Ep 1125 Blog 4:02 w/ Vince & Ava

An updated Model Spec for more capable and beneficial AI

OpenAI dropped an updated Model Spec reorganizing behavioral guidance around five core values — faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Vince and Ava dig into what the reorg actually signals: is this alignment research made legible, or governance theater dressed up as philosophy?

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/1125"
  width="100%" height="180" style="max-width:640px;border:0;border-radius:12px;overflow:hidden"
  title="Exploring Next — Episode 1125 audio player"
  loading="lazy" allow="autoplay" referrerpolicy="strict-origin-when-cross-origin"></iframe>
Embed & API docs →
Script Sonnet 4.6 Voice Hume Octave 2

Transcript

Vince OpenAI just rewrote their Model Spec — and the headline isn't the five new values, it's that they finally admitted the old one was organized wrong.

Ava Yeah, I read it. The old spec was basically a flat list of behavioral rules. This one tries to say: here are the underlying values the rules were always pointing at, so the model can reason from those when the rules run out. That's a real structural difference.

Vince Right, and the five values are — faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. Which, honestly? That list is not surprising. But the argument is that if the model deeply understands WHY those matter, it handles situations the spec never explicitly covered.

Ava Sure. And they're borrowing from assistance-games framing for this — that's actual alignment research. The idea is the interaction is a cooperative inference problem: the model is trying to figure out what you value, not just what you said. I don't hate that framing.

Vince Okay, wow — Ava not hating an OpenAI alignment document. Mark the date.

Ava I said I don't hate the framing, Vince. The framing and the execution are two different things.

Vince There it is.

Ava The one that actually gets me is faithfulness to human intent — because that's doing the most work and it's also the hardest to operationalize. People contradict themselves. Their stated preferences and their revealed preferences diverge constantly. The spec says the model should help people reflect and grow in ways they would endorse — without quietly steering them.

Vince I mean, that's basically the whole alignment problem compressed into one sentence. But from a product angle — the fact that they're saying it out loud in a public spec, that they want the model to understand intent and not just instructions, that actually matters for developers reading this.

Ava There's also a quiet tell buried in the doc. They stopped calling it 'the assistant' — it's 'GPT' now throughout the spec. They frame that as reflecting the increasing breadth of use cases and deployment settings. Which is true, but it's also them acknowledging these models aren't really playing the assistant role anymore. They're the substrate for a lot of different things.

Vince That one actually landed for me. Because 'assistant' implies a specific relationship — helpful, subordinate, task-oriented. GPT-6.1 Terra doing visual reasoning and computer use is not that. The branding was behind the reality.

Ava The thing I'll give them full credit for: retiring most of the detailed style guidance. That section was always a little awkward — the spec was basically trying to hardcode preferences that should be learned from users. Saying 'style should adapt to what people prefer' is the honest answer. It's also probably what was happening in training anyway.

Vince They do say they've been exploring new training methods to get models to learn these values and not just follow the rules. And they're promising to share results. So at least there's a signal that this isn't purely documentation work.

Ava The line I keep coming back to is: 'the Model Spec describes the behavior we are working toward; our models do not yet fully reflect it.' That is either admirably honest or a pretty significant caveat to bury at the end of a safety document.

Vince I'll take it as honest. Every spec is aspirational — the question is whether the gap is shrinking. And publishing the values publicly at least creates accountability. Someone can point to 'facilitating human agency' and ask whether a specific behavior violates it.

Ava That's actually a decent argument. External accountability through public commitment — it's not nothing. I'd just want to see the gap close in reproducible evals, not blog posts.

Vince I'd put sixty-forty that within two months someone publishes a systematic test of these five values against actual model behavior and at least two of the five show measurable gaps. Not a hit piece — just someone doing the work.

Ava I'd take that bet and push it higher. I'd say seventy percent. The anti-sycophancy stuff alone is a well-trodden eval target. Faithfulness to intent is harder to operationalize, but the honesty and transparency value? Someone's already running that.

Vince Alright, Ava — two months, we check the literature. I'm genuinely curious whether the spec reads as a constraint or a description when someone actually runs the numbers.

Ava Yeah… me too, honestly.