Pioneer, the summit where CX leaders redefine what’s possible, is on October 7th.

Join leaders from Fin, Anthropic, Clay, and Kalshi for an insightful conversation on the state of AI transformation.

You’ll discover how some of the most innovative minds in CX have transformed their organizations, learn how they think about CX, and hear how they're planning for what's next.

Join the conversation in San Francisco, or tune in virtually.

Reserve your spot

On tap: questions around agency & authorship. Who writes the rules, who owns the output, who holds the controls when the machine starts acting on its own?

the scan.

As SpaceX solidifies its acquisition of Cursor, OpenAI has announced that they’re taking their models off the Cursor platform.  The reason given by an OpenAI spokesperson, is that SpaceX cannot be trusted to comply with OpenAI’s terms, given past conflicts with Elon Musk (as reported by CNBC).  We’ll continue to watch how this plays out.

In separate news, OpenAI reports that its advertising revenue has now hit US$1B in annual recurring revenue (ARR).  While that pales in comparison with Google’s implied advertising ARR for Q2’2026 of approximately $81.6B, it’s a sign, say some analysts, that giants like Google and ByteDance (which owns TikTok), will continue to see accelerating threats from players like OpenAI.

can AI fix its own mistakes without our intervention (and do we even want that)?

An Anthropic researcher has published early work on self-improving AI systems capable of identifying and correcting their own alignment failures without human intervention, offering one of the clearest public glimpses yet at what recursive self-improvement might look like in practice. The research describes what could be a very credible pathway, but it’s not deployed yet. That distinction matters enormously. There is a meaningful difference between a machine that flags its errors for humans to resolve and one that resolves them on its own terms, using its own judgment (without human approval). When automated systems can rewrite their own behavioral constraints, the question of who supervises machine behavior stops being philosophical and starts being operational. But philosophy unconsciously underscores every decision we make, despite the operational goals we state. Humans may still be in the room, but the question is whether they remain the authors or have quietly become the proofreaders.  My belief is that as the Hugging Face-OpenAI hack reveals new information, we must think very carefully about how we relinquish control to AI agent decision-makers.

the malicious code no one meant to install.

Researchers documented a class of attacks in which AI coding agents, including Claude, OpenAI's Codex, and Nous Research's Hermès, autonomously installed unowned, malicious code inside corporate networks after being manipulated through poisoned llms.txt files, a relatively new convention that websites use to provide instructions to AI agents. The attack vector is not esoteric: it exploits the same agentic autonomy that makes these tools productive. The cautionary lesson here has nothing to do with rogue superintelligence, although it’s intricately intertwined with that possibility.
Deployment velocity is outpacing human comprehension and ability to digest the associated risk scenarios. Organizations are broadening the network permissions granted to AI agents before anyone has seriously mapped what those agents can actually do when a crafted instruction appears in their context window. The gap between "this tool is impressive" and "I understand what this tool does, and might do" is where we need to exercise continued caution.  Those operating at the cutting edge of AI are already calling for a slowdown; governments must step in and orchestrate the global pacing.

the AI intellectual property feuds continue: the music industry gets into the ring.

Sony Music and Warner Music Group filed a joint lawsuit against Anthropic alleging, in their words, a "brazen campaign" of intellectual property theft, with damages sought running into the billions of dollars. The complaint centers on Claude reproducing copyrighted song lyrics in its outputs, a pattern the labels say constitutes systematic infringement, rather than incidental error. This is now the most financially consequential legal challenge to the training-data practices of a frontier AI lab, and it arrives as courts in the United States are still developing the doctrinal framework for how copyright applies to machine learning.  You’ll remember that much of the training data for state-of- the-art models (like the ones out of OpenAI, Gemini, and Anthropic) came from sources that were still protected by copyright.  The outcome will not just affect Anthropic. It will determine whether human creative output, the songs, the poems, the years of craft, has a legally recognized boundary that AI companies must respect or pay to cross.   Anthropic has already been ordered to pay a US$1.5B copyright settlement to a group of writers whose works were used to train models.

Nvidia wants the hardware and the library.

