AI Reality Check — Issue 16

AI Reality Check - Issue 16

Friday 10 July 2026 | Kaye Nicholson | GrowthZone AI | growthzoneai.co.uk

Born analogue. Raised digital. 30 years of real business experience. Now explaining what AI actually means for work.

━━━━━━━━━━━━━━━━━━━━━━

1. Anthropic Plans to Develop Its Own Drugs

What happened: Anthropic plans to develop pharmaceutical drugs. The company has been building wet labs, hiring biologists, and recruiting staff from Big Pharma and academic institutions. It has launched Claude Science, an AI workbench for scientists working in biotech and pharma. The next step is developing drugs directly. Anthropic declined to say which diseases it will target first.

Why it matters: Anthropic sells AI tools to pharmaceutical companies. It is now planning to compete directly with those same companies in the drugs market. This is a significant change of position for a company founded to make AI safe and beneficial, and it raises questions about whether being a tool supplier and a product competitor is a sustainable combination for an organisation that needs pharmaceutical companies as customers.

Who it's for: anyone working in or around life sciences, pharma, or biotech. Anyone evaluating Anthropic's long-term strategic position. Any business thinking about the expanding ambitions of the AI companies whose tools they depend on. Anyone working with AI in a scientific context who wants to understand what Claude Science is and what it offers.

━━━━━━━━━━━━━━━━━━━━━━

2. Builder.ai: $445M and No Real AI

What happened: Builder.ai, a British startup, raised approximately $445 million from investors including Microsoft and the Qatar Investment Authority. It had a $1.5 billion valuation and claimed $220 million in annual revenue. In reality, revenue was closer to $50 million and the apps it built were not created by AI. More than 700 engineers working in Asia and Eastern Europe manually built every product on the platform. The Wall Street Journal raised questions about this as early as 2019. The company rebranded and moved on. In 2025, Builder.ai filed for insolvency. Employees were told in a remote town hall.

Why it matters: AI washing is not only a small startup problem. It happened at unicorn scale with major institutional investors and a British company. The warning signs were published in a major newspaper years before the collapse. In the AI boom, asking the basic questions became unfashionable. It should not have.

Who it's for: any UK business evaluating AI suppliers, tools, or platforms. Any organisation thinking about due diligence on AI claims. Founders, buyers, and procurement leads who want a framework for what questions to ask before trusting an AI product claim.

━━━━━━━━━━━━━━━━━━━━━━

3. Ford Rehired 300+ Engineers After AI Cameras Failed

What happened: Ford had replaced more than 300 veteran quality engineers with AI-powered cameras for quality inspection on its manufacturing lines. The cameras turned out to be less effective at quality control than the experienced human inspectors they replaced. Ford brought those engineers back, combined their expertise with available technology, and topped the JD Power vehicle quality rankings for the first time since 2010.

Why it matters: Ford's experience illustrates the gap between "AI can do this task" and "AI can replace the person who has done this task for 20 years." The cameras could inspect. They could not inspect as well as someone with years of contextual judgment and pattern recognition in that specific environment. The human expertise, combined with tools, outperformed the pure automation attempt.

Who it's for: any business owner or manager who has been under pressure to automate processes that involve experienced human judgment. Anyone in manufacturing, professional services, or any field where contextual expertise matters. HR and operations professionals thinking through AI deployment. Anyone who has been told automation is the inevitable direction of travel.

━━━━━━━━━━━━━━━━━━━━━━

4. OpenAI Built Its Own Chip and Split GPT Into Three

What happened: OpenAI designed a task-specific AI chip called Jalapeño in partnership with Broadcom. It is built for inference workloads, meaning running AI models at scale, rather than training them. The goal is to reduce OpenAI's dependence on Nvidia ahead of its IPO. Separately, OpenAI previewed GPT-5.6 as three distinct models: Sol for the most complex tasks, Terra for balanced everyday use, and Luna for speed, volume, and cost efficiency.

Why it matters: the Jalapeño chip confirms that AI companies are moving toward owning their full technology stack rather than buying key components from suppliers. The three-model GPT structure confirms that the tiered model approach is now the standard across the industry. The "right model, right task" skill applies to ChatGPT in exactly the same way it applies to Claude.

Who it's for: anyone using ChatGPT for business. Anyone managing AI spend across platforms. Tech leads and business owners choosing between AI providers. Anyone interested in where the AI industry's infrastructure is heading.

━━━━━━━━━━━━━━━━━━━━━━

5. ByteDance Found AI Can Get Smarter Without More Training Data

What happened: researchers at ByteDance published findings showing AI agents can double their learning rate every three months by interacting with real-world environments over long periods, rather than simply consuming more data during initial training. They built EdgeBench, a test suite of 134 complex long-horizon tasks including software engineering, scientific discovery, formal mathematics, and knowledge work, each requiring at least 12 hours of continuous operation.

Why it matters: the traditional AI scaling approach of adding more training data and more compute has practical limits. This research suggests a different path: AI that learns by doing rather than just by being trained. If this finding holds up at scale, it could sustain AI capability improvements past the point where the brute-force approach runs out of steam.

Who it's for: anyone following AI capability development. Business leaders thinking about where AI tools are heading in the next two to three years. Anyone trying to make sense of whether the AI improvement curve will plateau.

━━━━━━━━━━━━━━━━━━━━━━

Also this week: Micron broke ground on a $9 billion memory chip facility in Hiroshima to meet AI demand. Germany opened the world's largest power semiconductor facility at $5.7 billion. Midjourney asked a US court to force Disney, Warner Bros and Universal to reveal how they use AI internally. China's Zhipu AI released an open-weight model matching US systems on some cybersecurity tasks.

━━━━━━━━━━━━━━━━━━━━━━

This week in one sentence: Anthropic wants to make drugs, Builder.ai raised $445M on no real AI, Ford proved human engineers still beat AI cameras, OpenAI built its own chip and split its model into three, and ByteDance found a new way AI can keep improving.

Born analogue. Raised digital. 30 years of real business experience explaining what AI actually means for work.

— Kaye Nicholson | GrowthZone AI | growthzoneai.co.uk

Subscribe at growthzoneai.co.uk | Follow @GrowthZoneAI

#AIRealityCheck #GrowthZoneAI #UKBusiness #NorthEast