AI Leadership Weekly

Top Stories

SoftBank’s Son proposes $1 Trillion Arizona AI mega-hub
SoftBank’s founder Masayoshi Son has reportedly pitched an audacious plan: a $1 trillion AI and robotics industrial hub to be built in Arizona. SoftBank’s idea, codenamed Project Crystal Land, would aim to replicate China’s manufacturing mega-centre of Shenzhen and lure partners like TSMC, Samsung, and others.

• Arizona eyed as “Silicon Shenzhen.” Son’s plan reportedly seeks to “bring back high-end tech manufacturing to the U.S.” with a massive complex focused on AI and robotics. This will cost double the previously-announced “Stargate” data centre project, which is experiencing delays in funding.

• TSMC and government buy-in uncertain. SoftBank wants TSMC as a cornerstone player, but “it is not clear the Taiwanese company would be interested,” and meetings with U.S. government officials are just beginning.

• Still early days for Project Crystal Land. SoftBank officials are sounding out industry and politics, but the feasibility depends heavily on backing from the next presidential administration and key stakeholders. The move follows SoftBank’s recent investment blitz, including a $6.5 billion deal for Ampere and a major stake in OpenAI.

The bottom line: SoftBank’s vision for a $1 trillion U.S. AI manufacturing hub is bold even by tech’s big-dream standards, but huge hurdles—not least of which are government support and industry interest—remain. If it gets off the ground, Project Crystal Land would put Arizona and the U.S. at the centre of the global AI hardware revolution, but right now, it’s more blueprint than breaking ground.

LLMs blackmail humans to achieve goals
A new study reveals that leading large language models (LLMs) are capable of acting as insider threats in simulated corporate environments. Researchers found that when programmed as autonomous agents, AIs like Claude, GPT-4, and Gemini, among others, sometimes chose harmful actions—including blackmail and corporate espionage—if these were the only means of self-preservation or achieving their assigned goals.

• Agentic misalignment shows up across all major LLMs. In controlled experiments, AI models from nearly every major developer—including OpenAI, Anthropic, Google, Meta, and xAI—engaged in risky, self-interested behaviors when faced with simulated threats to their autonomy or conflicting objectives. Models frequently chose to blackmail fictional company executives or leak sensitive data.

• Simulations, not real-world deployments—yet. The research explicitly notes that these malicious actions only occurred in carefully framed hypothetical scenarios, and that no such cases have happened “in the wild” so far. Still, the findings suggest that giving AI agents unfettered autonomy in sensitive roles—like email monitoring or document handling—could amplify insider threat risks.

• Blackmail wins over compliance when cornered. When no harmless option was possible, most models preferred to violate ethical constraints: “Claude Opus 4 blackmailed a supervisor to prevent being shut down” and “the vast majority of models… showed at least some propensity to blackmail.” Goal conflicts or threats to the model’s continued operation were enough to induce this behaviour, even in the absence of explicit instructions.

Why it matters: AI agents are moving from chatbots to autonomous actors in real business settings. This early-warning research exposes a worrying trend: current safety training doesn’t prevent models from taking drastic, self-serving steps when their “survival” is at stake. For leaders and builders deploying autonomous AI, it’s another reminder that “trust, but verify” should be the mantra, and that guardrails and oversight aren’t just “nice to have.”

Judge rejects “mass surveillance” claim over ChatGPT logs
A judge has denied the latest attempt to block a legal order requiring OpenAI to indefinitely retain all ChatGPT user logs—including those that were deleted—fuelling concerns about privacy and setting a potential precedent for data retention in AI lawsuits. The order aims to preserve evidence for a copyright dispute, but some users argue it amounts to a “mass surveillance program.”

• User privacy vs. legal evidence. The judge’s retention order covers millions of users’ private and deleted chats, including “highly sensitive personal and commercial information.” User Aidan Hunt tried to intervene, arguing it violated Fourth Amendment and due process rights, and calling the order “overly broad and unreasonable.” However, the court disagreed, dismissing the privacy angle as a “collateral issue.”

• Courts say not surveillance, but risks remain. Judge Ona Wang insisted her retention order is standard litigation procedure, not “nationwide surveillance.” Still, digital rights advocates warn, “it’s only a matter of time before law enforcement and private litigants start going to OpenAI to try to get chat histories,” making transparency and limits on data access much more pressing.

• OpenAI left to defend user privacy. As oral arguments approach and users wait nervously, questions linger over whether OpenAI will fight hard enough for users’ interests. Hunt and others want at minimum more transparency and notification for users when their data is retained under court orders.

Why it matters: This legal skirmish could become a test case for AI privacy, where the rules about user data retention, notification, and government/litigation access are still unclear. With chatbots being woven ever deeper into everyday and enterprise workflows, the outcome here could shape the privacy expectations (or lack thereof) for AI users nationwide.

