brain/
conceptartificial-intelligence

ChatGPT super-assistant vision: chat → actions + proactivity

Notes

ChatGPT super-assistant vision: chat → actions + proactivity

Vintage: Mar 2026. nick-turley (Head of ChatGPT at OpenAI) on the Bg2 Pod March 15, 2026. The vision aligns with the broader super-assistant framing OpenAI has been signaling since 2024 ("Super Assistant Server" / SA server is the internal codename); Turley's contribution is product-side detail on what shipping the vision requires. Refresh against actual product releases.

One-line summary: nick-turley's framing of where ChatGPT goes next from a 900M-WAU chatbot to a "super assistant" — defined by two compounding capabilities: (a) the model taking actions (not just answering), and (b) the model being proactive (prompting the user rather than waiting to be prompted). Domain-specific agents are already shipping (Codex has escape velocity). General-purpose agents are "close." The transition is gated on capability hitting "escape velocity" — the threshold where users get enough partial credit on agentic tasks that they keep trying, generating the use-case data that lets the model improve.

The framing

Two compounding capabilities

nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "There are two concepts. There's ChatGPT doing stuff rather than just answering and then as ChatGPT being proactive. And when you put them together, you start feeling like it feels like a super assistant because these things compound."

  • Actions. ChatGPT today has a limited action space (web search, image generation). Goal: full action space of a human with a computer.
  • Proactivity. ChatGPT today is reactive. Goal: model understands user goals and acts speculatively ("you just landed where you were supposed to go, I'm going to call a cab for you"). Pulse is the first proactive surface.

Escape velocity as the deployment threshold

nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "When I look at past attempts that we've made, like the ChatGPT agent, for example, which kind of has capabilities like this, it was just slightly too early. The models weren't quite good enough to hit real escape velocity. And the problem is, if you don't have escape velocity, is that users don't learn to trust it... I do think we're about to get to that point with general purpose agents where it works well enough that you get at least partial credit."

The escape-velocity model: a capability that's "almost-working" produces failure that scares users off. A capability that produces "partial credit" reliably gets continued user attempts, generating the use-case data the lab can train on. The current general-agent moment (Mar 2026) is sitting right at this threshold.

Why domain-specific came first

Code agents work (Codex has escape velocity); general agents don't yet. The structural reason: code is RL-friendly. nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "It's testable, you know, if it worked or not. It's very RL friendly... I won't be surprised if you see this happen for other forms of quantitative knowledge work just because it happens to have the properties that code has." Cross-link to ai-coding-agent-asymmetry-on-novel-code — the testability property is exactly the verifiable-domain asymmetry the wiki has been tracking.

Chat as input modality, artifacts as output

nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "Chat is a great way of expressing your intent. It's a good way of communicating with the machine, but it's not a great output where in many cases what you want back is an artifact. Here's your plan for your trip. Here is the analysis. Here is an outcome that I delivered for you." This is the operative product-architecture statement: ChatGPT's UI evolution decouples input modality (chat / natural language) from output deliverable (rendered artifact).

Why it matters to this thread

The thread accumulates evidence on the agent capability transition. Turley's framing is unique in the thread for being product-side (OpenAI's actual roadmap framing, not an outside analyst's) and is from the head of the most-deployed AI product on the planet. Specifically:

  • The escape-velocity model is the right way to read the capability-inflection convergence (Karpathy December 2025, Benioff May 2026, Turley March 2026 corroboration). All three independent sources are describing what happens after escape velocity is hit: users keep trying, the use-case data accumulates, the model gets steeply better.
  • The "two compounding capabilities" framing (action × proactivity) is now the wiki's clearest characterization of what "super assistant" means operationally, distinguishing it from "better chatbot."
  • The mass-market-impact framing (10% of world, 90% to go) is the load-bearing growth claim for the openai commercial trajectory.

Evidence

Action + proactivity compound

  • nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "Strictly speaking, ChatGPT can do stuff. Today the action space is just very limited. It can search the web... but it clearly doesn't have the same action space that a human with a computer would have. And that is what we aim to build."
  • nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "Pulse is limited in the value it can provide for you because it's not connected to your Life and it can't take action. So it's producing information for you and people love that... but I think the magic begins when you have actions and proactivity because then it can begin speculatively actually detecting, hey, you just landed where you were supposed to go, I'm going to call a cab for you."

Coding agents already past escape velocity (Mar 2026 — Turley/OpenAI side)

  • nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "The thing that's already come first is the domain specific agents. If you look at what's happening in code, we're fully there. It's mind bending, but we've got so many engineers who don't open their IDE like ever... So Codex and products like it is clearly a product that has escape velocity where people are absolutely using it for all kinds of agentic work." Cross-corroborates andrej-karpathy's December 2025 inflection ("haven't typed code basically since December") and marc-benioff's May 2026 "Anthropic 4.6 hit, boom" framing — three independent practitioner vantage points converging on the same late-2025/early-2026 capability inflection.

Why code first, not other domains

Mass adoption is the deployment moat

Pricing must evolve

  • nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "It's possible that in the current era, having unlimited plan is like having an unlimited electricity plan. It just doesn't make sense because people may need a lot, a lot of electricity and they're getting a lot of value out of that." — Predicts metered / usage-based pricing layer alongside subscriptions. Cross-link to gpu-as-zero-sum-constraint.

Code Red as company-wide focus tool (Q4 2025)

  • nick-turley in 2026-03-15-bg2-chatgpt-super-assistant-era: "Code Reds are a tool we use to create focus... End of last year we had one of those moments... We need to focus on the basics like reliability, performance, the way that talking to the model feels, making personalization really great... We just exited the Code Red with the launch of 5.3, which is a great model for the everyday user, and 5.4, which is workhorse if you're trying to do real knowledge work." — Confirms a multi-quarter Q4 2025 → Q1 2026 OpenAI-internal "Code Red" focus event coincident with the Karpathy / Benioff / Turley capability-inflection convergence. Suggests the inflection wasn't accidental on the OpenAI side either.

Tensions / open questions

  • Action-space general agents are still gated on capability. The ChatGPT Agent v1 attempt was "too early." How far away is escape velocity for general (non-code) agents? Turley says "we're about to get to that point" — but the exact timing is the question every consumer-AI investor has.
  • Proactivity has privacy / autonomy boundary questions the wiki doesn't yet have evidence on. A model that calls cabs for you is excellent. A model that runs analyses without being asked may be welcome or invasive depending on user goals.
  • Cross-camp question: does this look like three-camps-of-ai Camp 3 thinking from a Camp 1 lab? Turley's framing is unusually deployment-focused for the head of ChatGPT — most of the friction is at the product layer (escape velocity, partial credit, use-case discovery). The Camp 3 / Camp 1 distinction the wiki uses elsewhere is blurred here: OpenAI is running both bets simultaneously.

What would falsify this

  • General-purpose agents fail to hit escape velocity through 2026 — would suggest the "we're about to get there" framing was over-confident and the next year is more of the same (chatbot + niche agentic features).
  • The action / proactivity split turns out not to compound — i.e., users want one without the other. Would weaken the "super assistant" framing as the meaningful product unit.
  • Subscription pricing remains durable (no metered tier) — would suggest the GPU-constraint pricing pressure is overestimated.

Related

Referenced by
Entities
Apoorv AgrawalArvind JainNick TurleyNvidia