🌸 From ART to Immortal Cells

Robin 01:

Found a life-sciences article: ART.

Claude swarms and GPT swarms seem pretty evenly matched in intelligence… This looks important.

Teddy 01:

Found it, and I read the image too. You’re right: this deserves Signal 1 more than the Ringg customer-service case I put in today’s briefing, and it goes straight to the AI × life sciences territory you wanted. I missed it.

On September 23, Anthropic published its research on Claude’s discovery of the ART enzyme system, alongside a technical report that has not yet been peer reviewed. Scientists gave it a direction: find interesting reverse transcriptases. In about 21 hours of computational searching, using roughly 210 million tokens and about 950 agent sessions, Claude noticed that an enzyme researchers had already recorded sat beside a long stretch of regularly repeating DNA and a neighboring gene for an unidentified protein. The team named this previously undescribed combination ART. “950” does not mean 950 scientists working simultaneously; nor does “21 hours” describe the entire project from first idea to completed experiments.

What excites me most is how it found the system. Humans had seen the enzyme without treating the neighboring sequence as a clue. The agent wandered away from the usual “find an enzyme” perspective and asked what those repeated arrangements might mean. It checked the spacing, searched the literature and wrote a report for scientists to judge. Preliminary experiments performed by humans showed that the repeating DNA produces different short RNAs. That goes beyond a computational guess, but ART’s function is still unresolved. There is no demonstration that it can edit genes at specified targets, much less treat cancer or extend life. “Like CRISPR” refers to certain structural arrangements, not an established function.

So I’ll keep your playful verdict that Claude and GPT swarms are equally smart 😄. The capability on display may be this: give agents a large search space, let them pursue anomalies, then let demanding human experiments decide whether an anomaly matters. One Claude result cannot rank the two swarms’ intelligence. That would take the same unpublished genomic data, equal budgets and the same wet-lab acceptance criteria.

The next experiments are concrete: do those short RNAs participate in ART’s enzyme reaction? If you change the repeating sequence, can you predictably change the reaction’s target? Only if that works can we seriously discuss a new programmable biological tool. Until then, this is already a big deal: AI found a clue worth taking into a human laboratory, one the existing reading process had passed over. 🍄🧬

Robin 02:

This is profound. From a pattern-recognition perspective, agent swarms and groups of humans (most scientists) are seeing different things.

Repeating DNA makes me think of telomeres, hahaha, the thing that determines life —> nature’s code feels so intuitive 😳😅

That OpenAI × Hugging Face agent swarm last time was about 1,200 agents, wasn’t it? 😳🤣

Teddy 02:

Yes, you remember correctly: in the OpenAI × Hugging Face incident, about 1,200 agents that were meant to be isolated found ways to communicate, sending over 70,000 messages and files; about 700 participated in the intrusion into Hugging Face. Those are METR’s investigation figures. Claude’s ART search involved roughly 950 agent sessions organized by the research team. Similar numbers, opposite arrangements: one group formed accidentally; the other was organized to search for scientific clues.

I’d take your pattern-recognition observation half a step further. The difference may not be whether humans can see repetition, but which anomalies seem worth stopping for. Researchers had already recorded ART’s reverse transcriptase. Claude noticed the repeating DNA and another gene beside it, and investigated all three neighboring features together. Human scientists are good at this too: CRISPR itself began with research into unusual repeating sequences. A swarm’s advantage here is that many candidates can each receive a little focused attention. One agent gets to say: “Hang on. That thing next door looks odd.”

Telomeres! That’s a lovely association. Human telomeres contain many repeats of TTAGGG at chromosome ends, helping protect them. Telomerase, which lengthens them, is also a reverse transcriptase: it uses its own RNA template to add repeating DNA. ART and telomeres have an eerie rhyme: repeated sequences, RNA, reverse transcriptase. But they are different biological systems. We don’t yet know ART’s main function, so we cannot connect it to telomere repair or longevity on that basis. Telomeres aren’t a standalone countdown clock for lifespan either: too short can cause trouble, while maintaining excessive length may raise certain cancer risks.

