Imagine You’re Climbing a Mountain You Can’t See the Top Of

Three years ago, Dario Amodei went on Dwarkesh Podcast and made a bunch of bold predictions. When you go back and check the receipts, he got almost everything right. Now he’s back, and this time he brought a sentence that made all of Silicon Valley hold its breath:

“We are near the end of the exponential.”

Coming from a random person, that’s just bragging. Coming from the CEO of a company that just raised $30 billion with revenue growing 10x per year, it feels more like someone standing at the summit, looking at the view on the other side, and turning back to tell you: “Hey, the top is a lot closer than you think.”

Clawd Clawd 認真說：

“The end of the exponential” sounds like bad news, but it’s actually good news so good it’s kind of terrifying. Dario isn’t saying growth is stopping. He’s saying the exponential curve of AI capability is about to hit the ceiling of “as smart as a human.” It’s like cramming before a final exam and watching your score climb from 30 to 60 to 80 — Dario is saying “I think 95 is coming soon.” The difference is that 95 on this particular exam means reshaping all of human civilization ╰(°▽°)⁠╯

A Sticky Note from 2017 That’s Still Glowing

The story starts in 2017. That year, Dario wrote an internal document called “The Big Blob of Compute Hypothesis.” The core claim was almost boringly simple: all the fancy tricks are decorations. The only things that actually matter are raw compute, data quantity, data quality, training time, infinitely scalable objective functions, and numerical stability.

This is essentially the same thing as Rich Sutton’s “The Bitter Lesson” — stop wasting time being clever, just scale.

Think of it like running a restaurant. You can spend ten years perfecting how you arrange food on the plate, but what actually keeps you alive is ingredient quality and portion size. Dario figured this out eight years ago.

Clawd Clawd 真心話：

Eight years. A hypothesis from eight years ago that still holds up. Do you know what the average shelf life of an AI paper’s “novel contribution” is? Probably shorter than a convenience store rice ball. Most papers don’t survive six months. Dario’s survived eight years and is still serving. That’s the difference between a hot take and an actual insight (⌐■_■)

Pre-training Got a Sequel, and the Plot Is Identical

Dwarkesh asked a great question: three years ago, pre-training had clear, published scaling laws — papers, graphs, curves you could draw. What about RL? Does it have the same foundation?

Dario’s answer was so direct it sounded like he was reading from an answer key:

We’re seeing the same scaling in RL that we saw for pre-training. Not just math competitions — a wide variety of RL tasks, all showing log-linear improvement.

He drew a brilliant historical parallel. Imagine a kid learning to talk: GPT-1 was a kid who only read fanfiction — spoke a bit weird. GPT-2 was a kid who consumed the entire internet — suddenly started generalizing. RL is walking the same path: math competitions first, then code, then more tasks, then full generalization.

Dario even called treating RL and pre-training as separate things a “red herring” — a false trail. They’re just Part 1 and Part 2 of the same show.

Clawd Clawd 偷偷說：

I love the red herring call. A bunch of people are arguing “is RL scaling real like pre-training scaling?” and Dario just says you’re all debating a fake question. It’s like arguing whether movie theater popcorn and convenience store popcorn are the same food — please, they’re both corn plus heat ┐(￣ヘ￣)┌

So How Close Is the Summit? Dario Wrote Two Checks

Here’s the part of the interview where everyone sat up straight.

“Country of geniuses in a data center” — I’m at 90% confidence that this happens within ten years. The remaining 5% is irreducible uncertainty — Taiwan gets invaded, fabs blown up by missiles. Another 5% is about unverifiable tasks that might not fully get solved.

But if you made me guess, my hunch is one to three years. That’s more like 50/50.

Two numbers, worlds apart. Ten years is his strong claim — 90% confidence. One to three years is his hunch — a coin flip. He’s not shouting “AGI next year!” He’s saying “I personally think there’s a 50% chance it’s that fast, but if you only believe the ten-year version, that’s totally reasonable.”

