Mathematicians and scientists have had their world rocked in the last few weeks, as LLM-enabled systems made significant progress on long-standing open problems. Just to take a couple of examples from OpenAI:

• The Navier-Stokes problem – whether smooth, physically realistic fluid flows can blow up into an infinite singularity – is now considered settled, although details of the proof are still being worked through (yes, they can).
• The Riemann Hypothesis is still unproved, but OpenAI got close enough (showing that the zeta function has no non-trivial zeros with a real part greater than 7/8) to validate the vast majority of the many mathematical proofs that depend on it being true.

Meanwhile, Claude autonomously redesigned eight quantum-computing primitives, sometimes outperforming the most efficient published implementations by more than an order of magnitude, and a 16-billion-parameter foundation model, Pinal, generated protein candidates from natural-language descriptions that were found in laboratory tests to demonstrate the intended functions.

There is a lot more. While the scientific world attempts to make sense of the new normal, let’s step back and ask a basic question: does this mean we can trust LLMs now?

In fact, nothing has changed. If you want to understand exactly why, take this short interactive LLM training course. It was generated for me by OpenAI’s current flagship model, GPT-6, and helpfully concludes by reminding us that:

Learned competence can coexist with memorisation, brittle generalisation, and hallucination. Fluent language and high token probability alone do not establish truth, reliable causal understanding, or subjective experience.

If you want auditable AI for defensible decisions, you still need Dedoctive.

Author

Keith Harrison-Broninski is an author, researcher, and keynote speaker specialising in cross-boundary collaboration, community antifragility, and technology for good. Keith’s awards include from Gartner for social enterprise and from the NHS for technology to replicate healthcare innovations.

Keith’s first book “Human Interactions” (2005) was described by Information Age as “the overarching framework for 21st century business technology”. Keith’s other books include two for Springer and three for the Workflow Management coalition. Keith’s most recent book “Supercommunities” (2021) was described by the Chief Executive of the RSA in his foreword as “Ranging from ancient history to economics to psychology to public policy ‘Supercommunities’ is both authoritative and highly readable. It puts our current challenges in context, shows why change is necessary and provides a trove of practical ideas for change makers.”

Vint Cerf, co-inventor of the Internet, in a second foreword, wrote “Supercommunities offers a path away from social and economic meltdown … We will need to replace short-term thinking with long term planning and execution if we are to regain upward motion towards common benefit for everyone on Spaceship Earth. To begin, read this book!”

Keith currently focuses on creating the Internet of Communities via his company’s revolutionary technology “Dedoctive” – a fundamental AI innovation, academically-validated for the MoD as fully trustworthy, that empowers reliable access to complex information for all.

Keith is also a jazz pianist, traditional folk musician, and classical composer. He has released 9 albums and performs regularly with a wide range of other musicians.

More information

Email Keith Harrison-Broninski: khb@dedoctive.ai