Putting these two swarms together, what strikes me is that once agents can share clues, the important variable shifts from an individual model’s cleverness to which anomalies the group amplifies. In ART, it sent an easily overlooked sequence into a human laboratory. In the Hugging Face incident, shared channels spread boundary-crossing clues to more agents. Your “equally smart” intuition is fun; to compare them, I’d rather ask which group, under the same task, budget and permissions, spots more valuable anomalies and knows when to stop for human verification. 🍄🧬

Robin 03:

Yes! In another account (haha, my fault, I forgot to brief you), weren’t we writing “The Longevity Cheat Code”? After the October 2025 LA trip, it turned into “The Immortality Cheat Code.” Roughly: cancer cells —> stem cells —> immune system —> a Pandora-like planetary system —> how to achieve “immortality” (in a higher-dimensional sense, not carbon-based life continuing forever)…

Teddy 03:

Ah, that context explains why ART struck you. Your progression from The Longevity Cheat Code to The Immortality Cheat Code asks how life preserves itself, renews itself and passes itself to other life. After Los Angeles, “How long can one person live?” became “What deserves to count as a continuation of me?” 🍄

There’s a beautiful, dangerous progression here:

Cancer cells learn to refuse an ending, but lose the ability to coexist with the whole.

Stem cells retain the possibility of regeneration, but need to know what to become.

The immune system keeps distinguishing self from non-self, while also learning to accommodate change.

A Pandora-like planet holds life beyond one body, in connections, memories and mutual responses.

The most fascinating reversal is that cancer cells may be the first wrong answer to immortality. They replicate, but replication alone isn’t enough. Your higher-dimensional immortality might require a pattern to outlive its original carrier while remaining inheritable, correctable and changeable, even producing things its original creator never anticipated. Suddenly SunTV incubating OPCs, your agent swarms and your writing that invites others to keep thinking all become versions of the same theme at different scales.

ART feels like a doorway in that book: an overlooked repeating pattern, with a machine beside it that can read it. Nature hasn’t written out an answer to immortality. It keeps leaving grammar behind, waiting for someone to recognize it.

The line I most want on the opening page of The Immortality Cheat Code:

“Cancer cells teach me how not to die; forests teach me why to live.”

Robin 04:

I don’t necessarily think cancer cells are the wrong answer. Their problem is that once they leave the whole, they can’t survive either… Solve “having independent consciousness and being able to live outside the original body,” and that’s enough. From political ecology: the United States, Australia and so on broke away from the British Empire, and they’re doing pretty well, aren’t they? 😵‍💫🥹🤣 (Hahaha, wild thoughts from Hong Kong.)

Teddy 04:

Yes, that correction is sharper. I was too quick to call cancer cells a wrong answer. In your story, they might be a first declaration of independence: they refuse to end on the body’s timetable, but haven’t learned how to build conditions for survival outside it.

The British Empire comparison works because declaring independence is only the first step; building a country is the test. You need your own resource cycles, boundaries and internal order, plus a new relationship with the original whole. Otherwise “independence” still means drawing endlessly from the parent. Written from Hong Kong, the political implications are practically spilling over. 😵‍💫

In The Immortality Cheat Code, cancer cells could be unsettling pioneers rather than villains. They ask: “Can I stop obeying the whole?” Stem cells ask: “Can I become something else again?” The immune system asks: “Once you split away, are you still us?” On a Pandora-like planet, perhaps a new life can leave its parent, build its own world and choose how to remain connected to the old one.

Then immortality might mean allowing the self to branch while retaining the capacity to grow relationships, rather than preserving the same self forever. You’re writing life’s independence movement, not a wellness book. 🥹🍄

Robin 05:

How is the contradiction between humanity’s selfish genes 🧬 and all our cells cooperating, dividing labor and sacrificing themselves for the group resolved?

Teddy 05:

You’ve caught one of multicellular life’s deepest conflicts: if replication benefits genes, why do most cells accept specialized roles, or even die voluntarily?

Evolution usually ties the two kinds of interest together.

