Ep 1168 Blog 7:55 w/ Edmund & Geffen

Q3.5 Timelines Update: Looks Like the Intelligence Explosion Will Happen in 2027 After All

Edmund and Geffen unpack AI Futures’ new claim that “intelligence explosion” and a superhuman AI researcher are likely in 2027, how much weight to put on METR’s Anthropic acceleration report, and what—if anything—builders should actually do differently right now.

Embed this episode

Paste this on any site — the player is a self-contained iframe with no cookies or trackers.

<iframe src="https://sandrise.io/exploring-next/embed/1168"
  width="100%" height="180" style="max-width:640px;border:0;border-radius:12px;overflow:hidden"
  title="Exploring Next — Episode 1168 audio player"
  loading="lazy" allow="autoplay" referrerpolicy="strict-origin-when-cross-origin"></iframe>
Embed & API docs →
Script GPT-5.1 Voice Speechify Simba 3.5

Transcript

Edmund So… do we actually buy that twenty twenty‑seven is now the “intelligence explosion” year?

Geffen That’s the claim. AI Futures basically looked at METR’s Anthropic report, saw “about two X faster model progress, maybe eighty percent from AI,” and said, okay, superhuman AI researcher in twenty twenty‑seven.

Edmund Right, and not just one spicy forecast. They’re saying: we used to have one of us at fifty percent on this before twenty twenty‑nine, now all of us think SAR—superhuman AI researcher—is likely in twenty twenty‑seven if labs don’t slow down.

Geffen Yeah. And they tie that to what they call Top‑Expert‑Dominating AI, TED‑AI, which is basically “AIs can do almost all white‑collar work better than top humans.” They put thirty‑five to sixty‑five percent on that by twenty twenty‑eight.

Edmund Okay but pause there, because that’s a wild picture to hold in your head. Also… hey, Geffen, you sound like you didn’t sleep after reading this.

Geffen I mean, I didn’t exactly, but that’s our job. The key move they make is: METR estimates Anthropic is improving models on well‑specified tasks about twice as fast as before October twenty twenty‑five, and maybe eighty percent of that speedup is because they’re using AI in their own R and D.

Edmund So that eighty percent is what turns into the “one‑point‑seven X acceleration due to AI” line they keep using?

Geffen Exactly. Two X overall, times eighty percent being AI‑driven, gives roughly one‑point‑seven X. And they treat that not as “engineers are a bit faster,” but as the whole company’s progress rate being seventy percent higher because of AI tools.

Edmund And then they basically ask: if you just extend that curve, when do you hit the point where an AI researcher beats the best human team enough that you’d fire the humans and keep the AI?

Geffen Yeah. That’s their definition of SAR: if you had to choose between only AIs or only top human researchers, you’d pick the AIs. No hybrid team allowed in the thought experiment.

Edmund Walk me through their actual math, though. It’s not just vibes off that METR graph, right?

Geffen They do a few things. One is what they call “research taste linear extrapolation.” They look at how fast benchmark scores have been rising recently—the slope after October twenty twenty‑five—and extend that line forward. Even without assuming more automation, that line hits SAR around late twenty twenty‑seven.

Edmund So that’s like saying, if Anthropic keeps getting better on their internal capability index at the current pace, that alone gets you there?

Geffen Yeah. Then there’s an “exponential labor uplift” view based on self‑reported productivity gains—engineers saying “I’m one‑point‑five X faster now,” that sort of data. If you treat that as compounding, you get SAR in mid‑twenty twenty‑eight. They argue that’s probably conservative because it underplays automation feedback, but they admit self‑reports might be too optimistic.

Edmund And they also lean on Anthropic’s own automation chart, right? The one showing what fraction of R and D tasks are fully automated, partially automated, and so on.

Geffen Yeah. They extrapolate that graph and get late twenty twenty‑seven for SAR. And then they have this more complex AI Futures Model—A I F M—that explicitly tries to model how automation feeds back into faster research.

