Public transcript

Sam Altman on Building the Future of AI

By: OpenAI
Length: 46 min
Language: English
Text from: the video's captions
Transcribed: 12 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This OpenAI forum conversation features Sam Altman, Josh Acklin, and Adrian Akyol discussing the superintelligence blueprint OpenAI released that morning. They explain why the document was published now — accelerating progress toward extremely capable models — and compare the moment to the early days of COVID. The discussion covers upsides like accelerated science and cheap abundance, layered resilience against cyber and bio threats, democratizing compute access, economic ideas like modernizing the tax base and portable benefits, and the March 2028 goal for an automated AI researcher. It closes with audience questions and details on how to give feedback.

Key points

Sam Altman says the rate of AI progress is accelerating and OpenAI expects extremely capable models quite soon, which is why they released the blueprint now to give society time to debate.
Altman compares the current moment to a cold night walk in early 2020 when OpenAI researchers saw COVID coming before the rest of the world did.
Resilience is framed as layered: safety evals and red teaming remain important, but society also needs incident reporting modeled on aviation, plus cyber and bio defenses.
Altman argues the world's software is more brittle and less secure than we think, and open-source models good at code will soon be good at security exploits, so AI must be used defensively.
Democratization of AI means both broad access to compute and giving people a voice in where the technology goes; Altman compares abundant cheap AI to the historical effect of cheap electricity.
Ideas in the blueprint include modernizing the tax base, portable benefits, a possible 32-hour work week, and counter-cyclical measures tied to AI disruption.
Adrian Akyol says the official goal for an automated AI researcher is March of 2028, which could accelerate further AI progress.
OpenAI is inviting feedback at newindustrialpolicy@openai.com, offering fellowships and research grants up to $100,000 plus up to $1 million in API credits, and opening a workshop in Washington, D.C. in May.

Questions it answers

00:00Why release a blueprint now

Why is OpenAI releasing a superintelligence blueprint now?

Sam Altman says progress is accelerating and OpenAI believes it is very close to extremely capable models, so the public and political system need time to debate ideas before decisions must be made. The document is meant as early ideas to start a discussion, and OpenAI may be wrong but expects a world of highly capable models quite soon.

The rate of progress is continuing to accelerate.
More debate time before decisions leads to better decisions.
The blueprint is meant as early ideas to start a discussion.

04:45The COVID analogy

How does Sam Altman describe the current moment?

He compares it to a cold night walk in early 2020 when OpenAI researchers saw COVID coming before the world did — the change has already happened but society has not digested it. He says it feels like that night walking through the streets again.

Working on exponentials primed researchers to see COVID early.
Altman feels the same acuity about AI now.
The models have already hit some level society hasn't digested.

06:45Upsides and responsibilities

What positive things do they see in the pipeline?

Potential upsides include a decade of scientific progress in a year, curing diseases, personalized medicine, new materials for cheap safe energy, and anyone being able to have AI implement a startup idea. Josh adds that AI could make broad benefits like food, shelter, and health care affordable for everyone.

Science acceleration is a central upside.
AI itself becomes a tool to mitigate downsides.
Broad benefits for everyone is a key responsibility of stewards.

10:30Resilience in layers

How do the speakers think about resilience to AI risks?

Adrian describes layers: classical safety like evals and red teaming, plus society-wide measures such as aviation-style incident reporting and hardening software. Sam adds that there will be many AIs, open-source models will be good at security exploits, and the world needs cyber and bio defense shields built with AI.

Incident reporting modeled on aviation is one proposal.
AI should be given to trusted defenders first.
Bio defense needs detection systems and rapid response treatments.

18:00Startups and democratizing access

How will AI change startups and access to AI?

Altman says every big drop in the friction of starting a startup, like the arrival of AWS, enabled amazing new things, and this coming shift is much bigger — two or three people and a ton of GPUs. Democratization means both broad access to compute and a voice in where AI goes, which he compares to making electricity cheap.

Small teams with abundant compute could build entire startups.
If compute stays scarce, the rich will bid up prices and concentrate power.
Energy abundance historically lifted quality of life; AI should follow.

25:00Workers, taxes, and the upside

How can society share the economic upside of AI?

Ideas include modernizing the tax base to tax AI rather than human income, portable benefits, a possible 32-hour work week, counter-cyclical unemployment insurance, and eventually new forms of ownership. Josh also suggests in-between institutions with mid-level accountability between corporations and governments.

Workers should co-author how AI is deployed, with unions empowered.
No one should lose their health care if they lose their job.
Capitalism's labor-capital balance may need evolution.

32:30The 2028 automated researcher

What is the timeline for an automated AI researcher?

Adrian says the official goal for an automated AI researcher is March of 2028, which brings a double disruption: a model capable of advanced cognitive work that may also accelerate further AI progress. He can't predict the pace after that point but expects it to exceed current progress.

March of 2028 is the official goal.
An automated researcher could compound AI progress.
There is significant uncertainty about these numbers.

34:00Audience questions on humans and care

What human qualities will matter most, and how can AI expand access to care?

They point to timeless qualities like character, compassion, and creativity, with Sam sharing his underwhelming robot cafe experience showing how much he values human interaction. On care, they see AI making health care navigable and scalable, freeing people to provide actual care, and cite Sam's blood test anecdote with ChatGPT.

Human connection matters more than people expect.
AI can make the best health care available at scale.
The care economy may grow in importance.

Notes

Why the Blueprint Now

Sam Altman says progress is accelerating and OpenAI believes it is "very close" to extremely capable models, so the public and leaders need time to debate before decisions must be made.
He compares the moment to a cold night walk in early 2020, when OpenAI researchers saw COVID coming before society did — "the change has already happened" but society has not digested it.

Upsides

Potential gains include a decade of scientific progress in a year, curing diseases, personalized medicine, new materials for cheap safe energy, and anyone implementing a startup idea with AI.
Josh Acklin frames AI as enabling broad benefits — food, shelter, electricity, health care — at much lower cost, with a special responsibility to help working and middle-class people and low-income countries.

Resilience in Layers

Adrian Akyol distinguishes classical safety (evals, red teaming, mitigations) from society-wide resilience: incident reporting modeled on aviation, and hardening software against AI-enabled cyber attacks.
Altman says there will be many AIs, not a few; open-source models good at code will be good at exploits, so defenders need differential access to capable models. A similar "bio defense shield" of detection and rapid response is needed.
Josh Acklin highlights supply chains and especially under-attended food supply chain bio risks.

Democratization and the Economy

Altman compares abundant cheap AI to cheap electricity: if compute stays scarce, the rich will bid up prices and concentrate power.
Blueprint ideas include modernizing the tax base, portable benefits, a possible 32-hour work week, counter-cyclical unemployment insurance, and eventually new forms of ownership in the upside.
The official goal for an automated AI researcher is March of 2028, which could compound progress.

Adoption and Human Qualities

Speakers note a capability overhang: people check AI every few months and miss reasoning models, underestimating progress.
Altman's robot cafe story and the care economy discussion emphasize that human connection remains central.

Text from the video's own captions. Summary and notes written by AI from the transcript, so check anything important against the recording.