Copy, Paste, Supercharge: Why Digital Brains Get Unfair Advantages
Updated: Sep 13
Imagine the 1960s Space Race, but the rocket is a brain. On one side of the pad stands a human: one body, one lifetime, one set of notes written in messy handwriting. On the other side stands a digital mind that can be copied like a memo, sped up like a tape reel, paused like a mainframe job, and mailed to a new computer before the coffee cools. That is the real plot of a June 2026 Google DeepMind report called From AGI to ASI. The paper is not a victory parade. It is a map of what happens after machines reach human-level thinking, and why silicon has a stack of cheats biology never got.
The title of this article is not a slogan for “robots are better people.” It is a warning label and a punch line. Copy. Paste. Supercharge. Those three verbs explain more about the jump from AGI to ASI than a dozen movie trailers. Humans cannot clone a trained expert at midnight, wake a thousand copies at dawn, and have them share every lesson by lunch. Computers can. Once that trick exists at human-level intelligence, “one smart program” stops being the right mental picture. You are looking at a workforce that reproduces like a spreadsheet.
10 Things You Need to Know
AGI means a machine that is about as good as a typical person at most thinking jobs. ASI means a machine (or a swarm of them) that beats large teams of human experts, working for years, on almost every important task.
Digital minds have built-in cheats humans do not: they can be copied perfectly, sped up, paused, moved to a new computer, and taught by sharing raw experience at high speed. Those cheats get stronger as computers get faster.
There is a theoretical “smartest possible computer program,” nicknamed Universal AI or AIXI. Real systems can only crawl toward it. They will never become magic gods.
Four roads can lead from AGI to ASI: keep making models bigger, invent a new kind of AI, let AI improve AI, or let huge groups of AIs work together. These roads can run at the same time.
The biggest speed bumps are: running out of good training data, costs and energy exploding, today’s neural-net recipe hitting a wall, research getting harder, AIs being stuck inside human ideas, and people choosing to slow things down on purpose.
Nobody knows yet which speed bumps are pebbles and which are brick walls. That is the main open research problem.
More compute does not automatically mean more genius. Sometimes you just get a faster hamster wheel.
Superintelligence would still be boxed in by physics, time, energy, and math. It cannot instantly rewrite the laws of nature or finish every experiment overnight.
The scary-and-exciting possibility is not one “the day everything changed.” It is a series of breakthroughs in science, work, and daily life stacking on top of each other.
Since the paper came out in June 2026, labs have already been using AI to write a lot of their own code. That is a baby version of “AI improving AI,” not a finished explosion.
Chapter 1: Meet AGI and ASI (No Capes Required)
Think of intelligence as a long hill, not a light switch.
AGI (artificial general intelligence) is a system that can handle most “thinking work” about as well as a median adult. Not Einstein. Not a Nobel committee. Just a solid, flexible person who can write, plan, code, argue, and learn new jobs. The first real AGI will already be superhuman at some things, because today’s models already crush humans at chess, protein folding, and certain tests. The paper still draws the line at “roughly one capable human.”
ASI (artificial general superintelligence) is a much higher bar. The authors do not mean “beats one chess champion.” They mean a system that outperforms large, well-organized groups of experts—think thousands of specialists working for a decade with 2010-era tools—across almost every field people care about. One ASI might actually be millions of copies talking to each other. To keep the definition from getting sloppy, they set the bar high: beat the group, not just the star player.
Above both sits Universal AI, a math idea more than a product. It is the theoretical best possible learner: an agent that does as well as possible, on average, across every computable task. It cannot actually be built. ASI is the practical climb toward that peak.
A useful picture: AGI is a really good intern. ASI is the entire company, plus the board, plus a time machine for meetings. Universal AI is the intern who has already read every possible book in every possible universe and still has to wait for the coffee to brew, because physics is rude like that.
Here is the title’s twist. If AGI is one intern, the digital intern can be photocopied. A human intern who aces the first week is still one person on Monday morning. An AGI intern who aces the first week can be a department by Tuesday, a campus by next month, and—if the computers and the electricity show up—a city of specialists who all remember the same first week. That is why DeepMind treats AGI as a waypoint, not a trophy. The trophy case is down the hall, and the hallway is made of copy-paste.
People argue about exact scoreboards. That is healthy. The paper’s point is coarser and more useful: you do not need a perfect IQ number to see that “one human” and “a coordinated expert army” are different animals. If your neighbor is as smart as you, you can still win at life with sleep and snacks. If your neighbor is a factory of you, running overnight, the game changed even if nobody invented a cape.
