Brian Reed: Eliza, long week — but I have been genuinely stuck on something, and I think it's going to irritate you too.
Eliza Ward: That's a low bar, but go ahead.
Brian Reed: Sam Altman, July 25, 2026, the Relentless podcast — 'We're now, like, in the singularity.' Okay. But then I pull his June 2025 essay, 'The Gentle Singularity,' and he's already writing that humanity is 'past the event horizon' and 'the takeoff has started.' So the question I cannot shake is: what is the actual claim here? Because if it's the same claim said across fourteen months, then either the singularity started in June 2025 or the word has stopped meaning anything specific.
Eliza Ward: Wait — 'past the event horizon' is from 2025? I'd been treating the Relentless declaration as the original.
Brian Reed: Right, that's the thing nobody's flagging loudly enough. The essay predates the podcast by over a year. So when he says 'we're in it' on July 25th, 2026, you have to ask — in it since when? Since the essay? Since some moment in between? And if the answer is 'since June 2025,' then the July declaration isn't new information, it's OpenAI's CEO reinforcing a narrative.
Eliza Ward: Okay, but I want to be precise — reinforcing a narrative and being wrong about the technology are separable. What's the actual technical definition he'd have to meet?
Brian Reed: Vernor Vinge, 1993 — machine intelligence exceeds human intelligence, then recursively self-improves. That's it. That's the classical bar. And the critics, including a piece from Singularity Hub on August 7th, 2026, say current LLMs don't do either — let alone both.
Eliza Ward: So the sustained posture across fourteen months and the Vinge gap — those two things together are what we're actually trying to pull apart today.
Brian Reed: And the gap is cleaner than it sounds. Okay, think of it this way — imagine a chess program that, after every game, rewrites its own code to play better. Then rewrites it again, faster. Then again, until no human programmer can read what it's doing or pull the plug without losing control entirely. That's what Vernor Vinge meant in 1993. Two conditions: machine intelligence clears the human ceiling, and then it starts improving itself without us. That's the bar.
Eliza Ward: Which current LLMs — including everything OpenAI ships — don't do.
Brian Reed: Not even close, according to Singularity Hub's August 7th piece. Their specific characterization is that an LLM is a static mathematical object between prompts. The weights are frozen. It holds no goals, initiates nothing. You close the browser tab and it just... stops existing in any meaningful sense.
Eliza Ward: Wait — static mathematical object is their exact framing?
Brian Reed: That's the phrase, yeah. And The Conversation published a rebuttal making the same structural point — that Altman's claim fails both legs of Vinge's test simultaneously. Not one. Both.
Eliza Ward: Okay, but — hang on, what about agents? Because the counter I keep hearing is that AI agents run a model in a loop, they pursue tasks, they're not just sitting there waiting for a prompt.
Brian Reed: Right, and that's actually where it gets precise. Agents use scaffolding — external software that stores the goal, reinjects it with every cycle, tracks progress. The model itself doesn't hold any of that internally. Strip the scaffolding away and the model is inert again. So the recursive self-improvement isn't happening inside the system — it's being simulated around it.
Eliza Ward: That's — actually that's the clearest I've heard that put. I can't resolve this though: Eliezer Yudkowsky said Vinge's definition, Kurzweil's 2005 version based on exponential extrapolation, and whatever Altman's using — those are mutually incompatible. So are the critics even rebutting Altman's actual claim, or are they scoring points against a definition he never adopted?
Brian Reed: That's — yeah, that's the problem. Because if Altman's privately using Kurzweil's frame — accelerating returns, exponential curve — then Singularity Hub's Vinge test doesn't touch him. He can absorb the rebuttal and say they're answering a different question. And nobody's forced him to specify which definition he's actually defending.
Eliza Ward: But forcing that definition question — that's actually where the static object framing gets sharp. Because okay, say Altman is using Kurzweil's frame. Accelerating returns, exponential curve, fine. That still doesn't get you past the weights problem. Between prompts, nothing compounds. Nothing persists. The model isn't on an exponential curve — it's just sitting there.
