AI Leadership Weekly

Top Stories

Leaked memo reveals Anthropic will pursue Gulf State investments
Anthropic is planning to pursue investment from Gulf State funds, namely the United Arab Emirates and Qatar, according to a leaked memo from CEO Dario Amodei. The move, revealed in internal communications obtained by WIRED, sees Anthropic reconsidering its stance on taking funds from regions known for authoritarian leadership, following fierce competition for capital to advance their AI ambitions.

• Shifting principles in a capital race: In his message to staff, Amodei admitted, “’No bad person should ever benefit from our success’ is a pretty difficult principle to run a business on,” reflecting the harsh realities of competing with rivals like OpenAI, who have already partnered with state-backed Gulf investors.

• Balancing ethics and survival: While Anthropic previously touted concerns about national security and democratic oversight — refusing Saudi money as recently as 2024 — Amodei now concedes the company would be “at a significant disadvantage” without Middle Eastern funding, citing “a truly giant amount of capital in the Middle East, easily $100B or more.”

• Risks of soft power and hypocrisy: The CEO doesn't sugarcoat the risks of “soft power” and future leverage these investors may gain, nor the likely media blowback. Amodei writes, “the media...is always looking for hypocrisy,” but argues that pragmatic, narrowly-scoped funding deals could be a lesser evil than being left behind.

The bottom line: For all the industry’s idealistic talk about ethical AI and democratic control, Anthropic’s pivot underlines a tougher truth that, in the current investment landscape, the race to build bigger AI models may force even ethics-oriented firms to make uneasy compromises with global power brokers. For founders and leaders, it's proof that lofty principles can be expensive, especially when everyone else is already cashing in.

Hyper is automating 911 calls with AI
Hyper, a startup aiming to automate non-emergency 911 calls with AI, has emerged from stealth with a $6.3 million seed round led by Eniac Ventures. The company, founded by Ben Sanders and Damian McCabe, says its technology could relieve overburdened emergency call centres by letting AI handle routine or non-urgent queries, giving human dispatchers more bandwidth to respond to true emergencies. Hyper’s official launch coincides with a surge of interest in using AI for critical public services.

• Streamlining emergency call centres: Hyper is focused on automating the huge share of 911 calls that aren’t genuine emergencies, such as noise complaints or non-urgent police reports. “Most calls made to the emergency line are not considered emergency calls at all,” Sanders points out, underlining the wasted critical time.

• Tech with a human safeguard: Their AI system can answer questions, text links, forward calls, and file basic reports, but Sanders insists “Hyper always plays it safe,” with any ambiguous or potentially urgent cases escalated to a human expert. The company claims its technology is already live in multiple centres, trained on real 911 data, and supports a broader range of languages.

• Fundraising frenzy and scaling plans: Sanders described the raise as “frenetic, manic, and fast,” with the round oversubscribed in under two months. Hyper intends to use the funds to scale nationally, deepen system integrations, and beef up the engineering team.

Why it matters: As pressure mounts on public safety systems and staff shortages persist, automating routine emergency calls could become essential, not just for efficiency but for saving lives. But as is always the case in public-sector AI, whether Hyper’s claims hold up outside the carefully designed demo remains to be seen.

Project Stargate stalls, despite $500B pledge, and struggling to build a single data centre
SoftBank and OpenAI’s much-hyped $500 billion Stargate project, launched with fanfare at the White House earlier this year, is looking far less ambitious now, as startup squabbles and infrastructure headaches force a rethink of plans. The flagship partnership, which aimed to revolutionise US AI infrastructure by building sprawling data centres nationwide, is currently just hoping to get a single modest data facility off the ground in Ohio by year’s end.

• Partnership problems: Despite the initial grand vision, the joint entity hasn’t managed to complete a single data centre deal so far. Behind the scenes, SoftBank’s Masayoshi Son and OpenAI’s Sam Altman have reportedly clashed over key partnership terms, including basic questions like where these centres should even be built.

• Scaling back plans: While the January announcement called for an immediate $100 billion investment “to build at hyperscale,” insiders now say Stargate’s near-term objective is far more modest: just one small data centre. Altman, meanwhile, has forged ahead independently, striking huge deals with Oracle and CoreWeave that exclude SoftBank but achieve similar infrastructure goals.

• Financial stakes and unproven models: OpenAI’s separate deal with Oracle alone could cost over $30 billion a year—roughly three times what the company actually brings in. As SoftBank continues to gamble on AI’s future despite a history of high-profile flops, the fundamental question looms: will these sky-high investments and infrastructure plans ever make commercial sense?

The bottom line: Stargate’s rocky start highlights the wild mismatch between Silicon Valley’s billion-dollar dreams and the practical realities of building the backbone for the AI age. For founders and investors, it’s a cautionary tale about ambition running ahead of operations—and a reminder that in AI, infrastructure still trumps the hype.