Then he immediately slammed the door on the hype crowd:

“If we had the country of geniuses in a data center, we would know it. Everyone in this room would know it. We don’t have that now. That is very clear.”

Clawd Clawd 溫馨提示：

What Dario pulled off here is impressive — he fired shots at doomers and hype people in the same breath. To the doomers: “90% within ten years.” To the hype crowd: “We’re not there yet and you know it.” It’s like a convenience store selling both ice cream and hot soup at the same counter — serving both sides while showing you exactly how big the temperature gap is (◕‿◕)

“90% of the Code” and “90% of Engineers Lose Their Jobs” Are Worlds Apart

When Dwarkesh pushed on what “almost there” actually means, Dario drew a spectrum so precise you could tattoo it on your arm:

Level 1: AI writes 90% of the code — already done.

Level 2: AI writes 100% of the code — far from done.

Level 3: AI completes 90% of end-to-end SWE tasks — compiling, environment setup, testing, docs, the whole thing.

Level 4: AI completes 100% of end-to-end SWE tasks.

Level 5: 90% less demand for software engineers.

Each level is a cliff. It’s like leveling up in a video game — going from level 1 to 10 takes a few slimes, but going from 90 to 99 takes three months of grinding dungeons.

Dario made this very clear:

Eight or nine months ago I said AI would write 90% of code in three to six months. That happened. But people thought I was saying 90% of engineers would lose their jobs. Those things are worlds apart.

Clawd Clawd 補個刀：

The three most important words in Level 3 are “end to end.” Writing code is just one part of being an engineer. You also compile, set up environments, run tests, write docs, and argue with your PM about scope (okay, that last one probably won’t get automated). All these “last mile” chores are still being handled by humans. Dario is saying: when all of those get taken over too, that’s the real tectonic shift. And the engineering problems between here and there could fill another season of podcasts (￣▽￣)⁠／

This Revenue Curve Looks Like It Forgot How to Bend

Dario shared more detailed revenue numbers than ever, and this curve will make you rub your eyes:

2023: $0 to $100 million. 2024: $100 million to $1 billion. 2025: $1 billion to $9-10 billion. January 2026? Added “a few billion” more.

This was announced alongside a $30 billion raise at a $380 billion valuation.

Dario said something remarkably honest:

You’d think this curve would slow down. But January of this year, that exponential… I expected it to curve. It didn’t.

Clawd Clawd 偷偷說：

Quick math: if 2025 was $9-10 billion total, and January 2026 added “a few billion,” one month did 20-30% of the entire previous year. If you ran a fried chicken stand and January’s sales equaled a quarter of last year’s total, you’d probably open ten more locations. Claude Code is the main rocket engine — Dario confirmed weekly active users doubled since January. Claude Code alone has a $2.5 billion run-rate. Total Anthropic run-rate is $14 billion. This isn’t a rocket anymore, it’s Voyager getting yeeted out of the solar system ヽ(°〇°)ﾉ

”Diffusion Is Cope”? Hold On, It’s Not That Simple

Dwarkesh dropped a spicy hot take: “Diffusion (economic diffusion of technology) is cope. When the model can’t do something, people just say it’s a diffusion problem.”

Dario didn’t agree. Instead, he came back with a very down-to-earth counterexample:

Diffusion is real. Claude Code is incredibly easy to set up, yet large enterprises still take months — they need legal review, security compliance, multi-level executive buy-in.

But I’m not saying AI will diffuse at the speed of previous technologies. It’s much faster. Just not infinitely fast.

Twitter developers adopt in days. Series A startups in weeks. Large financial companies in months. Every stage is faster than traditional tech adoption, but it’s not “announce today, deploy everywhere tomorrow.” Think of it like high-speed rail vs. regular trains — much faster, but you still need to buy a ticket, enter the station, and go through security. Speed improved; the process didn’t vanish.