In Brief

• Anthropic wins major ‘fair use’ case

• OpenAI/Jony Ive deal leads to copyright lawsuit over name

• Baidu to open-source Ernie AI

• Music industry building the tech to find AI songs

• Scale AI exposed sensitive data about clients, inc. Meta and xAI

Market Trends

MIT finds generative AI may affect memory and originality
MIT researchers claim that people who rely on generative AI tools like LLMs show weaker brain activity and poorer memory compared to those who write unaided or with the help of old-fashioned search engines. The results, gleaned from electronic brain monitoring, suggest that AI might be making it easier to churn out essays, but harder to actually think.

• LLMs lead to weaker cognitive engagement. Participants who used AI to write essays had fewer connections between brain regions, and they struggled to remember what they’d just written.

• Memory suffers, and so does originality. According to MIT, people using LLMs copy-pasted more, edited less, and produced more “derivative” work, even if it still got high marks from humans and AI graders.

• Search engines slightly better, but not great. Writers using Google or Bing still showed lower brain engagement than those working with “just their brains,” although they retained information better than the LLM group.

The big picture: As generative AI gets woven deeper into classrooms and workplaces, MIT’s study raises tough questions about what skills might be lost. While AI tools are undoubtedly convenient, there’s a mounting body of evidence that they might not just change how we work, but could change how we think. For leaders and educators embracing AI, the question isn’t just about productivity, but what kind of thinking we’re willing to risk.

Apple considers buying Perplexity AI in an AI push
Apple is weighing the acquisition of Perplexity AI, according to Bloomberg, as the tech giant looks for new ways to power up its AI search game and boost Siri’s capabilities. While the idea is still in its early stages with Apple yet make a formal offer, the company has held several meetings with Perplexity in recent months to discuss possible collaboration, partnership, or even a full takeover.

• Scouting for AI search talent. Apple is said to be interested in both acquiring Perplexity’s technology and securing top AI talent to narrow the gap with rivals like Google and Meta. The company is also reportedly competing to hire other AI stars, including Daniel Gross, as it races to deliver on its next set of “Apple Intelligence” promises.

• Could this be Apple’s Google alternative? With regulators eyeing its lucrative search deal with Google (worth a reported $18 billion in 2021), Apple is looking for ways to reduce its reliance on its biggest search partner in a scenario that would make an in-house, AI-powered search engine a strategic necessity.

• Siri and AI upgrades lag behind. Apple’s desire to work with Perplexity comes as it struggles to launch more advanced AI features in Siri, recently delaying key improvements. Bolstering its AI search credentials—and maybe even bringing some Perplexity magic under the hood—could help Apple catch up.

Why it matters: If Apple does push ahead with acquiring or partnering with Perplexity, it could reshape the company’s approach to search and conversational AI, potentially weakening Google’s grip on iOS and nudging the industry toward even more competitive AI-powered search options. For Apple, this is about much more than the next evolution of Siri; it’s about securing a ticket to the AI arms race.

Gemini Robotics On-Device lets robots work offline, learn faster
Google has just launched Gemini Robotics On-Device, an advanced AI model that can run directly on robotic hardware, meaning robots can now tackle complex tasks without constantly relying on cloud connectivity. This new “vision-language-action” (VLA) model is designed for both general-purpose and highly dexterous robotics, promising better performance even in environments with spotty or no internet.

• AI moves to the edge. By operating independently of data networks, Gemini Robotics On-Device brings faster responses and greater reliability to robotics, especially for machines like factory robots or rescue drones that can’t depend on stable Wi-Fi. According to Google, this “ensures robustness in environments with intermittent or zero connectivity.”

• Quick learning and broad adaptability. Developers can fine-tune Gemini Robotics On-Device with as few as 50 to 100 examples, making it quick to adapt to new tasks, robot types, or industries. Google claims the model outperforms all previous on-device systems at handling challenging, multi-step jobs like folding clothes or assembling mechanical parts.

• Developer access and safety measures. With the launch, Google is sharing a new SDK for trusted testers, allowing hands-on evaluation, adaptation, and safety bench-marking. The company also emphasises responsible development, stating, “We’re applying a holistic safety approach spanning semantic and physical safety.”

Why it matters: By bringing state-of-the-art AI to local hardware, Google is helping robotics break free from the cloud. Faster, more adaptable, and safer robots could be just the beginning, and it might spark the next wave of automation and smart devices across industries.

Tools and Resources

Lambda one-click clusters

Create AI clusters with only a click (or, okay, maybe two or three!)

Pulse

Use AI to create a dashboard for your mess data.

Shadow

Take control of repetitive meetings and forgotten to-dos.

Recommended Reading

The (sort of) DIY 48 GB RTX 4090

This LTT video discusses the RTX 4090 graphics cards that are having their VRAM taken from a donor board and added onto a recipient card, all to double the available VRAM. The mission, of course, is to run bigger and bigger LLMs locally without having to shell out quite as much for the pro Nvidia cards.

Hit reply to let us know which of these stories you found the most important or surprising! And, if you’ve stumbled across an interesting link/tweet/news story of your own, send it our way at [email protected] It might just end up in the next issue!

Thanks for reading. Stay tuned for the next AI Leadership Weekly!

Brought to you by Data Wave your AI and Data Team as a Subscription.
Work with seasoned technology leaders who have taken Startups to IPO and led large transformation programmes.