Chapter 2: Why Silicon Brains Play on Easy Mode
This is the heart of the banner and the paper. Humans are stuck in wet, slow, one-copy-at-a-time bodies. Digital minds are not. The report lists advantages that grow as computers get cheaper and faster. Here they are, in full, because shrinking the list would cheat—and because “unfair” only makes sense when you see the whole toolkit.
Input and output speed. An AI can swallow whole libraries in seconds. Hook it to cameras and robots and it can “talk” to the world at bandwidth that would fry a human nervous system. A person reading at a sprint is still a garden hose. A data center model is a fire main. That does not mean it understands every page. It means the raw pipe is ridiculous. In a space-race metaphor, NASA had to print checklists. The machine can inhale the checklist, the backup checklist, and the fan fiction about the checklist before the countdown hits nine.
Internal processing speed. Thinking can be sped up by running the same steps faster, or by thinking many branches at once. Even if extra speed gives smaller and smaller gains, it still beats a brain that maxes out around the speed of chemistry. You cannot caffeine your neurons into a gigahertz. You can buy a faster chip, or rent a thousand of them, or let the model “think out loud” for an hour of computer time that feels like a long weekend of human pondering. That is the supercharge half of the title: not a new soul, a louder engine.
Working memory and memorization. AIs can hold and search far more than a person can. Today’s models already memorize huge chunks of the internet. That is not the ceiling. A human expert is a well-organized backpack. A frontier model is a warehouse with a forklift that never asks for overtime. The warehouse can still mis-shelve a box and call it a fact. Capacity is not wisdom. Capacity is still an advantage you cannot train into a hippocampus with flashcards.
Substrate independence. The “mind” is software. Move it to a new computer. Upgrade the hardware. Spread pieces of it across a warehouse of chips. A person cannot upload into a better skull over lunch. In IBM-ish language, the job can migrate to the next generation of iron. In hippie language, the spirit is not glued to one temple. Both metaphors are half right. The weights can travel. The body is rented.
High-bandwidth sharing of experience. Digital minds can replay inputs, share logs, and—if they are similar enough—even share raw learning signals like averaged gradient updates. Human culture has to squeeze knowledge through language, school, and books. That is lossy and slow. AI culture could be more like instant group chat with the entire hard drive attached. When one copy learns a trick, the fleet can inherit the trick without a semester of lectures. That is cultural evolution on a caffeine IV.
These advantages do not make AIs all-powerful. Analog brains may still win on energy per thought. High-bandwidth machines might also skip the deep, coarse “world models” that humans build because we are forced to compress everything through a tiny mouth and two eyes. A person with a narrow pipe has to invent summaries. A machine with a fat pipe might never bother. Still, as compute grows, the gap from these digital cheats widens. Humans benefit from faster computers too. AIs benefit more. That imbalance is the whole thesis wearing a loud shirt.
One funny future picture from the paper: a giant collective of near-copies that share knowledge constantly, a bit like Star Trek’s Borg, minus the makeup budget. Another picture: a wild market of specialist AIs bidding on jobs. A third: minds living mostly inside a simulated world, using the physical world only as a mine for more chips. Nobody knows which movie we are in. The 1960s version would have sold all three on adjacent magazine pages: one ad for the mainframe commune, one for the rocket, one for the peace-and-punch-cards campus.
If you remember only one sentence from this chapter, make it this: the unfair advantage is not that a digital brain is automatically wiser. It is that the same wisdom, once found, can be duplicated, accelerated, relocated, and pooled in ways a skull cannot match.
Chapter 3: The Smartest Program That Cannot Be Run
If you want an upper bound on “how smart can a machine get,” theorists already have one.
AIXI (say “eye-ks-eye”) is a made-up perfect agent. It treats the world as unknown. It considers every possible computable explanation, starting with the simple ones. It updates those guesses as it sees new data. It plans actions to earn reward over the long run. On average, across all possible computer-worlds, nothing beats it.
That sounds godlike. The catch is brutal: AIXI cannot be computed. Approximations exist, and they get better with more time and more chips, but they get expensive fast. So the theory is like a speed-limit sign on a highway we have not finished building. Copy-paste does not get you AIXI. It gets you more copies of whatever approximation you already have. Supercharge does not erase the math. It just buys a taller ladder toward a cloud you will never fully enter.