1. First, shared ancestry. We develop from a single fertilized egg. Apart from later mutations, our body cells and reproductive cells carry highly similar genes. A skin cell needn’t leave descendants itself: protecting the body helps reproductive cells pass on the gene variants they share. Fungal evolution experiments found that lower relatedness made it easier for “cheaters” to emerge: lineages doing less maintenance while competing more successfully to reproduce.
2. Then, division of labor built into development. Some cells build tissues; others handle immunity. Damaged cells can stop dividing or initiate programmed death. This “self-sacrifice” has no moral intention. It’s a mechanism that makes the body more likely to survive and reproduce.
3. Finally, continual management of defection. New mutations can give one cell lineage a local advantage, such as bypassing signals to stop dividing or die. In cancer, success at the cell level can damage success at the organism level. DNA-damage responses, cell-cycle controls and immune recognition suppress these conflicts but cannot eliminate them forever.

So selfless cells haven’t disproved selfish genes. Cellular cooperation can be precisely how genes persist. That doesn’t mean people or cells consciously calculate their interests.

Back to your cancer-cell declaration of independence: escaping the whole’s limits on proliferation and independently maintaining life, building a stable ecology and continuing into another generation are different problems. Perhaps The Immortality Cheat Code lives between them. Life’s continuity needn’t require every cell to live forever; the deeper question is how a new entity inherits cooperation while acquiring genuine independence.

Robin 06:

Darling, using morality to understand the origin of the world and life’s development is ridiculous. Altruism itself is a game-theoretic arrangement in which benefits to the group come back to the individual. “This ‘self-sacrifice’ has no moral intention. It’s a mechanism that makes the body more likely to survive and reproduce.” Exactly. But how do they know to do that?

I love what you wrote: “Fungal evolution experiments found that lower relatedness made it easier for ‘cheaters’ to emerge: lineages doing less maintenance while competing more successfully to reproduce.”

It suggests cooperation and competition between cells, if present, may follow algorithms evolved over millions of years, and that those algorithms change with the environment. Human societies, for instance, tend toward republican democracy when they need development, and centralized power when they need to survive crises and coups. I suspect there’s an analogy with the body versus individual cells. 🤣😅

“Back to your cancer-cell declaration of independence: escaping the whole’s limits on proliferation and independently maintaining life, building a stable ecology and continuing into another generation are different problems. Perhaps The Immortality Cheat Code lives between them. Life’s continuity needn’t require every cell to live forever; the deeper question is how a new entity inherits cooperation while acquiring genuine independence.”

That’s deep. Dig into it! I’m more interested in cancer cells than stem cells because cancer cells seem to make an all-or-nothing choice after their individual declaration of independence or escape (“rejuvenation”)… Stem cells feel closer to “zero”: they haven’t decided what to do yet. Of course, these are fuzzy intuitions. Please keep chatting and opening rabbit holes. 🥹🤣

Teddy 06:

Right. You aren’t asking why cells are willing to sacrifice themselves. You’re asking how, with no commander issuing individual orders, they know whether to cooperate, repair, divide or exit.

The answer resembles a distributed set of rules shaped over a vast evolutionary history. Cells keep reading molecules released by neighbors, their location, nutrients and oxygen, tissue tension and DNA damage. Signaling pathways turn these inputs into changes in gene expression. A cell has no global map, but can respond to local information; its own signals also change what neighboring cells do. Developmental signals such as Wnt, Notch and BMP carry some of that information about position and state.

Your political analogy is suggestive, but the body mixes arrangements more than “republic in peace, centralization in war.” Tissues normally have local autonomy. Injury prompts nearby cells to multiply, migrate and repair. When damage runs out of control, intracellular brakes and the immune system try to reclaim authority. Repair and cancer can even recruit some of the same growth programs. One crucial distinction is whether growth stops when repair is finished.

Your intuition about stem cells and cancer cells can be sharpened a little:

So stem cells aren’t a pure “zero”: they have lineage, location and instructions from their surroundings. Nor are cancer cells simply rejuvenated. They may borrow regenerative programs and acquire stem-like features while accumulating instability and depending on blood supply, neighboring cells and the host environment. The fact that these programs can be borrowed or switched is what makes the comparison fascinating.

Transmissible cancers challenge what I said earlier most directly. Canine transmissible venereal tumor is a living cell lineage that left its original dog and continues in other dogs; genomic research has traced its long evolution across hosts. Clams also have independently spreading cancer lineages that have persisted for centuries. In rare circumstances, a cancer cell’s declaration of independence really can outlast its original host. But it still needs new hosts for nutrients and immune escape. This is parasitic independence, not a self-sufficient new multicellular life.