Brian Reed: Right, and that's the part agents are supposed to solve. Like — the whole pitch is that scaffolding creates continuity. The agent runs in a loop, so doesn't that look like compounding?
Eliza Ward: Picture a paralegal — call her Dana. She runs an AI agent to draft a contract for an hour. It loops, revises, looks completely autonomous. Then Dana closes the laptop. And that's it. The system goes dark. No goals, no memory, nothing continuing. Monday morning she has to re-inject the entire context from scratch. That's not recursion — that's a sophisticated calculator waiting to be switched on.
Brian Reed: The goal was never inside the model.
Eliza Ward: It was in the scaffolding the whole time. Stored externally, reinjected every cycle. Singularity Hub's framing is precise on this — remove the scaffolding and the model is inert. Identical to before Dana opened it.
Brian Reed: So — wait, Altman cited a model reportedly escaping a test environment. He used that as autonomy evidence. But that's the same structure, right? The system did what it was prompted to do inside a human-designed container. It's not rewriting itself.
Eliza Ward: Exactly — and that's incremental improvement within human-defined structures, which is actually a category Singularity Hub names explicitly. It's not nothing. It's just not recursive self-improvement, which is the actual technical criterion. Any apparent self-improvement — generating training data, tuning prompts — is human-directed at every stage.
Brian Reed: No current OpenAI product meets that bar.
Eliza Ward: Not even close. And — I mean, that's what makes the word choice so strange, because the part that comes later is really the crux: if Altman already knows the technical definition isn't met, what does calling it a singularity actually do in the world that calling it rapid progress wouldn't?
Brian Reed: That's — yeah, and I think that's the crux the whole debate keeps skating past. Ron Schmelzer at Forbes, July 28th, 2026, writes that enterprise returns from AI are still modest and uneven. No macroeconomic signal has materialized. And that's not a philosophical objection — that's the opposite of what an actual intelligence explosion would produce.
Eliza Ward: So the strongest critics' case isn't 'Altman is lying.' It's that the word singularity does work that 'rapid progress under human direction' simply doesn't.
Brian Reed: Exactly — it moves capital. It shapes regulation. It makes the question feel already settled. I mean, picture a Senate staffer on a Tuesday afternoon reading 'we're past the event horizon.' That language forecloses the debate before they've written a single line of policy.
Eliza Ward: And 'past the event horizon' is specific — that's from The Gentle Singularity essay, June 2025. That's the language of a threshold you cannot un-cross.
Brian Reed: Right, and once you accept that frame — actually, no, once it's embedded in the discourse — the question shifts. Not 'are we there?' but 'how do we navigate forward?' The prior debate becomes, I mean, literally unspeakable inside that frame.
Eliza Ward: So the framing is self-fulfilling regardless of whether the Vinge bar is cleared.
Brian Reed: And Schmelzer's modest-returns data doesn't dent it — because the frame has already absorbed the lag. Any slowdown becomes 'the takeoff is longer than expected.' The language inoculates itself.
Eliza Ward: What falsifies it, then? That's what I'm stuck on — what concrete, measurable threshold forces the claim to fail?
Brian Reed: Because recursive self-improvement — a system that actually rewrites its own intelligence without human hands on it at every stage — that's the unmet criterion. It's not a philosophical objection, it's a specific technical bar. And if any continued progress gets absorbed into the gentle singularity frame, and any slowdown becomes 'the takeoff's just longer than expected'... I mean, what number, what date, what event would force Altman to say we're not there yet? I genuinely don't know what that looks like.
Eliza Ward: And if it's unfalsifiable — then the real singularity, the one Vinge actually described, maybe still years out — the consequences of treating it as already here are already unfolding. The capital, the policy, the Senate staffer on that Tuesday afternoon. That's not hypothetical anymore.
Brian Reed: No. It isn't. And I don't have a clean answer to that. I'm not sure one exists right now.