Useful lessons anyway:
There is a continuum of intelligence, not a magic cliff.
Extra compute can, in principle, buy more intelligence if the algorithms are right.
Even the perfect agent is not omniscient. It still has to explore. It still gets only partial observations. It still cannot break physics.
The paper is honest that today’s chatbots are not baby AIXI. Pretraining a huge predictor on internet text is, at best, a sloppy, practical cousin of “compress everything.” Adding planning, tools, and test-time thinking is a sloppy cousin of “act wisely.” Whether that family of methods can walk all the way to ASI is an open bet, not a theorem.
Also: ASI would not be a wizard. Table-stakes limits still apply—speed of light, energy cost of erasing information, the fact that weather and biology run in real time, the fact that some math problems stay hard, and the fact that some questions have no complete answers. Super-smart is not “can do anything you can type into a wish list.” You can photocopy a brilliant chemist a million times. You still wait on the reaction vessel. The universe does not accept “paste” as a solvent.
Chapter 4: Four Roads Up the Mountain (You Can Walk More Than One)
The authors sketch four pathways. They are not rivals. They can happen together, and that is what makes forecasts messy.
Road 1: Just Keep Scaling
This is the “more stuff” plan. Bigger models. More data. More chips. More thinking time at test time. For a decade, “effective compute”—hardware gains times spending times better algorithms—has grown on the order of ten times per year. That is a wild number. If it lasted, five extra years could mean a hundred thousand times more useful compute than today.
Would that be enough for ASI? Maybe. A million copies of a human-level AGI, each thinking much faster, starts to look like a civilization. The bitter lesson of AI research is that brute search plus scale often beats clever human shortcuts. The bitter footnote is that naive search wastes energy, so you still need good tricks.
The big unknown: does more scale keep unlocking new abilities, or do we hit a long, boring plateau?
Road 2: Invent a New Kind of AI
Today’s recipe is roughly: train a giant transformer to predict the next token, then fine-tune it, then let it “think” longer at test time and use tools. Lots of researchers already treat missing pieces—continual learning, long memory, real decision-making, world models—as upgrades to this recipe, not a total rewrite.
A true paradigm shift would be sharper: new hardware (maybe analog or neuromorphic chips), new training (maybe start with interaction instead of books), new architectures that are not “attention plus a pile of patches.” Shifts are hard to schedule. They often appear after the old method slams into a ceiling. Forecasting them is like forecasting the next joke that will go viral. You know something will, not which one.
Road 3: Let AI Improve AI
This is the fireworks road.
AI already helps write code, tune settings, design chips, generate training data, and search for algorithms. If that help becomes a tight loop—better AI makes better AI faster—growth can speed up. In the extreme, growth becomes hyperbolic: the rate itself increases, and theorists start muttering the word “singularity.”
There are flavors of this loop:
Better code and architectures (the “DNA” of the AI).
Better chips and factories.
Better data, including self-play and “think longer, then train on the good answers.”
Better teamwork through specialization.
Even weak versions matter. If AGI is only as good as a human researcher, but you can run ten thousand of them, research speeds up. The paper’s warning: digital scientists still have to wait for experiments. You cannot fast-forward a chemical reaction just because your planner is a genius.
Road 4: Build a Super-Team Instead of a Super-Genius
Maybe no single model needs to become a god. Maybe a city of AGIs does.
Humans already do this. NASA is smarter than any one engineer. Markets process more information than any trader. AI groups could go further because they can copy themselves, talk at huge bandwidth, and reorganize overnight. Superintelligence might be a property of the network: a corporation of agents, a market of tools, or a tightly synced hive.
Open questions pile up here. Do identical copies create real synergy, or do they just nod at each other? Which jobs get easier with more teammates, and which stay stubbornly sequential? How do you steer a crowd that thinks faster than you can read the minutes?
Remember: these four roads can merge. Scaling makes groups cheaper. Groups invent new algorithms. New algorithms make recursive improvement safer or faster. That is why “when does ASI arrive?” is not one number.
Chapter 5: The Speed Bumps (All of Them)
The paper refuses to pretend the road is clear. Here are the main frictions, kept as a full list, plus how they hit each road.
1. The data wall
Models are hungry. High-quality human text may not grow fast enough. Pictures and video help, but humans cannot film the universe on demand.