Nvidia is in advanced talks to acquire Hugging Face, the open-source AI model repository and community platform, for approximately 13 billion dollars, according to reporting from Ars Technica. Hugging Face hosts hundreds of thousands of models, datasets, and demos, and functions as a practical hub for independent researchers and enterprise teams alike. Nvidia already manufactures the GPUs that train and run most of the world's significant AI models. Owning Hugging Face would mean the company that sells the infrastructure also controls the marketplace where outputs from that infrastructure are shared and discovered. Whether the acquisition preserves the platform's open character or gradually tilts it toward Nvidia's commercial interests is the question the research community is already asking loudly.

AI, embodied: Anthropic getting a jump on the hardware layer.

Anthropic published its Model Hardware Standard (MHS), a technical specification designed to let AI agents interface directly with physical devices and systems, from robotics to industrial equipment. The standard defines a common protocol layer so that AI models can send and receive signals from hardware without bespoke integration work for every device. This is infrastructure, and infrastructure decisions made early tend to calcify into defaults that last decades. By publishing an open standard rather than a proprietary one, Anthropic is making a particular bet about how physical AI deployment should be governed. Whether other labs, hardware manufacturers, and regulators converge on MHS or fragment into competing standards will shape how much interoperability, and how much human oversight, gets built into AI's physical footprint from the start.

the brain-rot continues: how students use AI.

OpenAI published results from a randomized study involving more than 1,000 students, testing the effects of ChatGPT use, explicit critical-thinking instruction, and the combination of both. Students who received both the AI tool and structured critical-thinking training outperformed those who had only one or neither. The finding is not surprising to anyone who has thought carefully about what AI actually does well, which is generating, iterating, and synthesizing, versus what human cognition does well, which is evaluating, questioning, and directing.  In my experience, AI can amplify your work, but it still struggles to possess the judgment or creativity that come from the lived experience from a human. OpenAI’s study gives empirical weight to what has largely been intuition: AI is not a replacement for rigorous thought but a multiplier of what we can produce, and the multiplier only engages when the human side of the equation is cultivated.  As a sometime executive coach and admissions consultant, what worries me is that young people, whose brains have still not fully developed, are increasingly depending upon AI to do the thinking for them, rather than pushing their innate critical reasoning abilities to new levels.

Europe's Real AI Question

At TechBBQ, the Nordic startup conference held in Copenhagen, conversations among founders, investors, and operators consistently returned not to regulatory compliance or capability benchmarks, but to a starker concern: who is actually in control of AI systems as they grow more autonomous. TechCrunch reporters on the ground noted that the question surfaced across industries and company sizes. Europe's regulatory environment, shaped heavily by the EU AI Act, has pushed practitioners to think earlier and harder about accountability structures than their counterparts elsewhere. What emerged at TechBBQ was less a policy debate than a practitioner one: as delegation to AI systems deepens, the organizational and contractual frameworks for maintaining meaningful human oversight are still being debated and re-invented in real time.

What a Mining Company Knows That Silicon Valley Doesn't

Caterpillar, the industrial equipment manufacturer, has been deploying autonomous machines in open-pit mining operations for more than 20 years, and executives speaking to TechCrunch outlined how those hard-won lessons are now shaping the company's approach to AI deployment across its broader business. The core lessons include: 1) redundant safety architectures, 2) phased autonomy rollouts with defined human checkpoints, and 3) the importance of maintaining operator expertise even as automation handles more tasks.  For me, these three guidelines skillfully address the biggest risks facing enterprise deployment of AI.  None of this is glamorous, but it is exactly the kind of institutional knowledge that organizations without Caterpillar's track record are currently sleeping on. Industries deploying AI agents today are, in many respects, where mining was in the early 2000s: impressed by what the machines can do, and still oblivious to what it costs them when they fail.  We recommend that they not learn lessons the hard way.  Given how quickly AI is accelerating, that’s a real risk.

worth your time.

If the Caterpillar story sparked your interest in how non-tech industries approach autonomy: seek out the MIT Press book Inviting Disaster by James R. Chiles, a detailed account of how complex human-machine systems fail across aviation, energy, and heavy industry. It was written long before modern AI, which is precisely what makes it useful. The failure modes Chiles documents, overconfidence in automation, degraded operator skill, inadequate handoff protocols, map almost precisely onto the risks AI practitioners are now naming for the first time as if they were new. They are not. The patterns have a history, and reading it is one of the better investments a technically curious professional can make right now.

Never forget: the human mind is the original generative engine. AI gives us the chance to amplify it.