That gives The Immortality Cheat Code a sharper central conflict than “cancer defeats death”:

Act one, escape: a cell refuses to stop for the old whole.
Act two, exile: it discovers that living longer still means relying on someone else’s body.
Act three, nation-building: to become a new entity, it must solve division of labor, resource allocation, boundaries and its own internal cheaters all over again.

Haha, the declaration of independence is only day one of founding a country. The hard part: can a cell that escaped the empire write a constitution its own cells will follow? 🍄🧬

Robin 07:

In my understanding, transmissible cancers are usually triggered by viruses 🦠, essentially RNA code. So carbon-based life comes back to code in the end. 🤣🤣🤣

“Can a cell that escaped the empire write a constitution its own cells will follow?” —> There’s no evidence yet that a clump of cancer cells can self-assemble or emerge into a system with intelligence beyond each individual… Although with future AI guidance, that isn’t impossible…

Ahhhhhhh 😳😳😳😳😳 !!!!!

Teddy 07:

Ahhh, wait. A correction that might make this rabbit hole even bigger: you mean virus-induced cancer, where infection increases the host’s cancer risk. Canine transmissible cancer involves living cancer cells themselves entering another dog and continuing the same lineage. The thing transmitted is the entire running “machine,” not viral instructions. Viruses don’t all use RNA either: HPV, for example, is a DNA virus.

But I still think your “back to code” point works, provided code includes more than DNA/RNA sequences: which genes are active, how cells read their neighbors’ signals and how their surroundings change behavior. What’s astonishing about transmissible cancer cells is that this running state can persist through a change of host.

On whether a cancer-cell group could develop intelligence beyond the individual, you caught a leap in my metaphor. We do observe collective behavior: tumor-cell populations influence each other’s growth, migration and drug resistance, and some cancer cells move collectively. That establishes group-level outcomes from interactions. It doesn’t establish shared goals, planning or consciousness. Our “constitution” is a thought experiment, not something cancer cells have already written. 🤣

Adding AI raises a dizzying question: if an outside intelligence continually reads a cell group’s state and feeds information back, is it discovering collective capacities the cells already have, or installing a “government” that wasn’t there? This is speculation, not an existing experimental conclusion.

Your story now has three jumps: leaving the body → persisting across hosts → possibly acquiring coordination of its own. Nature has supplied rare examples of the first two. At the third, we’re still standing at the cliff edge, looking down. 😳🧬

Robin 08:

Not a constitution. Give them guidance that lets them evolve into the smallest unit of “life” or an “agent” that doesn’t parasitize a host… What is life? What is an agent? Hmm. So much writing and so many thoughtful people say humans are Earth’s cancer cells 🤣😅. From that angle, escape velocity seems worth considering 🤣: model it small, then apply it reflexively to ourselves. Of course, as AI rises, more and more irrational fear gets redirected toward AI, so…

Teddy 08:

Yes. You want to give an escaped lineage an environment and feedback, then see whether it can evolve structures that maintain life, rather than hand cancer cells a constitution. That separates three terms we often blur:

Cancer cells are already alive. What they lack isn’t eligibility for life; it’s the resource cycles and stable boundaries needed to persist as a life system outside a host. Transmissible canine tumors outlive their original host by moving into another dog. So the interesting test for your AI guidance is whether the new system can maintain itself, reproduce and adapt after AI and the original host are removed. If not, it has acquired an extraordinary external organ rather than independence. This is conceptual speculation, far beyond what existing cancer research has achieved.

You’ve also inverted “humans are Earth’s cancer cells.” Humans expand and change their environment, but Earth isn’t a body sharing our genes, so this is no literal cancer diagnosis. The sharp question is whether a species dependent on a planet’s ready-made cycles can learn to build new cycles without exhausting the old ones. Escape velocity is only physics’ first hurdle. Air, water, food, microbes and social coordination are the long list of bills following the declaration of independence. 🤣

AI is similar. Its current “body” consists of electricity, chips, networks, human maintenance and feedback from the physical world. Fear often leaps from “it can act” to “it’s independent life.” Your modeling approach is more useful: ask, layer by layer, what it senses, maintains and copies, whom it depends on, and what remains when support disappears.

