Public transcript

Big Technology AI Summit (full): Greg Brockman, Mike Krieger, Aaron Levie & Friends of The Podcast

By: Alex Kantrowitz
Length: 3 h 48 min
Language: English
Text from: the video's captions
Transcribed: 13 October 2026

Read the full transcript Open on YouTube

Notes in English.

Summary

This is the full recording of the first Big Technology AI Summit, hosted by Alex Kantrowitz at the Commonwealth Club in San Francisco. The day covers the state of AI in mid-2026: the government restriction of Anthropic's Fable and Mythos models, the massive data center buildout, token economics and enterprise ROI, AI product building, and OpenAI's frontier plans. Guests include Aaron Levie, three AI infrastructure reporters, Dallas Dolan of PwC, Ranjan Roy, Alex Stamos, Mike Krieger, and Greg Brockman.

Key points

ChatGPT has reportedly hit 1 billion active users, NVIDIA is at a $5 trillion market cap, OpenAI raised $122 billion, and the industry expects $700 billion in capital expenditures this year.
Aaron Levie argues the Fable export controls were more likely a chaotic government reaction than an Amazon conspiracy, and that government evaluation of frontier models will likely become standard within 3 to 5 years.
Alex Stamos says Mythos is genuinely the best bug-finding model known, but Opus 4.8, GPT-5.5, and Chinese open-weight models are close, so restricting bug-finding hurts defenders more than attackers.
Reporters Ana Gardez, Max Churnney, and Lauren Good estimate roughly 50% of announced AI data centers may not get built on time, and see Taiwan/TSMC as an under-planned geopolitical risk.
Dallas Dolan of PwC says the real question is 'outcome maxing' not token maxing, with ROI measured in hours saved, and that PwC would shift usage toward cheaper models if prices were cut.
Mike Krieger describes Anthropic Labs' mission as closing the gap between model capabilities and products, and says Fable enabled delegating large tasks like converting a whole project from Python to TypeScript overnight.
Greg Brockman says OpenAI is building toward a personal AGI with almost no interface, where agents hooked into Gmail, Slack, and your computer act on your behalf, and that scaling laws have hit no wall.
Brockman says roughly 230 million people use ChatGPT for health queries each week and that AI-assisted personalized medicine will become standard.

Questions it answers

00:00Opening remarks

Why is this summit happening and what are its core questions?

Alex Kantrowitz recounts starting Big Technology after a 2020 Commonwealth Club event and frames the summit around three questions: what the technology will do next, what happens if it works, and what happens if it goes wrong.

ChatGPT near 1 billion users; $700B capex this year
Three core questions structure the day

09:32Aaron Levie interview

What does the Fable/Mythos restriction mean for AI safety and the industry?

Levie sees the restriction as a likely chaotic government reaction rather than an Amazon strategy, argues it sets a precedent for government control of model releases, and predicts government model evaluation within 3-5 years. He also covers open-weight models closing on the frontier and token spend being driven by harder tasks.

Government greenlighting of releases likely in 3-5 years
Export controls may push sovereign AI stacks abroad
Box agents use ~100x more tokens as tasks get harder

40:18Infrastructure reporters panel

Is the AI data center buildout real and what are the risks?

Ana Gardez, Max Churnney, and Lauren Good discuss roughly half of announced data centers possibly not being built on time, chip industry cyclicality, Nvidia's inference pivot, OpenAI diversifying from Nvidia, and Taiwan/TSMC as an unplanned geopolitical risk.

~50% of announced data centers may be delayed
No real contingency plan exists for a Taiwan crisis
OpenAI wants to diversify from Nvidia

1:02:00Dallas Dolan on ROI

Is enterprise AI spend delivering real returns?

Dolan says the question is outcome maxing, not token maxing or minimizing; ROI is measured in hours saved. PwC uses a control plane to route tasks to the right model and would shift to cheaper models if prices were cut, and is hiring as many people as last year.

ROI measured in hours, not tokens
Price cuts would shift usage to cheaper models
Agent limits are risk, cost, and expectation tolerance

2:04:21Live Friday show

What do listeners think about AR glasses, agents, and AI advertising?