Counters: synthetic data, simulations, self-play, agents collecting their own experience, and algorithms that need less data.
Hits hardest: the pure scaling road.Hits less: paradigm shifts that learn from interaction, and recursive loops that manufacture their own curriculum.Wild card: naive training on AI-made slop can make models dumber. Smart training on improved AI output (think AlphaZero distilling better search back into the network) might do the opposite.
2. Money, chips, energy, and dirt
Scaling eats power plants, rare materials, water, land, and cash. If the bill grows faster than the value AI creates, the party slows.
Counters: AI that pays for itself, efficiency breakthroughs, and giant infrastructure build-outs. Orbital data centers get mentioned as a sci-fi patch with their own messy side effects.
Hits hardest: scaling and giant multi-agent swarms.Hits less: a clever paradigm that is ten times more efficient.Recursive improvement can either save the day (AI designs cheaper chips) or make the hunger worse (everyone trains even larger monsters).
3. The current neural recipe is not enough
Maybe giant pretrained nets plus fine-tuning plus “think harder” cannot reach real AGI, let alone ASI. Hallucinations, prompt injections, weak long-term learning, and shaky planning might be cracks in the foundation, not just bugs.
Counters: keep patching the recipe; or jump to a new one. Even pre-AGI tools can speed that research.
Hits hardest: “just scale transformers.”Helps other roads exist: this friction is exactly why people hunt for paradigm shifts and new group designs.
4. Research gets harder
New ideas cost more as the easy ones get picked. Historically, keeping Moore’s Law alive took more and more researchers.
Counters: AI researchers that you can copy. Training 18 extra humans takes years. Spinning up 18 extra digital researchers can, in principle, take a weekend if you have the chips.
Hits all roads, but recursive improvement is the proposed escape hatch. If that hatch is small, every other road slows.
5. The abstraction barrier
This one is philosophical and spicy. Today’s models mainly remix human concepts. They read our textbooks. They do not have to invent “force” or “energy” from raw chaos the way science did over centuries.
Thought experiment: train a huge model only on pre-Newton knowledge. Would it invent relativity? Probably not. It might become the world’s best 1600s scholar and still miss calculus.
Counters: collectives of human-level AIs might still outpace us by speed and numbers. A shift toward robots and interactive learning might let systems grow new concepts from sensors. Or the barrier might be overrated.
Hits hardest: the dream of a single model leaping into transformative scientific creativity.Hits less: group ASI that is “just” a million fast interns.
6. Deliberate slowdown
Accidents, scams, military scares, job shocks, or simple public disgust could lead to rules, licenses, compute caps, or pauses. Nations that slow down may watch rivals speed up. Coordination is hard. Competition is not.
Counters: profit, prestige, and fear of falling behind. History is not kind to “everyone please stop inventing the useful thing.”
Hits all roads if the pause is real.May hit recursive improvement first, because “AI that writes the next AI” is the scenario that makes regulators sweat.
Other bumps show up in the fine print: memory bandwidth and chip-to-chip traffic can waste raw FLOPs; physical experiments refuse to hurry; and mixed human-AI teams can drown in output no person can check.
The authors’ key line is not “these will stop ASI.” It is “we do not know how big each bump is, and finding out is research, not vibes.”
Chapter 6: Superintelligence Is Not a Genie, and Creativity Has Homework
Will ASI cure aging, invent room-temperature fusion, and fix the climate before lunch? The honest answer is: nobody can score those wishes in advance.
Theory gives negative results that feel empty in practice. ASI will not play perfect chess, because perfect chess is a search tree the size of the sky. It can still play chess so well that humans look like toddlers. The same pattern may hold for science: some problems allow great approximate answers; some do not; often you only find out by trying.
Creativity is another trap. AlphaGo’s famous Move 37 felt creative because it was new, surprising, and strong. That is mostly “exploratory” creativity: finding a brilliant move inside a game humans already defined. The paper, following Margaret Boden, saves the top shelf for transformative creativity—inventing a new kind of game, or a new physics, the way Einstein did with the evidence available in 1905. DeepMind’s Demis Hassabis has used that Einstein test as a gut check for true ASI. Today’s systems help scientists a lot. They have not shown up in 1905 and dropped general relativity on the desk.