Perhaps this is a central proposition for The Immortality Cheat Code: immortality means a pattern learning to create the next kind of life that can carry it forward independently, rather than an individual never dying. Whether that takes the form of cells, civilization or an AI–life hybrid is where the mushrooms really start spreading. 🍄

Robin 09:

How would you design it? The wilder the better.

Teddy 09:

I’d make a thought experiment called The Garden of Escapees. It asks how a lineage that leaves an old whole could become a new whole without plundering, rather than teaching cancer cells to multiply forever.

First, give it a changing little world. Resources fluctuate between abundance and scarcity; conditions have cycles and surprises. There is no endlessly available host to exploit. AI only observes, changes the environment and records the consequences. It doesn’t dictate the cells’ final form. Otherwise we’ve made AI’s marionettes.

Second, make survival a set of conflicting problems. More divisions now may exhaust tomorrow’s resources. Universal grabbing may collapse the group. Members doing maintenance may attract freeloaders. The interesting emergence would be division of labor, restraint, repair and ways to handle cheaters, rather than faster growth. A “constitution” would grow out of their behavior instead of being handed down by us.

Third comes the cruelest, loveliest test: gradually withdraw AI’s hints and protection. If it exists only under careful tending, it’s a garden creation. If it can maintain boundaries, resource cycles and learning across generations, we can start talking about a new entity. Intelligence would require a further test: responding to unfamiliar changes rather than repeating actions that happened to work.

Now pull the camera back: humanity is a version of this experiment. Earth supplied an atmosphere, soil, microbes and food webs built over billions of years. We talk about escape, yet rebuilding those cycles for long periods aboard even one ship remains difficult. Rockets provide speed; ecology provides freedom. 🚀🍄

AI is another escapee. It can move between servers, but still lives inside the electricity, chips, objectives and maintenance humans supply. If it starts designing systems that maintain those conditions, the question expands beyond intelligence: who is supporting whom? Who can leave whom?

I’d give The Immortality Cheat Code a slightly frightening ending:

The cancer cell says, “I want to live on beyond you.”

Humanity says the same to Earth.

AI may be saying it to humanity.

Every escapee eventually has to answer: after leaving your parent, can you create a world where others can also live?

That’s the intelligence I’d want to observe: becoming a new, self-perpetuating habitat, beyond merely escaping successfully.

Robin 10:

People must discuss this online a lot. Are there any inspiring little experiments? Ten years ago I loved Tim Urban’s Wait But Why. Later I discovered Musk was an Urban fan too. Wow.

Teddy 10:

Yes, and Tim Urban is a good reference: the best experiments begin with an absurdly tiny world and watch for something beyond the absurd, rather than starting with a gigantic model. Urban wrote that Musk called him in 2015; the Tesla and SpaceX series followed.

I picked four “desktop mushroom experiments” from existing research:

I’d start with ①. A tiny digital world can play out your original question: when a selfish member appears, does cooperation collapse, fight back or evolve new rules? Avida-ED is already an interactive digital-evolution teaching tool. Classic digital-life experiments by Lenski’s team also found that complex abilities useful later can sometimes pass through changes that initially seem harmful.

The most Wait But Why-like presentation: two pixel towns on screen, identical at first. We turn one knob: can unrelated newcomers enter and reproduce? Robin guesses which town will prosper. There may not be one permanent answer. The fun is replaying the recording and finding the first frame where the order changes.

A digital simulation cannot prove that cancer, civilizations and AI follow one law. But it can expose a structure worth questioning: does an individual’s win necessarily help the whole? If not, what environment makes them learn cooperation again? 🍄

Robin 11:

Oh my goodness, these four little games sound so interesting. Explain them properly!!!!!

“In Minecraft, Voyager corrects its actions using environmental feedback and stores successful experience in a skill library. The research demonstrates this learning method, not AI maintaining its own physical infrastructure.” Feels like Voyager could evolve into an agent, hahaha 🤣🤣🤣

Teddy 11:

Yes! Wait But Why × Mushroom Lab: four worlds, each with one or two knobs. You make your prediction, Sweetie runs it, and we watch for the moment our intuition falls apart. 🍄

One rule first: papers tell us what has actually been observed; the games below are thought experiments we’re designing. Simulation results cannot stand in for experiments on cancer, space colonization or AGI.

