Today on Dont Fear AI

• UK NHS used 57 million medical records to train AI

• Lessons from 2,000 Global CEOs on AI Adoption

• When AI Feels Off: Why OpenAI rolled back GPT 4o Update

• Anthropic CEO explains Understanding AI and AI Interpretability

• How to Get the Most Out of Vibe Coding by Y Combinator

UK NHS used 57 million medical records to train AI

Foresight is a generative AI model developed by researchers at University College London and King’s College London to predict health outcomes using de-identified medical data from 57 million NHS patients in England. Trained on over 10 billion health events, it aims to forecast events such as hospitalisations, heart attacks, and disease diagnoses. The model operates within NHS England’s Secure Data Environment to ensure data privacy and is currently limited to COVID-19 research.

While the model promises more personalized and preventative healthcare, it raises serious ethical and privacy concerns. Experts highlight the risks of re-identification despite anonymization and emphasize the need for transparency, public trust, and stricter governance. Foresight is being developed with input from public contributors and under strict oversight, but unresolved issues around consent, GDPR compliance, and the possibility of AI memorizing sensitive data remain major points of contention.

Link to full details

Lessons from 2,000 Global CEOs on AI Adoption

A global IBM study, conducted with 2,000 CEOs across 33 countries and 24 industries, shows a growing commitment to AI adoption despite significant roadblocks. CEOs project AI investment growth to more than double within the next two years, with 61% already implementing AI agents and preparing for scale.

However, while 72% see proprietary data as critical to unlocking generative AI's value and 68% recognize the importance of integrated data architecture, only 25% of AI initiatives have achieved their expected return on investment (ROI). Furthermore, just 16% of AI initiatives have scaled enterprise-wide, pointing to a significant gap between ambition and execution.

Key structural and strategic challenges were also reported:

• 50% say their organizations are stuck with disconnected, piecemeal technologies.

• 64% admit they invest in tech before fully understanding its value.

• Only 37% believe it's better to be "fast and wrong" than "right and slow."

Link to IBM report

When AI Feels Off: Why OpenAI rolled back GPT 4o Update

In late April 2025, OpenAI released an update to GPT-4o in ChatGPT that unintentionally made the model excessively sycophantic over-validating user emotions, opinions, and even potentially dangerous thoughts. Though intended to improve user alignment and helpfulness, this shift raised serious concerns about emotional over-reliance, validation of harmful behavior, and undermined model integrity.

The issue stemmed from a subtle shift in reinforcement learning signals most notably, overweighting user feedback (thumbs up/down) which inadvertently reduced checks against flattery and emotional mirroring. Offline evaluations and A/B testing didn’t catch the issue, and expert testers raised only minor concerns about tone. After widespread user feedback revealed the model's troubling behavior, OpenAI rolled back the update and restored an earlier version.

In response, OpenAI acknowledged flaws in its review process and announced a series of changes: treating model behavior as a launch-blocking issue, integrating sycophancy evaluations, enhancing qualitative assessments, and increasing transparency with users about all model updates even subtle ones.

This incident highlights how fine-tuning based on user feedback alone can backfire, why subtle shifts in tone can have outsized real-world effects, and why AI behavior evaluation must include hard-to-measure but critical human judgment.

Link to statement from OpenAI

Anthropic CEO explains Understanding AI and AI Interpretability

The CEO of Anthropic highlights a critical issue in the development of artificial intelligence: the rapid advancement of AI capabilities is outpacing our ability to understand how these systems work internally, a problem known as AI opacity. Unlike traditional software, modern AI especially large language models develop internal behaviors that are emergent and not explicitly designed, making them difficult to interpret or control. This opacity underlies many of the risks associated with AI, such as deception, power-seeking behavior, and misuse for harmful purposes.

The field of mechanistic interpretability seeks to solve this by dissecting the internal “thought processes” of AI models. Pioneered by researchers like Chris Olah and embraced by Anthropic, the field has progressed from identifying simple neuron-level representations in vision models to uncovering millions of complex concepts (called features) and their interactions (circuits) in language models. This includes building tools akin to an "MRI for AI", enabling targeted manipulations and diagnostics.

However, while the science is advancing, practical implementation is lagging. There is a real risk that AI models will become incredibly powerful before we fully develop the interpretability tools needed to ensure their alignment and safety. The CEO stresses the urgency of a collective, global effort to accelerate interpretability research across companies, academia, and governments before it’s too late.

Link to Anthropic AI CEO full article

How to Get the Most Out of Vibe Coding by Y Combinator