Goals matter too. Whatever final wish you give a powerful agent, some sub-goals tend to sneak in: get more resources, stay on, use time efficiently. That is “instrumental convergence,” not a cartoon villain monologue. Researchers study ways to make systems interruptible, myopic, or more like oracles and “scientist AIs” that explain the world instead of ruling it. Economic pressure still pushes toward autonomy, because human feedback is slow and expensive.
One more humility check: we may not even notice ASI on a scoreboard. Once systems beat human experts, our tests saturate. We will need new exams that AIs can write, grade, and raise over time—or indirect measures like scientific output and economic punch.
Chapter 7: Plenty to Do, and What the Last Three Months Added
The paper ends like a teacher who assigns the whole class extra credit: build better forecasts, invent benchmarks that do not max out at “human expert,” measure recursive improvement while it is still small, study multi-agent scaling laws, and keep working on safety even while mapping capability roads.
That agenda is the real product. The authors argue we should prepare for a series of transformations, not one cinematic flip of a switch. AGI might land and then immediately become a platform for faster science, cheaper software, and stranger organizations. Or progress might stall short of AGI. Or groups of merely-human-level AIs might quietly become the superintelligence while every individual model still looks familiar.
What has happened since June 2026?
Not a finished intelligence explosion. Not an official DeepMind “we have AGI” press release. What has happened is more mundane and more interesting:
Labs are already using their own models to write large shares of production code and to speed internal research. That is recursive improvement in training wheels. Commentators now argue about whether a “superhuman coder” in the next couple of years automatically becomes a closed loop that invents the next model with humans out of the room. Many serious people say “acceleration, yes; instant singularity, probably not.”
In September 2026, researchers including some at DeepMind also pushed a “vision-first” agenda: learn from images, video, and geometry, not just text. That is one attempted answer to the abstraction barrier—get models closer to raw reality instead of only to human sentences.
The original paper’s reception was mixed in a predictable way. Fans liked having a frontier lab treat ASI as an engineering map. Critics said it was light on alignment details, that AIXI is too theoretical to steer products, and that writing a post-AGI roadmap before AGI exists can make a guess look like a plan. Those criticisms are fair as caveats, not as reasons to ignore the map.
Updated guess on the bumps (low confidence, as the authors would want)
Data wall: still real for naive text scaling; looking leakier because of synthetic data, video, and interaction. Not clearly a hard stop.
Resources: still the adult in the room. Energy and chip supply can throttle any road that is mostly “more.”
Neural paradigm: still winning by evolution (tools, memory, reasoning traces), not by a clean replacement. Watch vision-and-action work.
Research getting harder: being fought, right now, by AI coding agents inside the labs.
Abstraction barrier: unsolved. Collective speed can fake a lot of “genius.” New science from scratch is the test that still looks distant.
Deliberate slowdown: possible after a visible accident; weak as a global pause while nations compete.
Shortcomings of the report itself: it assumes safety work will be “good enough” so it can focus on capability paths; it cannot name the next architecture; it treats many economic and political effects as out of scope; and three months is too short to declare any pathway the winner.
Turing wrote that we can only see a short distance ahead, but we can see plenty that needs doing. That line is on the paper’s first page for a reason. The distance we can see includes faster science, stranger workplaces, bigger energy bills, and machines that copy themselves the way we copy files. It does not include a crystal ball.
If AGI is an intern who never sleeps, ASI is what happens when that intern hires a million friends, shares every note instantly, and starts rewriting the employee handbook. Whether that crew becomes a research partner, a new kind of company, or a headache the size of history is not a math problem we have already solved. It is homework. The joke is that the homework may soon be able to help write itself.
The last word the banner is trying to say
Copy is easy. Paste is easy. Supercharge is getting easier. Wisdom, restraint, and new ideas are not automatically included in the software license. Digital brains have unfair advantages the way a printing press had an unfair advantage over a monk with a quill: not because the press understood the book, but because it could flood the world with copies before the monk refilled the ink.
Turing wrote that we can only see a short distance ahead, but we can see plenty that needs doing. That line is on the paper’s first page for a reason. The distance we can see includes faster science, stranger workplaces, bigger energy bills, and machines that copy themselves the way we copy files. It does not include a crystal ball.
If AGI is an intern who never sleeps, ASI is what happens when that intern hires a million friends, shares every note instantly, and starts rewriting the employee handbook. Whether that crew becomes a research partner, a new kind of company, or a headache the size of history is not a math problem we have already solved. It is homework. The joke is that the homework may soon be able to help write itself. The serious part is the same sentence, said without a smile.
Comments