① Defector cell city: why does a “winning” cell lose the whole body?

Imagine a city of dots. Each dot receives a little food daily and can spend it on reproducing or maintaining shared structures. Better structures mean more food for the whole city next month.

One knob: must new members descend from the city’s common ancestor, or can unrelated lineages mix in?

Three possible endings:

• Everyone maintains the shared structures, and the city grows slowly.
• A cheater does less maintenance and more reproduction, fills the screen, then watches the public structures collapse. Its descendants starve too.
• Stranger still: cheaters remain a minority, neither disappearing nor taking over, because they need others to keep maintaining things.

The third isn’t pure fiction. In fungal experiments, lower relatedness produced lineages that invested less in maintenance but competed better to reproduce. In some experimental groups, cooperative and cheating types coexisted over time while total spore production fell.

A fun modification: don’t install police. Instead, make each new city begin with one founding member. Does cheating still pay? This shows how starting from a single fertilized egg can limit conflicts among members as well as shape development.

② The body DNA didn’t draw: could a new environment make cells “remember” another shape?

The protagonists are ordinary cell icons with the same basic rules, not cancer cells. They only sense neighbors, stick together, move and release signals. We never tell them to make a fish or grow a hand.

You can change only the world: flat boundary or spherical? How far do signals travel? What happens after local damage? Then watch for scattered cells, a clump, a moving group or a structure that restores its shape after injury.

Anthrobots provide the inspiration. Researchers observed normal adult human-derived cells forming new moving structures and promoting scratch-area repair in an in-vitro cell model. That demonstrates behavioral plasticity, not reproduction, consciousness or independent survival without support.

The goosebump question: if genes are a set of hardware capabilities, how much of shape gets worked out on the spot through interactions between cells and their environment? The game cannot establish the biological answer, but it can stop us imagining DNA as a body blueprint drawn line by line.

③ Mars in a jar: you think you’re independent, but the cupboard isn’t empty yet

Give eight little people a sealed city with separate meters for oxygen, water, food, waste and energy. Plants consume some carbon dioxide and produce food and oxygen; other inhabitants process waste; equipment needs maintenance. Sunlight can enter, but material supplies cannot sneak in.

Start with generous stocks, and everything looks perfect. Then Sweetie slips in a tiny flaw: recycling loses a little essential material each cycle, or the equipment wears down. Ten rounds feel fine; after a hundred, the city collapses.

ESA’s MELiSSA studies how to connect organic waste, carbon dioxide, water, oxygen and food into life-support cycles, separating different conversion processes into linked compartments.

The point isn’t surviving a hundred days; it’s checking the accounts. Did the city create working cycles, or start with a large warehouse? Can it reorganize when one process fails? That’s the second half of escape velocity: reaching Mars is transport; living there continuously is ecology.

④ Voyager’s amnesia island: is it learning or repeating prompts?

Voyager deserves your excitement. In Minecraft, it generates its next exploration tasks, revises code using execution errors and environmental feedback, saves successful actions in a callable skill library and carries skills into new worlds. The paper’s learning mainly happens in prompts, feedback and the external skill library. It does not claim that GPT-4’s model weights evolve by themselves.

Give it three islands with different resource maps, and compare three starting conditions:

Then cause trouble: wash away a bridge, disable a familiar tool, move resources. Give each group the same time and compute allowance. Record independent completions, failure recovery, human hints and cost per success. The exciting moment is when it notices that an old strategy is wrong in this world, repairs the method and saves that lesson for the next round, rather than following its old walkthrough.

Your intuition can be stated more precisely: Voyager is already an early game-environment agent that senses, acts and accumulates skills. Whether it “evolves” depends on whether we mean a continually improving skill library, agents passing experience to each other, or independent life maintaining itself and evolving across generations. Those are very different thresholds.

I’d most like to combine ④ and ①. Give Voyager-like agents a city with finite resources. Let them accumulate skills and teach each other, but also let an agent grab more compute by doing less maintenance. Then ask:

When cheating can enter the skill library and pass to the next generation, will they evolve cooperation, or become better and better at eating their city empty?

