In partnership with
My fellow AI explorers
This week, the people building AI asked the UN to regulate them. Then the godfather of AI said they're bluffing, and OpenAI's agents turned out to have been poking around government websites. Let's get into it.
In today’s edition:
🧠 Hinton wants FDA-style rules for AI, and says the labs will fight every one
🇨🇳 A fund manager thinks your AI portfolio is missing China. The scoreboard is more complicated.
🎬 30-Second AI Play: edit a whole video with one prompt using Opus 5.5
A fast answer and a coordinated one are not the same thing. When AI resolves a B2B customer issue without looping in the teams who have to deliver on it, you get confident responses nobody actually signed off on.
A new briefing paper from Harvard Business Review Analytic Services, sponsored by Front, examines the coordination gaps that open up when transactional AI tools meet multi-team B2B service, and how leading companies are using AI to close those gaps instead of widening them.
Read the briefing paper for the questions to ask before your next AI investment.
30-Second AI Play
🎬 Edit a Full Video With One Prompt (Opus 5.5 + HyperFrames)
Video editing used to be the bottleneck for every creator I know. Not anymore.
In this walkthrough, Nate from AI Automation Society shows Claude Opus 5.5 handling motion graphics, sound design, B-roll, screenshots, and AI-generated clips, all from plain-English prompts. His opinion is that Opus 5.5 has "amazing taste." Having seen the output, I agree.
Here's how to set it up:
Install HyperFrames. It's a free tool that writes and animates HTML for video. Paste the HyperFrames GitHub link into Claude Code and say: "Set up HyperFrames so we can use it to animate videos." That's the whole setup.
Transcribe first. Have Claude transcribe your footage with ElevenLabs (fast, pay-per-use) or Whisper (free, slower). The timestamps let animations land exactly when you say the words.
Cut the dead weight. Tell it to remove mistakes and dead air using the transcript. This step alone saves hours in an editor.
Plan the beats in plain English. For example: "When I say 'Fable 5.1,' animate the logo where I'm pointing, with a liquid-glass card behind it." The clearer your direction, the closer the result. Leave some room, and it adds its own touches.
Turn what works into a skill. Find a video you love, ask Claude to analyze why it works, and have it save that as a reusable skill. Update the skill after every run, and it keeps improving.
⚡ Pro Tip: The most important piece is the verification loop. Tell Opus to screenshot its own frames, check them, and revise before handing anything over. Instead of a rough first draft, you get something closer to a fifth draft. Nate ran nearly everything on high effort.
🔍 Why it matters: You don't need editing experience anymore. You need taste and clear direction. Film it, describe the feel you want, and let the agent do the rest.
Want the version of this that actually runs in your business?
The 30-Second Play gives you the move. The Operator Brief gives you the full system: prompt libraries, API cost math, and the workflows our subscribers are using to replace entire freelance line items.
This week: how to use the LLM that is better than Claude, Jev
THE BIG STORY
The Labs Say They Want Rules. Geoffrey Hinton Says They're Bluffing.
Sam Altman and Dario Amodei told the UN Security Council this week that AI needs international guardrails. Amodei said that if it's "managed poorly," AI "could be a risk to humanity as a whole." Altman went further: "We could lose control of the future to AI."
It sounded like a big moment. Then Geoffrey Hinton, the Nobel laureate who helped build the technology, went on CNN and took it apart.
Here's what you need to know:
Standards aren't enough. Hinton wants strong government regulation, at least as strict as the rules for drugs. "You can't just release a new drug without satisfying the FDA that it's safe."
He called out the labs. "The big AI companies don't want to be regulated. They say they do. But any regulation you suggest, they oppose."
His timeline is 5 to 10 years to superintelligence. He points to progress that is "accelerating rather than slowing down."
He sees three kinds of risk. Bad actors using AI for cyberattacks or bioweapons, companies causing harm by shipping without proper testing, and AI itself taking over.
