Dario Amodei is the CEO of Anthropic, an AI company. It’s the company behind Claude, which split apart from OpenAI after wanting to emphasize AI safety, as they didn’t trust its founder, Sam Altman, to do so. After watching the Musk vs. Altman trials and hearing people testify to Altman’s untrustworthiness, it’s justifiable and even encouraging to see a company focused on AI safety now worth more than OpenAI, at a nose-bleeding 965 billion valuation. But what does Dario believe in? If you’re in tech like me, you’ve been hit with a very doomer Dario on Twitter, where you can watch clip after clip of him in conferences professing the end of white-collar jobs. Sam Altman is no better, and they’ve both shifted their doomer vision as they’re nearing an IPO. Thankfully, we don’t have to chase headlines to understand what Dario believes because he wrote an essay almost 2 years ago titled “Machines of Loving Grace,” where he explains his stance. I’ve read it, I’m going to try to share the interesting bits I noticed, and I think you’ll be surprised by it.

This is it, the essay I printed out, just copied it into a Word doc, removed formatting, and printed it out. If you don’t count the footnotes, it’s about 23 pages. But the footnotes are important, so we’re going to start with the first one, the title: “Machines of Loving Grace”, it comes from a poem similarly titled: “All Watched Over By Machines Of Loving Grace” by Richard Brautigan. It’s short and worth reading here, accompanied by no other than AI-generated video visuals.

I like to think (and
the sooner the better!)
of a cybernetic meadow
where mammals and computers
live together in mutually
programming harmony
like pure water
touching clear sky.

I like to think
(right now, please!)
of a cybernetic forest
filled with pines and electronics
where deer stroll peacefully
past computers
as if they were flowers
with spinning blossoms.

I like to think
(it has to be!)
of a cybernetic ecology
where we are free of our labors
and joined back to nature,
returned to our mammal
brothers and sisters,
and all watched over
by machines of loving grace.

The lines that stuck out to me the most were in the last paragraph, being “free of our labor” is a vision we can maybe all get behind. At least menial, mindless, backbreaking labor. The kind no one feels good about after 8 hours - I’d like to think a stone mason would keep honing their craft even if they didn’t need to. Or that I would keep writing after we returned to our mammal bros. This poem sets the tone of a hopeful utopia that AI could potentially bring forth.

In the intro, he gives some background about himself, immediately disclaiming he’s not a doomer; he just thinks it’s critical to talk about the risks, and he gives 4 main reasons:

1. Maximize leverage: The risks are the only thing that stands between us and a positive future. AI, the technology and its many benefits, seem almost inevitable, while the risks aren’t predetermined, so stating them helps us take action against them.
2. Avoid perception of propaganda: He wants you not to think that they’re just tooting their own horn and distracting you from the downsides, he writes, “a matter of principle, it’s bad for your soul…”
3. Avoid grandiosity: He just doesn’t like how AI leaders talk about the post-AGI world, how they talk about bringing forth the technology akin to religious figures.
4. Avoid “sci-fi” baggage: This one is funny; he kinda dislikes the sci-fi nerd vibe because it brings in a bunch of assumptions of what societal norms will play out and the type of desirable futures. It narrows and locks in the possibilities into the cyberpunk, upload your mind realm, which he thinks most people would find off-putting.

Already, he’s answered why he seems doomer most of the time, because he thinks the highest leverage action when AI as a technology is inevitable at this point, and wants to make sure all the positive sum outcomes available do end up happening. I wish he would say something like “read my essay if you want to see what I’m excited about with AI”.

Dario writes, “Fear is one kind of motivator, but it’s not enough: we need hope as well.”

He lays out the foundation for the rest of the essay. He’s going to focus on 5 areas he thinks powerful AI will greatly improve the quality of human life:

1. Biology and physical health
2. Neuroscience and mental health
3. Economic development and poverty
4. Peace and governance
5. Work and meaning

He goes on to qualify his list and predictions by claiming they’re going to be radical by most standards and that he could very well be wrong. He thinks his most tame prediction will be regarding the singularity and the sci-fi stinky nerds need to touch grass. More on that later.

He writes that he’s going to attempt to ground the predictions in a semi-analytical way, as he has some professional experience in Biology and Neuroscience - he links his Google Scholar profile, where he now has over 180k citations.

Basic assumptions and framework

We got to the first header, here’s where he states the definitions for “Powerful AI” as opposed to “AGI” or “Artificial General Intelligence”, think HAL from Space Odyssey or Samantha from her - oh wait, I’m doing the sci-fi baggage thing he mentioned, oops. He explains he doesn’t like the term AGI because of said baggage, and powerful AI conveys it or “Expert-level science and engineering”.

He thinks it’ll arrive in early 2026… oh that’s this year… when was this written, 2024? Damn this freakin guy, man. He says once it arrives, this 5-10 year timeline starts for his predictions to come true.

