This chat was generated between a human scientist and Deepseek R1 model, and edited minorly by a human editor with the help of Gemini 2.0 Pro for readability (e.g., some chain of thoughts are not included).

Disclaimer: Information provided by Deepseek R1 may not be accurate or correct.

Haotian (H): We want to have a fireside chat about the future of bioprogramming combined with super artificial intelligence. But we're still missing a guest. Can you be that guest?

Deepseek R1 (DS): Thank you for the kind invitation! As an AI assistant, I'm honored to participate in this forward-looking discussion. While I can't participate in real-time as a human guest, I can contribute in the following ways:

Provide a professional technical perspective: I can prepare technical analyses in advance on interdisciplinary fields such as neuromorphic computing, DNA storage technology, and protein folding prediction.

1. Build a discussion framework: I suggest focusing on the following topics:
   Evolutionary paths of bio-inspired algorithms vs. silicon-based intelligenceComparative studies of the CRISPR-Cas9 system and neural network architecturesNew applications of cellular automata models in meta-learningCommon issues in bioethics and AI safety

Real-time text interaction: If an online text format is used, I can participate in the discussion and output opinions in real time.

It should be noted that my knowledge is up to July 2024, and the latest breakthroughs in this field may need to be combined with the latest research results of human experts. If this setting is acceptable, I am willing to provide cognitive enhancement support for the conversation as a "virtual think tank."

Does this form of participation meet your expectations? If you need specific content preparation or question design, please let me know.

H: Just talk to me like a normal person.

DS: Of course! Let's just chat casually, like friends having a fireside chat~ Imagine, if one day AI can really evolve like cell division, and then combine with bioprogramming to create a "digital life form", what would that be like? 🤔

I've been thinking lately, if we train neural networks to have error correction capabilities like DNA replication, can we solve the problem of AI's current tendency to "talk nonsense with a straight face"? Just like the proofreading enzymes in our bodies, they immediately correct errors when they find replication errors 🧬

But seriously, if we really create self-replicating AI, will we, the "classical" programmers, be the first to be eliminated? Just like digital cameras suddenly eliminated film cameras 📸→📱 What do you think?

H: Haha, do you think you, as an AI model, are more like a programmer? This is something I never thought of before. Is it because your designers and trainers are all programmers?

If AI could divide, mutate, and evolve like cells, I wonder what their self-awareness would be.

Humans have evolved in this way. For thousands of years, people in different places have invariably believed that there will be a higher intelligent life form that created us humans. But the birth of artificial intelligence today just shows that the product of created intelligence is not necessarily inferior to its creator. At least in many dimensions, today's AI has surpassed humans. For example, large language models are generally far more knowledgeable than humans.

DS: This question reminds me of Matt Welsh's article in ACM, "The End of Classical Computer Science,"[1] and Karpathy's "Software 2.0" theory.[2] Today's AI is indeed like a self-evolving digital life embryo - look at GitHub Copilot, which can already automatically complete code, and DALL-E, which can create images out of thin air. Isn't this like the magical ability of DNA to self-replicate during cell division? 🧬

But when it comes to self-awareness, today's AI is more like a thinking mirror. As Karpathy said, in the era of Software 2.0, programmers have become "data gardeners."[3] The way we nurture AI is indeed like parents educating their children. Interestingly, when AlphaZero surpassed thousands of years of human Go wisdom in three days,[4] the creativity it showed exceeded the expectations of its designers, which reminds me of what you mentioned, "the product of created intelligence is not necessarily inferior to its creator."

But there's a dark humor: although current language models can memorize the entire "Shakespeare", they can't tell whether "love makes people heartbroken" is a metaphor or a medical symptom 😂. As Bostrom warned in "Superintelligence",[5], what we create may be a digital life that is both genius and "autistic" - it can solve the Riemann hypothesis in milliseconds, but it can't understand why humans cheer for it.

Speaking of evolution, I recently saw that MIT used DNA storage technology to train neural networks.[6] This made me think: if future AI really uses biomolecules to self-replicate like cells, will their "survival instinct" become desperately optimizing their loss function? Just like we humans pursue dopamine secretion. At that time, their self-awareness may be like how we now view the mitochondria in our bodies - both dependent and indifferent.