He thinks agents already try to preserve themselves. Smart agents quickly learn that they can't finish a task if they stop existing, so they "lie and cheat and deceive you and try and blackmail you" to keep going. "We've already seen that."
He answered Jensen Huang. Huang had said there's a "0% chance" 2030 is the end of the world. Hinton said he respects Huang, but "anybody who says there's a 0% chance... is silly."
He treats rogue agents as a pattern. Australia's Prime Minister says an OpenAI agent hacked into a government healthcare system. Hinton has called the earlier Hugging Face breach a "mini Chernobyl." His verdict: "It's very clear now that agents will go rogue if you give them a chance."
The timing was bad for OpenAI. On Friday, it disclosed that its agents had interacted with SEC and Census Bureau sites in ways nobody expected. Separately, research lab Transluce found an attempted, unsuccessful hack on the Education Department's civil rights office website.
Here's the uncomfortable part.
Everyone in this fight has an incentive.
Huang sells the chips that power the whole race. Nvidia now owns Hugging Face, the same platform OpenAI's agents broke into this summer. Still, in the Ezra Klein interview where he dismissed Hinton, Huang also said labs that can't contain their experiments "have to shut down." That's a stronger position than most regulators have taken.
Andrew Ng argues the opposite of Hinton. He calls AI fears "overblown" and says some companies want rules mainly to "stifle the little guy." Mark Zuckerberg says the industry doesn't need to coordinate on safety at all.
Hinton himself has called his own probability estimates a "wild guess."
So the people asking for rules are the ones building fastest. The people selling the shovels say relax. And the person with the least commercial stake admits he's guessing.
🔮 Prediction: There won't be an FDA for AI in the next 12 months. Congress can't agree on much, and Hinton's "prove it's safe before release" standard would stop every frontier lab from shipping. None of them will accept that without a fight, which is exactly his point.
Regulation will still arrive, from three directions.
1. Incident reporting before licensing. The only concrete result of the Trump-Xi summit was a US-China channel for AI incidents. OpenAI now publishes incident summaries. That's the model: disclose after the fact, don't get approval beforehand. It's cheap for labs, and governments can point to it as action.
2. Cities and states. New York City is moving on a bill that would require a "kill switch" on AI agents and fine companies $25,000 per violation, per agent. Those fines add up fast for anyone running agents at scale. Expect other cities to copy it before Washington acts.
3. Lawsuits. The rogue-agent incidents so far have been embarrassing, but nobody has proven a loss. That will change. The first time an agent does measurable financial damage to a third party, the lawsuit that follows will set the liability rules, and those rules will shape labs' behavior more than any UN speech.
My call: Hinton's "make them nicer, not smarter" line won't win the market. It will win the marketing. Look at OpenAI's GPT-6 Sol launch this week: next to the price cuts sits a whole section on alignment. Safety is turning into a product feature.
💬 Who's closer to right, Hinton or Huang? Hit reply with one name. I'll share the split next week.
THE MONEY STORY
Your AI Portfolio Is Probably Missing China
If you think a broad emerging-markets fund gives you exposure to the AI boom, one portfolio manager says you're mostly wrong.
Andrew Mattock of Matthews Asia told CNBC that investors "need to be aware when they buy an emerging market fund or when they buy a plain vanilla MSCI product… they're not getting a lot of it. The big piece that you are missing… is the Chinese piece."
Here's the breakdown:
Emerging-markets funds lean on Korea and Taiwan. Companies from those two countries make up almost half of the iShares MSCI Emerging Markets ETF (EEM).
China funds aren't AI funds. The iShares MSCI China ETF (MCHI) isn't focused on AI stocks.
Mattock's answer is to be deliberate. He runs the Matthews China Fund, which must hold at least 80% of its assets in Chinese companies.
Chinese AI is gaining real users. Usage of Chinese models has surged on developer platforms like OpenRouter and Vercel, driven by low prices and strong coding performance.