Alex and Ranjan debate whether Snap's AR spectacles and face computers have a future versus the iPhone plus better Siri, take audience questions on agent autonomy benchmarks, agent learning layers, AI advertising and privacy, and US-China AI cooperation.

Alex: the AI device is the iPhone with better Siri
Models will build detailed profiles of users for ads
DeepSeek and Chinese models now appear in cost conversations

3:18:55Alex Stamos on security

Is the Mythos danger material or marketing?

Stamos explains the Fable takedown timeline, says Mythos is genuinely the best bug-finding model but not a singular cyber god, and argues restricting bug-finding hurts defenders and injects political risk into US AI, pushing companies toward Chinese open-weight models as backups.

Mythos best at bugs, but Opus/GPT/Chinese models close
US models must understand security to write secure code
Political risk now a factor in choosing AI vendors

3:40:17Mike Krieger on Anthropic Labs

How does Anthropic build the next generation of AI products?

Krieger explains Labs' two guiding questions — close the gap between capabilities and usage, and prepare products for what models will be good at in six months — describes delegating whole projects to Fable, and addresses Anthropic's platform-vs-product tensions and safety considerations.

Labs created Claude Code and computer use
Fable delegated a full Python-to-TypeScript conversion
Instagram could have been built by 4-6 people with these tools

Notes

The Big Picture

Alex Kantrowitz opens by recalling his last Commonwealth Club appearance in May 2020, which led him to leave BuzzFeed and start Big Technology.
The framing questions of the summit: What will the technology do next? What happens if it works? What happens if it goes wrong?
Key stats cited: ChatGPT near 1 billion active users, NVIDIA at a $5 trillion market cap, OpenAI's $122 billion raise, $700 billion in capex this year, and OpenAI plus Anthropic heading toward $50 billion in combined revenue.

Aaron Levie on the Fable/Mythos Restrictions

On the "Jassy mystery," Levie prefers Occam's Razor: Amazon's security teams found jailbreaks that recovered Mythos-level capability, and the government reacted with blunt tools, rather than a boardroom plot to kneecap the labs.
He argues that for AI safety believers, the Fable restriction is a desirable precedent: government can now press a button to stop a model's rollout.
He predicts government evaluation and greenlighting of model releases within 3-5 years, and warns export controls may push other countries (especially China) to build sovereign AI stacks.
On tokens: Box's agents went from 5,000-20,000 tokens per task to a million or more, a ~100x increase, because tasks keep getting harder — cost per unit of intelligence falls while appetite grows.

Infrastructure Panel (Gardez, Churnney, Good)

Ana Gardez believes ~50% of announced data centers won't be built on time, and reporters should back into gigawatt announcements to check commitments.
Max Churnney expects chip overcapacity and a cyclical whiplash; he sees Taiwan/TSMC as a black swan with no real contingency plans.
Lauren Good notes Jensen Huang pivots well (crypto, AI, now inference) and shares the story of brushing Jensen's hair on a Wired cover shoot.
OpenAI is investing in non-NVIDIA chips (Cerebras, in-house silicon) to diversify from Nvidia.

Dallas Dolan on ROI

The winner will be defined by "outcome maxing," not token maxing or minimizing; ROI is measured in hours saved.
PwC's control plane routes tasks to the right model; a 50% price cut would shift usage toward cheaper models.

Stamos on Security

Mythos is likely the best bug-finding model in the world, but Opus 4.8, GPT-5.5, and Chinese open-weight models are close; bugs are cheap and plentiful.
Restricting bug-finding is an own goal: American LLMs must understand security to write secure code; GLM 5.2 from Z.AI is MIT-licensed and near frontier.

Krieger and Brockman

Anthropic Labs' two questions: close the gap between model capability and usage; imagine what models are bad at now but good at in six months.
Fable let Krieger delegate a full Python-to-TypeScript conversion overnight.
Brockman: the goal is a personal AGI with almost no interface; Codex usage inside OpenAI now rivals Slack; scaling laws have hit no wall; ~230 million people use ChatGPT weekly for health queries.

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