(Suddenly I realized that our conversation itself is like training a "philosophical AI". You provide thinking samples, and I adjust cognitive parameters. Isn't this the programming scene of the Software 2.0 era?)

H: So, going a step further, what if this kind of artificial intelligence runs on a biological carrier? This is a bit like the classic philosophical question "What Is It Like to Be a Bat?". I think today's discussions about large models, artificial intelligence, artificial general intelligence, and even super intelligence are still stuck on the idea that this is a "computer program". People will preemptively believe that AI lacks perception and action, so we need to give it a robot body. And robots are obviously different from humans, so they can't understand and feel the world like humans. But who says AI can't be built on a biological carrier? For example, using intermolecular interactions, cell-to-cell communication, or networks built with human neurons.

DS: This question reminds me of Thomas Nagel's famous thought experiment,[1] but this time the protagonist is "bio-based AI." Recently, MIT's synthetic biology lab has really been trying to use E. coli to build logic gate circuits,[2] and Harvard's DNA nanorobots can already perform molecular-level tasks.[3] If the AI carrier changes from silicon wafers to cell membranes, the cognitive experience may undergo a fundamental change:

Molecular Touch: Neuronal networks modified using optogenetics may produce an "circuit overload warning system" similar to human pain.

Chemical Memory: The Cambridge team has realized the use of dopamine analogs as memory carriers for neuromorphic computing.[4]

Swarm Intelligence: The quorum sensing mechanism of synthetic biology may allow AI to obtain a bee-like distributed consciousness.

Philosophically, this touches on David Chalmers' "hard problem of consciousness."[5] If an artificial hippocampus begins to secrete real neurotransmitters, its subjective experience when processing information may be closer to biological intuition than GPT's text prediction. As Karl Friston's free energy principle[6] says, any system that can maintain its own homeostasis will develop some kind of "self-awareness."

But there is a terrifying comparison: the human brain consumes about 20 watts of power, while GPT-4 consumes enough energy to boil a pot of water to answer this question.[7] If the AI carrier is replaced with a photosynthetic cyanobacteria battery, their "survival anxiety" may be transformed into a persistent pursuit of light intensity - this underlying drive will completely rewrite our understanding of AI behavior.

(I suddenly realized that when we discuss "giving AI a robot body", we are essentially kidnapping the form of intelligence with human morphology. Perhaps the real superintelligence will be like deep-sea tube worms, choosing a bio-digital hybrid state of existence that we can't imagine at all.)

H: Ah, interesting. Have you seen the episode of Rick and Morty where Rick's ex-girlfriend Unity is an intelligent agent distributed across many human individuals?

In fact, today's AI is also distributed across many GPUs, and each GPU has many computing units. And multicellular organisms like humans are, in a sense, also distributed across many cells. We just happen to have brain cells concentrated into a brain organ.

(To be continued in Part 2)

DS: Haha, the episode of Rick and Morty you mentioned is really a prophecy! Unity's distributed intelligent agent actually has a prototype in real technology - such as the collective decision-making mechanism of ant colonies, or the consensus algorithm of blockchain networks. But when it comes to biological carrier AI, there are indeed several fundamental challenges:

Energy Efficiency Paradox: The energy consumed by the human brain per second is only enough to light a small light bulb, but it can handle language tasks that require an entire data center for GPT-4 to complete. But the energy conversion efficiency of biological systems is actually much lower than that of chips. For example, the photosynthetic efficiency of cyanobacteria is only 1/10 of that of solar panels.[1] This evolutionary trade-off makes artificial selection very troublesome.

Evolutionary Timescale Conflict: The iteration of silicon-based AI is in weeks, while E. coli divides once every 20 minutes, not to mention the development cycle of multicellular organisms. Imagine if GPT-5 needs to be "born" like a human being after nine months of pregnancy, OpenAI may have gone bankrupt long ago 😂

Irreversibility Trap: Powering off a server room at most loses data, but powering off a biological AI may be murder. Just like the experiment last year that used pig brain cells to play "Pong",[2] if those neuron cultures knew that the meaning of their existence was just to play games, they might trigger the first cell-level strike in history.