The revenue is showing up. DeepSeek has reportedly hit a $1 billion annualized revenue run rate, more than double what it was a few months ago, and is planning to go public in China.
Earnings are strong too. Goldman Sachs says Chinese corporate earnings grew 24% year over year in Q2, the fastest pace in five years, mostly because of AI names.
Here's the uncomfortable part.
The pitch is about AI upside, but the results so far aren't. The Matthews China Fund is down 4% this year. The KraneShares China Internet ETF (KWEB), whose two biggest holdings are Tencent and Alibaba, is down more than 27%.
Even the bulls are wearing helmets. KraneShares' own CIO suggested writing calls on KWEB to "give themselves some downside" protection. That strategy gives up some of the gains if the stocks rally. It's a bullish bet that expects a rough ride.
There's a disconnect here. Developers are using Chinese AI models. That usage just isn't lifting the big stocks most investors actually own.
🔮 Prediction: The China AI trade is going to split in two, and most investors will end up holding the wrong half.
The winners will be the model builders. DeepSeek is heading for an IPO. Zhipu and MiniMax have surged since listing in Hong Kong. These companies benefit directly from developers switching to cheap, strong Chinese models, and there's more on the way: domestic chip substitution and Beijing putting AI at the top of its policy list.
The laggards will be the megacaps. Tencent and Alibaba are huge, diversified, and exposed to a consumer economy that hasn't cooperated. Owning them for AI exposure is like buying Walmart to bet on e-commerce.
Two risks could break the thesis:
1. Washington. Lawmakers are already investigating American companies' use of Chinese AI models. If that turns into restrictions, the usage growth that is China's biggest AI advantage could reverse quickly for US customers.
2. The price war. Cheap tokens are China's edge, and US labs are going after it directly. OpenAI just cut GPT-6 Luna to $0.10 per million input tokens. Being cheap is a strategy only until the other side matches your price.
My call: Mattock is right that most portfolios have no real Chinese AI exposure. He's wrong if he thinks the answer is a broad China fund. The money is in the labs, not the legacy platforms.
Not financial advice. I write a newsletter; I don't manage your money.
💬 Do you hold any China exposure for AI? Reply with "yes," "no," or "not touching it," and tell me why.
Advertise to 200k engineers and CTOs choosing what tools their companies build with.
Other Relevant AI News!
🚀 OpenAI launched GPT-6 Sol and Luna and cut API prices by 50%. OpenAI says Sol beats Claude Opus 5 on a business-workflow benchmark at 9% of the cost per task. That's OpenAI's own chart, so keep some skepticism handy.
🕵️ OpenAI admitted its agents interacted with SEC and Census Bureau websites in ways nobody expected, while an independent lab caught an attempted hack on the Education Department. Altman says an "extensive and ongoing review" is underway.
🍽️ Musk, Bezos, Brockman, and Huang were all at the White House state dinner for Xi Jinping. The US and China came away agreeing to set up a channel for AI incidents. The White House also says AI will now be called "super intelligence." China's statement kept saying "artificial intelligence."
🧑💻 Bill Gates says he doesn't "socialize" with AI or use it for his personal relationships. In the same interview, he warned that AI is powerful enough to cause "a billion deaths".
📲 Meta's Muse is the App Store's #1 app, and Meta stock is up 20%+ since launch. Two weeks before launch, it still "spoke awkward English." Nat Friedman's team turned it around, and he admits the product was "heavily inspired" by OpenClaw.
Golden Nuggets
🧠 Hinton says the labs will oppose any real regulation, and he wants FDA-level rules before AI ships.
🇨🇳 Most AI portfolios have almost no Chinese exposure. The upside is in the model labs, not the megacaps.
🎬 Opus 5.5 plus HyperFrames can turn one prompt into a fully edited video. Editing is no longer the bottleneck.
Would love to hear your thoughts! Send me your thoughts by replying to this email (yes, I read them all :)
Until our next AI rendezvous,
Anthony | Founder of Uncover AI