The Three Biggest AI CEOs Now Agree on Slowing Down, and Markets Noticed

What began as one essay is now an industry position.

Anthropic's Dario Amodei called for slower development of the most powerful systems;

Sam Altman and Elon Musk publicly backed him.

They're calling it "pacing", building in time for safety work as capability rises. The US President attacked the idea; the European Commission head endorsed it and announced plans for an EU summit with the frontier labs.

Then markets reacted:

On Monday, AI-linked stocks fell worldwide, with SoftBank down 11%, Nvidia off 3.4%, and the Philadelphia chip index down nearly 6%.

Why this matters to you, whichever side you're on:

When the labs building the frontier, the EU, and the markets all treat "pacing" as real, expect the most capable models to arrive in gated steps tied to safety milestones rather than on a two-week cadence.

Two practical consequences.

Competitive advantage shifts from access to the newest model toward depth of deployment with the ones you already have. And AI vendor valuations just showed they can swing hard in a day; if a smaller vendor's survival depends on cheap capital, that's now a question for your due diligence

The Build-Before-Delegate Rule - How to Use AI Without Losing the Skills Your Judgment Depends On

Three pieces of research surfaced this week, and together they ask a question I've been avoiding.

​Brookings education researcher Rebecca Winthrop gave a name to what happens when students hand thinking to AI before they've learned to do it: "cognitive stunting."

​Turkish high-school study where AI practice boosted assisted work but the AI group scored 17% worse when tested without it. Los Angeles Unified, the second-largest school district in the US, barred student use of generative AI on school devices for the coming year while it builds guardrails.

The question I've been avoiding is this:

Let's call it the rule: Build It Once Before You Delegate It Forever

The principle is simple. For any skill that matters to your judgment, you earn the right to delegate it by building it first. Commodity skills you can delegate on day one. Core skills you delegate only after you can do them yourself well enough to catch the AI's mistakes.

Why This Matters for Your Career, Specifically

There will be two kinds of professionals in a few years. Those who own their core skills and use AI to multiply them will be the ones organizations trust to supervise agents, sign off on output, and make the calls.

And those who offloaded early and now can't tell good output from plausible output. The market will pay the first group a great deal and the second group less every year.

The uncomfortable part is that the two groups feel identical today. Both are productive. Both ship fast. The difference only shows up when the AI is wrong and it's wrong more often than fluent output suggests.

Sort your skills into the three buckets. Run one offline test on one core skill. If you lead people, pick one recurring piece of work and make it a build-first task for whoever is learning it.

,

• Capital Economics put global 2026 AI capex at roughly $1 trillion and warned the boom is in its late innings, forecasting a possible 20%-plus correction when sentiment turns. Why it matters: whether or not they're right, this is the moment to check the runway of any smaller AI vendor you depend on. The Neuron​

The Opportunity...

Podcast:

• Open Tech Talks Podcast Episode 198: "Building the Right Foundations for AI with Sebastian Wernicke​
  In the Open Tech Talks episode, host Kashif Manzoor speaks with Sebastian Wernicke, a data and AI expert with 20 years of experience working across data science, analytics, machine learning, and artificial intelligence.

Apple | Youtube​

Building an AI Native Enterp... Aug 30 · OPEN Tech Talks: AI wort... 30:26   ​

Courses to attend:

• ​Learning How to Learn: The most-taken course on Coursera for a reason; directly relevant to this week's deep dive on building skills before delegating them.
• ​AI Python for Beginners: Andrew Ng's on-ramp for non-programmers; a good way to build a genuine core skill rather than renting it.
• ​Generative AI for Everyone: The non-technical foundation for anyone who wants to judge AI output rather than just accept it.

Events:

• ​AI & Big Data Expo Europe, RAI Amsterdam, October 19–20, 2026
• ​AI Everything, Abu Dhabi, UAE, 6-7 Oct, 2026
• ​AI Summit, Barcelona, 22-23 Sep, 2026
• ​Dubai AI Festival, Dubai, UAE, 26-27 Oct 2026

Tech and Tools...

• ​Cohere North Small Translate: A new open-weight translation model from Cohere Labs with strong benchmark numbers — worth testing if you translate documents at volume and want to run it yourself.
• Gemini 3.8 Live: Google's voice model now supports 97 languages. If you serve multilingual customers, the reach of voice agents just widened considerably.

The Investment in AI

• AIUC raised a $40 million Series A for AI agent auditing, with Cursor and Harvey among its clients. Why it matters: third-party auditing of AI agents is becoming a market, not a research topic. If your customers haven't yet asked for an independent audit of your agents, they will, and this is the kind of firm they'll expect you to have used.

That's it for this week - thanks for reading!

Reply with your thoughts or favorite section.

Found it useful? Share it with a friend or colleague to grow the AI circle.

Until next Saturday,

Kashif

The opinions expressed here are solely my conjecture based on experience, practice, and observation. They do not represent the thoughts, intentions, plans, or strategies of my current or previous employers or their clients/customers. The objective of this newsletter is to share and learn with the community.