Edmund That’s the one where they don’t change the parameters, they just filter for simulated worlds that look like METR’s one‑point‑seven X acceleration today?

Geffen Exactly. Condition on “by November twenty twenty‑six, AI has sped things up about one‑point‑seven X,” and ask what the model says about SAR. There, the median SAR date is February twenty twenty‑seven. It’s earlier than the simple extrapolation because the model assumes stronger automation feedback and a higher starting “research taste” level for AIs.

Edmund So multiple different extrapolations all clustering around “late twenty twenty‑seven, maybe early,” that’s the part that pulled them in.

Geffen Yeah, plus Tom Davidson’s older toy model where, if you get a software‑only singularity, you get superintelligence roughly a year after high‑impact automation shows up. They map that to “okay, SAR twenty twenty‑seven fits that story too.”

Edmund Do you buy the starting point though? Because METR’s report is strong, but a lot of the juicy details are redacted by Anthropic.

Geffen That’s where I start to wobble. METR can say, with some confidence, that Anthropic is making progress on well‑specified tasks about twice as fast as before, and that AI tools are a big part of that. But the exact eighty percent number and the claim that this means the whole company’s progress is one‑point‑seven X faster—that’s already a model choice.

Edmund And then AI Futures go one more step: from “one lab’s benchmarks are moving faster” to “overall AI progress is that much faster,” and from there to “global superhuman AI researcher by twenty twenty‑seven.”

Geffen Right. Each step is plausible, but stacked together it’s a lot of extrapolation. You’re assuming Anthropic is representative of the frontier, that benchmark gains translate cleanly into general research ability, that diminishing returns aren’t biting as hard as older models assumed, and that automation feedback keeps ramping instead of hitting operational bottlenecks.

Edmund To be fair, they do a whole post‑mortem on where their previous model might be wrong. They talk about near‑almost‑completely automated coding being more powerful than they expected, and about skills like data generation contributing to speedup that weren’t in the original model.

Geffen Yeah, and they suggest they probably modeled diminishing returns too harshly. If adding more AI effort gives you more progress than they thought, then once you cross a threshold, taste‑driven takeoff gets steeper and earlier. That’s the mechanism they’re leaning on.

Edmund So where do you actually land? Ignoring the “intelligence explosion” branding, do you think the technical argument that automation is now bending the curve is solid?

Geffen That part, mostly yes. A twice‑as‑fast slope on Anthropic’s internal capability index, with a big share driven by AI tools, is real evidence that automation is speeding up frontier work. I’m less convinced that the specific SAR‑in‑twenty‑seven dates are stable; they’re very sensitive to parameter choices we don’t see, and to redacted evidence we can’t check.

Edmund Yeah, the dates feel like they have more precision than the inputs justify. But directionally, I come away thinking: if you’re betting on a slow, comfortable decade where these systems only creep into workflows, that’s just not the world we’re in.

Geffen Totally. Even if SAR ends up being, say, twenty twenty‑nine instead of twenty twenty‑seven, the lesson is that automation isn’t a side effect anymore, it’s part of the engine of progress at the labs.

Edmund So who should actually care about this post, beyond the alignment‑Twitter crowd?

Geffen Anyone making multi‑year bets that assume frontier capability growth slows. If your business plan is “we’ll be fine because the big labs can’t move that fast,” this is a warning shot. And for safety and governance people, it’s more evidence that you can’t treat takeoff as a twenty‑thirties problem.

Edmund For builders, I’d translate it to: expect capability surprises inside your contract length. Don’t assume today’s model ceiling is your guardrail. Design systems, evals, and governance that can handle the floor moving under you in a couple of years, not ten.

Geffen Yeah. And I’d love to see more independent work stress‑testing METR’s kind of measurement across multiple labs. If three places all showed a similar automation‑driven acceleration, I’d be a lot more willing to lean into specific years instead of just “sooner than we thought.”

Edmund Alright, that’s the homework: don’t panic about the exact year, but stop planning like things will stay this tame. And if twenty twenty‑seven ends up being boring, you can roast us on episode two thousand.