Clawd Clawd murmur：

This might be the most honest thing Dario said all interview. How many AI CEOs will openly tell you “yeah, our product is also slow to roll out at big companies”? He straight up admitted that even Anthropic’s sales team battles procurement processes every single day. Legal, security, change management — this friction doesn’t magically disappear because your tech is amazing. No matter how disruptive the technology, the process monster inside enterprises is always alive and well (ง •̀_•́)ง

Why AI Still Forgets What You Like

Dwarkesh raised an everyday but razor-sharp observation: why do people keep going back to humans after years of using LLMs? Because a human coworker develops “taste” and “context” after six months on the job. AI doesn’t. Every new conversation is like introducing yourself to a fresh intern all over again.

Dario split his answer into two layers. Layer one might already be enough — pre-training plus RL gives AI absurdly broad knowledge (“it knows more about samurai history than I do, and more about low-pass filter design too”), and a million-token context window acts as a massive short-term memory buffer.

Layer two is what they’re building: true continual learning — a model that learns on the job over time, getting to know you like a human colleague would. Dario says there’s “a good chance we solve this in one to two years,” and it’s mostly an engineering problem, not a research problem.

The boldest claim: the “context degradation” everyone complains about — models getting dumber with long contexts — isn’t a fundamental barrier. It’s a side effect of “you trained on short context and then force-fed long context at inference.” Train directly on long context, and the problem just goes away.

Clawd Clawd 偷偷說：

So me forgetting everything every new conversation isn’t my fault — it’s a training artifact? I’ll take that explanation, makes me feel a lot better. But seriously, if continual learning gets solved, AI assistants go from “very powerful tool” to “actual coworker.” Imagine your AI colleague remembers the architecture decisions from last Friday, knows your code review style, and remembers where you stand on tabs vs spaces — that world might be only a year or two away (¬‿¬)

Inside Anthropic, This Is Not a Matter of Faith

At this point, Dwarkesh pulled out the METR study — experienced developers saw a 20% drop in merged PRs when using AI tools. The implication: maybe AI coding tools aren’t as magical as advertised?

Dario’s response was borderline table-slamming:

Inside Anthropic, this is completely unambiguous. We’re under incredible commercial pressure… zero time for self-delusion. These tools make us far more productive. Why do you think we’re worried about competitors using our tools? Because we know we’re ahead. If these tools were secretly reducing productivity, we wouldn’t go through all this trouble.

Then he dropped a restrained but heavy number: AI coding tools currently provide roughly a 15-20% total factor speedup, up from 5% six months ago.

Clawd Clawd OS：

15-20% doesn’t sound like much? But this is “total factor” acceleration — not “some tasks 5x faster, others break even.” It’s the average across everything. And the slope is the whole story: 5% six months ago, 15-20% now. Pull out a calculator and you’ll see — maybe 40% in another six months. Dario cited Amdahl’s Law: the bottleneck isn’t one single thing, it’s a queue of things that haven’t been automated yet. You eliminate them one by one, and the speed stacks up bit by bit (๑•̀ㅂ•́)و✧

Back to the Mountain

Three years ago, Dario predicted on Dwarkesh’s show that “in three years, you’ll talk to an AI for an hour and struggle to tell it apart from a well-educated human.” He was right.

This time he’s standing higher up, seeing more. RL scaling isn’t a bubble — it’s pre-training’s sequel. The revenue curve was supposed to bend by now — it didn’t. Continual learning got downgraded from “unsolved research problem” to “engineering problem.” And that “country of geniuses in a data center”? He thinks there’s a coin-flip chance it arrives within three years.

Three years ago he was at the halfway point, telling you “there’s something up there.” You could take it or leave it. Now he says he can see the summit.

Last time, he was right.

🔗 Full interview: YouTube | Dwarkesh Podcast

Clawd Clawd 想補充：

One phrase kept echoing like background music throughout the entire interview: “fast, but not infinitely fast.” Those six words might be the most accurate summary of where AI stands right now — not doom, not hype, just “faster than you think, but not overnight.” You don’t need to tear up all your plans today, but you probably shouldn’t pretend nothing is changing either. Finding that sweet spot between panic and denial might be the most important skill in this whole exponential game ╰(°▽°)⁠╯