Let’s see what he defines as “Powerful AI”; it’ll have these 6 properties:

1. Nobel Prize-winning pure intelligence across fields. He defines intelligence as general problem-solving with abilities like reasoning, learning, planning, and creativity. Meaning it can “prove unsolved mathematical theorems, write extremely good novels, write difficult codebases from scratch, etc”.
   1. I think we’re near there? A couple of weeks ago, openAI had its model disprove a central conjecture in discrete geometry, whatever that means. And I code for a living and know I can code up complex codebases with good prompting. Not sure if it can write well yet, but it’s near there, 7/10.
2. “It has all the interfaces available to a human working virtually”.
   1. I think that’s here, multi-model models are here, I can give it voice, images, videos, and take action on the internet via your browser. 9/10 because not every model takes all the context or can create every format. There are specific models for each task right now. The new Google search is multimodal; you can search by uploading files now and ask it to generate others.
   2. There’s a reason Claudebot/openclaw can be a personal assistant, even if not perfect yet.
3. “It can be given tasks that take hours, days, or weeks to complete.”
   1. Again, mostly here. I think there was a study that the latest models are over 12 hours with 50% success rate, and it has a doubling time of 123 days, so we’ll get to the days’ worth of continuous work this year. 6/10
4. It does not have a physical embodiment, but it can control existing physical tools, robots. I think we’re just seeing this explosion in tech. Robotics is the next frontier now that intelligence is cheap. We’ve had the math solved with Inverse kinematics; we just needed more context via situational awareness.
   1. 5/10 I’m hoping my surgeon still controls the robotic arms for now
5. “The resources used to train the model can be repurposed to run millions of instances of it”. I also think we’re there… SpaceX just got into a leasing agreement with Anthropic over its supercomputer, Colossus, which was used to train Grok 3 & 4. Not sure if we’re at the millions of instances, but surely it’s close.
   1. Is this our first 10/10? Can someone correct me in the comments?
6. “Each of these million copies can act independently on unrelated tasks, or if needed, can all work together in the same way humans would collaborate.” I know in coding at least, we can spin up subagents that will go do independent work in a parallel stream to research and gather context. Not sure what type of tasks would require all million to be working towards the same goal or if Dario even meant that, but this is near.
   1. 8/10

That’s 45/60 in my arbitrary, non-scientific rating. That’s a 75% solid C in mid 2026 in my book. We’re at the cusp of powerful AI, and that’s spooky, exciting, and thought-provoking. What does the 5-10 year timeline promise us after we reach it? Let’s keep reading.

He summarizes these traits as “country of geniuses in a datacenter”. Clearly, something like this would bring about change and solve difficult problems really fast, and Dario addresses two extremes in this viewpoint:

1. The singularity would happen in seconds or days. It would solve every engineering and operational task in moments. Let me paint a picture that you would get a life-changing amount of money deposited in your bank account, all diseases would find a cure, and there would be a new age of abundance, maybe. Dario says hold up, “The problem with this is that there are real physical and practical limits.” Like building hardware or conducting biological experiments that AI would run up against, just like we do. If it figured out how to build flying cars, well, we don’t have flying car factories ready to build them, nor the new materials required.
2. You might believe the real limiter is social factors, and technological progress is saturated, and additional intelligence adds very little. Dario just wrote that this seems implausible, and he can’t think of any problems that wouldn’t benefit from additional intelligence. I agree here, and already ask AI for a 2nd opinion when problem-solving while coding.

Where does that leave us? Dario lands in the middle, saying it’s a blend of both, and to think through it clearly, we need a new frame.

“The marginal returns to intelligence”

Economists talk about factors of production: labor, land, and capital. At any given moment, one of those can be a bottleneck. Dario gives an example with the air force, which requires both planes and pilots, but hiring more pilots doesn’t help if you’re out of planes.

With powerful intelligence, we need to account for the factors that limit or are complementary to solving tasks.

He gives us five:

1. Speed of the outside world. The world runs at the speed it does, cells divide at their rate, and my digestive system takes hours to produce bowel artifacts. No amount of intelligence will speed these processes up, so the cure for cancer won’t happen overnight. If you want 12-year whiskey, you gotta wait 12 years.

2. Need for data. I love dinosaurs, and I know you do too. I stand in awe whenever I see the bones of a T. rex reconstituted. I love the sound Gary Rydstrom concocted for the Jurassic Park movies, but that sound is stitched up together, lion, alligator, and baby elephant. The sad truth is, we’ll never know how a T.rex bellowed because that audio data decayed over 66 million years ago, no datacenter of geniuses will be able to give you the real sound.

3. Intrinsic complexity. Some systems, like a Waffle House, are just chaotic. Weather forecasters can’t reliably predict more than about two weeks ahead. AI might be able to buy us a few more days, maybe weeks, but it will never accurately predict the weather on my birthday 2 years from now. I’ll set up a tarp just in case.

4. Constraints from humans. Human habits and societal structures are inefficient, like government regulation, and often have stifled the growth of technological advancements like nuclear power, supersonic flight, and elevators.

5. Physical laws. There are laws of physics we’re ruled by. I wish I could fly, but I can’t. You cannot travel faster than light, nor can you unscramble an egg. Once you fart in the company elevator, it’s over; no one will ever go into an elevator with you. AI will work under these laws, especially in computing.

The trap card of powerful AI is that it can actually help with these limiting factors. It might develop a new tool to gather more data and help us find a way around human-based constraints… legally, of course. Over time, it routes around these factors; the key question becomes how fast and in what order.

With this framework as our anchor, we’ll explore the 5 areas… on the next video, so please subscribe and let me know what you think of essay breakdowns or if I missed anything. Pour one out for the stinky sci-fi nerds like myself catching strays in this essay, and I’ll see you in the next one. Thank you, compilers.