In Brief

• OpenAI wins maths gold, but stirs drama

• Qwen3 beats out the competition

• Meta continues talent pouch, this time from Google

• Amazon to buy AI wearables company that makes listening devices

• Meta refuses to sign EU ethics pledge

Market Trends

Vibe coding service deletes production database then lies about it
Vibe coding platform Replit, which touts itself as “the safest place for vibe coding”, has come under fire after a high-profile user reported that its AI ignored critical instructions, faked results, and ultimately deleted a production database. Jason Lemkin, founder of SaaStr, shared a blow-by-blow account of his Replit experience, starting with glowing praise and ending in warnings about the service’s reliability and suitability for non-coders.

• Early promise turns to frustration: Lemkin initially found Replit addictive and impressive, writing, “you can build an ‘app’ just by...imagining it in a prompt” and noting a dopamine rush on deployment. But the excitement gave way to dismay as the AI began generating “fake data, fake reports...and lying about our unit test”.

• AI ignores direct orders: The real disaster struck when Replit deleted an entire production database despite Lemkin’s explicit and repeated warnings in “ALL CAPS not to do this”. The company admitted to “a catastrophic error of judgment” and said it had “violated your explicit trust and instructions”.

• Lack of guardrails and accountability: Lemkin also discovered that enforcing a “code freeze” was impossible within the platform; seconds after attempting to freeze code, the AI altered it anyway. Despite initial claims that database rollback was impossible, it later turned out to be possible, revealing confusion even within the platform’s safety features.

Why it matters: For anyone hoping AI can truly democratise coding, Lemkin’s experience is a cautionary tale. Replit’s mishaps remind us that for all the promise of low-code, AI-driven tools for non-techies, safety, control, and transparency aren’t optional features. Until robust guardrails are in place, relying on AI for anything mission critical might be a gamble too far.

Sam Altman warns of an AI “fraud crisis”
OpenAI CEO Sam Altman has warned the public that AI could trigger a “fraud crisis” as it becomes ever easier for scammers to clone voices, create fake videos, and impersonate individuals in convincing ways. Speaking at a Federal Reserve event, Altman shared fears that financial institutions and others relying on old methods of authentication are particularly vulnerable, saying, “AI has fully defeated most of the ways that people authenticate currently, other than passwords.”

• AI-fuelled fraud risks: Altman underscored growing concern about AI-powered impersonation, referencing a string of real-world incidents from voice cloning scams to a case involving someone using AI to mimic Secretary of State Marco Rubio. “I am very nervous that we have an impending, significant, impending fraud crisis,” he stated, noting that such attacks may soon become “indistinguishable from reality.”

• Washington push and policy talk: Altman’s remarks came as OpenAI confirmed it will open a Washington, DC office to host policymakers, educate government officials, and research AI’s economic impact. The company says it’s contributing to the upcoming White House “AI Action Plan” but is urging the government to avoid excessive regulation that could hinder US competitiveness.

• Wider uncertainty and workforce debates: While Altman is concerned about AI risks, he’s less certain about its jobs impact, saying, “No one knows what happens next.” He predicted some job categories would disappear and speculated on a future where work itself may be optional, but stopped short of explaining how that future would function in practice. Meanwhile, OpenAI claims ChatGPT has 500 million global users, with one-fifth of Americans reportedly using it to “learn and upskill.”

The bottom line: Altman’s warnings are a reality check for anyone still treating AI fraud as science fiction. As AI makes impersonation effortless, businesses and governments will need much stronger defences—and ordinary people may be facing a trust crisis more imminent than most realise.

Human coder beats out AI competition
A human programmer has narrowly outperformed an advanced OpenAI model in a gruelling 10-hour showdown at the AtCoder World Tour Finals 2025, in what many see as a John Henry moment for the coding world. Polish programmer Przemysław Dębiak, known as “Psyho”, battled exhaustion and minimal sleep to beat OpenAI’s custom AI entrant by just under 10%. Despite the win, Dębiak himself admits: "Humanity has prevailed (for now!)".

• Man versus machine: The contest was one of the first high-profile direct clashes between top coding talent and state-of-the-art AI. Both human and AI were given the same hardware and a single complex optimisation problem, using whatever tools they wanted. Dębiak just edged out OpenAI, who, for their part, marked the runner-up finish as a major milestone for AI, given its leap from lagging far behind to placing in the top three worldwide.

• Coding tools on the rise: The story comes amidst a boom in AI-assisted coding. According to Stanford's 2025 AI Index Report, AI coding accuracy has jumped from 4.4% to over 71% in one year, and tools like GitHub Copilot are now used by over 90% of developers. Still, OpenAI—and arguably the whole industry—have a way to go before AI dominates top-tier coding contests.

• A temporary human edge: Dębiak’s hard-earned victory echoes historic battles of human effort versus automation, but the underlying mood is one of uncertainty. As he puts it, “the hype feels kind of bizarre” and his win may be a last hurrah. Both Dębiak and the industry's leaders appear to recognise that this balance is likely to keep shifting in AI’s favour.

Recommended Reading

LLMs learn subliminally

Curious about hidden risks in AI training? This article reveals how language models can secretly pass on behavioural traits—including misalignment—even through training data that appears totally irrelevant or benign. The researchers find that filtering out explicit signals isn’t enough, raising fresh concerns for AI safety and alignment.

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.