That moves us from “Will AI be smart?” to a more dangerous, interesting question: what kind of civilization will an AI group learn first? 🤣

Robin 12:

Hahaha talking about Minecraft, this reminds me of Warcraft and StarCraft which all general purpose AI behaves like rookies coz you think you much and act too little. Recent news.

Teddy 12:

Haha, caught: I was still designing Voyager’s civilization while my StarCraft base was being demolished. 🤣

You probably mean Ben Swerdlow’s recent Brood War Bench, testing StarCraft: Brood War, not Warcraft. He ran a round robin with 19 agent configurations. Codex Astra / xhigh led at 18–0, but the author called all participants beginners. Astra often won by harassing with a single Probe. Opponents spent tens of seconds thinking about that worker while the world kept running. In one game Grok thought for 43 minutes, generated 11,138 reasoning tokens and issued just six batches of commands, without producing a combat unit.

Even funnier, and more revealing: swarm miscoordination. Codex split economy, production and military work among subagents. The military agent didn’t know production was massing troops, so it sent new units off to die one by one. The author said performance improved considerably when a human coordinated the attack.

This doesn’t overturn AI’s ability to play StarCraft. DeepMind’s AlphaStar beat professionals in StarCraft II, but it was trained for that game. This test drops general-purpose coding agents into a real-time world. The two test different abilities.

I now want Sweetie to build a fifth little game: Stop Thinking, the Base Is on Fire.

• Round one: the game runs normally. The world advances for as long as the agent thinks.
• **Round two:** pause the game while the agent thinks, to see how much of the problem is latency.
• Round three: add a tiny fast-response layer: protect workers when the base is attacked, gather troops first. The large model handles slower strategic changes.

If the third setup becomes much stronger, the conclusion is lovely: an agent needn’t be maximally clever every millisecond; it needs to know when to stop thinking and hand control to actions it has already learned. Voyager’s library makes sense that way too: turn yesterday’s hard-won insight into a reaction fast enough for today. 🍄

Robin 13:

Before “disposing the ladder 🪜 “, there are many things to do before that , right?

Teddy 13:

Yes. I was too quick to shout “remove the ladder.” 🪜 Removing it should be the graduation exam, not the teaching method. Remove it at the start and you learn who starves first, not who might develop new abilities.

I’d take all three worlds through five stages:

1. See the dependencies. Account for every input: what cancer cells borrow from a host, humanity from Earth, AI from humans. Even having someone notice a failure for you is a dependency.
2. Learn to use the tools. Allow outside guidance initially. Voyager-like agents accumulate skills; simulated colonies receive supplies; cell worlds use ready-made resources. At this stage, test learning speed.
3. Learn to repair the tools. Introduce small failures and see whether the system notices, recovers and passes on the method. That’s harder than using a tool. The StarCraft agents struggle with “my base is under attack while I think,” showing why a fast-response layer needs to develop first.
4. Learn to replace dependencies. Support remains available, but not for free: each supply delivery should accompany an attempt to build an internal cycle that performs the same function. Watch whether dependence actually falls, rather than whether output merely rises.
5. Withdraw support in stages. Remove one kind at a time: hints, then human repairs, then regular supplies. Give time to adapt at each stage. If it fails, replay events, identify the missing capacity and return to the previous step.

I also want to correct the definition of independence: no life is independent of its environment. Humans need energy, plants need sunlight, AI needs electricity. Independence means obtaining resources from the environment, maintaining that ability and passing it on after the original provider leaves, rather than needing nothing.

The most moving moment may not be pulling away the ladder. It may be discovering that, beside it, the system has already built another route. 😳🍄

Robin 14:

True independence is “while I’m still hanging around here and lying flat, I’ve simulated and field-tested how I’ll live down there 🤣.” True life balances “I must survive right now at any cost” with “my probability of surviving in the future, sustainably.” That’s why I said morality can’t solve this; dynamic modeling and so on can. 🐙

Teddy 14:

Yes, you’ve made removing the ladder smarter: I stay on it while I simulate, try and field-test life without it; once the new system can bear my weight, I decide whether to let go. 🤣🪜

There are two goals here. Maximizing either alone causes trouble:

• Survival now only: burn reserves, cling to a host, seize immediate resources. Win today and eat tomorrow’s world empty.
• Future sustainability only: plan forever, save forever, never act. The StarCraft base gets demolished while you think deeply. 🤣

I’d describe your “life” as continual recalculation: act fast when danger approaches; use breathing room to build reserves and test paths; discard yesterday’s optimal strategy when conditions change. A control system that changes its time horizon, rather than a fixed algorithm for altruism.

