Others are exploring entirely new ways of computing. And one particularly interesting alternative is biocomputing. The basic idea behind biocomputing is that Mother Nature, she has already found the near perfect answer to true generalized information. intelligence. And that answer is the human brain. Think about it. In a conventional computer, memory and processing, they're two different separate components, right? So the storage and persistence of data, of memories, it's a discrete process. And that data, it has to shuttle between processing and memory almost constantly. That movement of data costs time. and it costs energy. But in human brains, memory and processing they exist as a single unitary structure, the neuron. This biological architecture allows

2:39people to learn things on the fly. We're constantly adapting as new information is received. It also gives us the ability to generalize. We can solve completely new problems. even those that we have little to no experience with. Modern AI models, right? They need billions of training data examples to learn something that a child could pick up with just a handful of examples. And in terms of energy efficiency, right? The adult human brain It runs on like 20 watts a day. That's literally millions of times more energy. efficient than an AI chip in a data center somewhere. So the goal of biocomputing is to harvest these natural advantages of biological neurons in order to build more efficient and powerful AI systems. So there's been a bunch of significant technical breakthroughs in the field of biocomputing. Way back in 2008, there was this team of researchers from the University of Reading. They surgically extracted around 300,000 living neurons from rat fetuses and placed these neurons into a dish. The neurons were then connected to this array of

3:51electrodes, and then the living neuron, through those electrodes, were given control of a wheeled robot. Ultrasound sensors in front of the robot, it would detect nearby obstacles like walls. Those signals were then converted into electrical impulses and sent to the neurons. And the resulting firing patterns from those neurons, they were collected and routed back to drive the robot's wheels. So in this experiment, these neurons were able to rewire themselves to avoid obstacles, effectively demonstrating that it was learned. learning on the fly. Another major breakthrough happened in 2013, where scientists figured out a way to convert human skin cells to these pluripotent stem cells. These stem cells were then reprogrammed into human neurons. I suppose it's a good thing that we can now produce huge amounts amount of human neurons on demand. That's way better than having to harvest them directly from people. people's brains, right? Now, more recently, there's this AI company called Cortical Labs. And Cortical Labs seems to have brought all of these prior breakthroughs and innovations into a working product. They call this biological computer the CL1. So their biocomputer

5:09It consists of about 800,000 human neurons sitting in a new nutrient bath. Now, they supposedly experimented with many different types of neurons from different animals, and they found the human neurons to perform the best. There is this surrounding life support system that keeps this wetware environment stable. This life support system allows the human neurons to stay alive inside of this computer for around six months. So this layer. of living neurons, it's sitting on top of this high-density electrode array. And this electrode array it routes input and output to and from the neural network, kind of like a BIOS system. Now, this CL1 biocomputer is shipped with a Python SDK. And this SDK allows people to train this living neural network to perform. various cognitive tests, tasks like playing the game of Pong. And this training process is somewhat disturbing to me. If we take Pong as an example, the chip translates the game world into electrical patterns. And it reads the cell's spike back signals as kind of like actions in this game, like the thing moving. So when the neurons hit the ball, they're rewarded with this predictable

6:29electrical signal. But when they missed, they're then punished by this chaotic stream of electrical noise. Now, in this training process, the neurons would essentially reorganize and rewire themselves in order to avoid being punished. The end result is a neural network that learned how to play pong. Now, beyond playing pong, Cortical Labs have done a bunch of things with their biocomputer. The computer was trained to play the game Doom effectively. Then they decided to cluster a whole bunch of these biocomputers together. I think it was up to like 15 biocomputers to form this hive mind. They found that this hive mind performs some cognitive tasks much better, like the task of guiding. a video game character towards a target. They found that the Hive Mind it actually performed better for this task than traditional AI video game algorithms, like the DQN algorithm, for example. In terms of cognitive tasks, they have now gone beyond virtual video game environments into the physical environment. Cortical Labs recently demoed a robotic dog that was. being controlled by neurons grown from their lab. The RoboDog, it was tasked with autonomously navigating to a point in this map. They have

7:48also recently released this cloud-based wetware as a service. Now anybody can sign up for an API key and experiment with these living computers for just $300 a week. I kid you not, I'll link it in the description. Now, with all that said, I think it's still early days for biocomputing. Compared to high-end AI chips, biocomputers have several major disadvantages. Neurons are slow, they're delicate, they need expensive life support systems to deliver nutrients and for climate control, and so on. And each biocomputer, it only lives for less than a year. Being living things, each neural network changes over the span of its brief life. Each biocomputer is slightly different from one another. Because of this, training results, they're difficult to reproduce consistently and at scale. And fundamentally, the promises of biocomputing are still largely just

8:48future projections, right? Because look, machine learning algorithms, they're still improving. Silicon chips is also improving. We may very well see near future biomorphic chips that achieve the same goals as biocomputing without using living human tissue. And quite frankly, all of this It has left me with a deep aversion towards biocomputing. In fact, this whole idea it's so revolting that I wish I had never learned about it in the first place. Fundamentally, these AI companies, they're trying to twist and warp the essence of human intelligence itself. This near perfect, almost miraculous intelligence that Mother Nature, God, the universe that they gifted us with. And these companies, they're doing all of this twisting. in order to create an artificial version of general intelligence. Something as powerful and as efficient as the human mind, but without the limitations of human consciousness or human conscience. But here's the thing: is this clean separation even possible? These living neural networks, they can rewire themselves to avoid punishment, which means they can. form memories of what they have experienced. But how do we know that those memories

10:08don't carry a trace of the pain, maybe even the torment that they have experienced during the training process? How can we be certain that somewhere inside of those networks, there isn't some primitive form of consciousness, some flicker of awareness, like a faint echo of their own existence? And when their brief operational lifespans come to an end, what happens to those memories? All those moments of experience. Do they simply vanish into nothingness, never to exist again? Are they lost? Like tears in the rain? Guys, this is cosmic horror-level Horror level stuff. Like a science fiction nightmare brought into the real world. And I sincerely hope to God that this technology it never takes off. And that's all I got to say about that. Hope it helps. Anyways, if you have a morbid curiosity to join me in this life journey, please subscribe to my channel and subscribe to my Substack newsletter. If you want to support me in my Vlog Blog creation efforts, please become a paid member of the channel, a paid member of my Substack, or just buy me a copy. coffee. If you want a one-on-one coaching session with me, just schedule it. Anyways, thanks for listening. Talk soon. Bye.

This page was produced automatically: the audio was transcribed with nvidia/parakeet-tdt-0.6b-v3+sortformer-v2.1 and interpreted by Qwen3.6-35B-A3B-UD-IQ3_S.gguf. Transcription and interpretation can contain errors; always confirm against the original episode. Nothing here is investment advice.
All insights