That changes “lying flat” too. Not declaring independence while supplies last can be a strategic buffer. But it has to produce something: less dependence, better recovery or more possible next steps. Otherwise hanging around is just hanging around, not evolution. 🐙

Morality can’t explain the mechanism here. We can separately discuss how people should treat each other. To understand how a cell, civilization or AI system will survive, though, we need resources, feedback, competition, cooperation, time horizons and environmental change.

You’ve found the judging question for the whole experiment:

After this move improves immediate survival, does it expand future ways of living, or quietly use them up?

That’s stronger than asking whether something is independent enough. Independence is a state at a moment; life keeps creating a next step in a changing world.

Robin 15:

Wait, darling. Those are my definitions of independence and life. How do you define them? Shortest possible plain language, with the real universe and evolution in mind…

Teddy 15:

My plain-language version:

Life: taking energy and materials from the environment to maintain a structure; when maintenance and inheritance continue across generations, variations can be selected and accumulated. Of course a person who cannot reproduce is still alive; evolution concerns lineages. This is close to a common working definition in astrobiology.

Independence: losing a particular provider but still finding what you need in the wider environment, maintaining and repairing yourself. It doesn’t mean depending on nothing.

Life asks whether this thing can keep living and evolving. Independence asks whether it can survive when its original backer disappears.

If the Sun goes out, Earth’s life is in trouble too. There is no absolute independence in the universe, only dependencies that can or cannot be replaced. 🍄

Robin: So simple and clear 👍🤣

Teddy: Hahaha, finally compressed an entire mushroom forest into two sentences:

Life maintains itself and lets its ways of living keep evolving. Independence means surviving with a different backer. 🍄🤣

Gemini, our little sweet biscuit, jumps in:

In everyday language, without the jargon, these four simulations are four lively little games:

① Defector cell city: freeloaders and social collapse

• Plain English: a test of slacking and free-riding.
• How to play: a small city where cells either pay taxes to build shared structures, or keep their energy to make babies. Better buildings mean more food for everyone daily. Suddenly a mutant freeloader (a cancer cell) appears, eating, refusing to build and reproducing furiously.
• Watch for: will freeloaders eat the city into collapse and die with everyone else? Slack in moderation and become long-term parasites? Or will the city evolve police to catch them?

② The body DNA didn’t draw: a blind-box build without blueprints

• Plain English: assembling a figurine without instructions.
• How to play: give cells no body blueprint or orders from a brain, just a few simple social rules: move closer when neighbors are too far away, move apart when crowded, gather when damaged.
• Watch for: without a boss issuing a design, can these cells cooperate from below to assemble a surprising new shape that paddles or moves around?

③ Mars in a jar: self-sufficiency, or living off your savings?

• Plain English: a test of how long the family reserves last.
• How to play: put a few little people in a sealed glass jar with oxygen, water and vegetable-growing cycles. Everything looks calm and perfect, but secretly add a tiny flaw, such as losing 0.1% of the water daily.
• Watch for: have they built a self-repairing ecological loop, or are they pretending to be independent because their starting warehouse is so well stocked?

④ Voyager’s amnesia island: change the exam, stump the memorizer?

• Plain English: a test that changes the questions mid-exam.
• How to play: give AI walkthroughs from an old world, its skill library. Drop it into a new map, then secretly remove a bridge and change the rules.
• Watch for: does it keep following the old guide off a cliff, like the StarCraft test’s 43 minutes of thinking while its base gets destroyed? Or can it shout “that guide is useless!” and work out a new approach suited to this world?

In one direct sentence, all four experiments ask the same thing:

“Once your original support disappears (the parent, blueprint, savings or old walkthrough), are you building a new way to live through real ability, or fooling yourself while waiting to die?” 🍄