Beyond language with stochastic parrots

Leave a reply

Decorative image of a summer flowerSome months ago, I wrote in unflattering terms about artificial intelligence applications (AI apps) and large language models (LLMs), (see ‘Ancient models and stochastic parrots‘ on October 1st, 2025).  My view is changing, probably as AI apps develop and my user skills improve.  I have started using a couple of different free AI apps as research assistants in three ways.  First, when I am writing administrative documents, such as a job description for a Coordinator of AI in Education, for which a job title was sufficient for the app to generate a first draft that only required light editing and tailoring to the specific context.  Second, using a different AI app, to answer questions about phenomena which have allowed me to construct explanations for observations made of new and, or, complex systems – I could have delved into textbooks and monographs or searched research articles but AI does this much more quickly.  The third way I have used AI apps is to identify gaps in knowledge that could be fruitful topics for research.  This is a more difficult task because AI apps only know about stuff they can find on the internet in the form of language or text.  Hence, I have to ask questions with answers that reveal something unknown or not understood.  This is not straightforward because LLMs are fundamentally constrained by language.  In ‘The Years’, Annie Ernaux wrote that ‘language will continue to put the world into words’.  Yann LeCun, Meta’s former chief scientist, has suggested that to understand how the world works, a model would need to learn from videos and spatial data, not just language, and that without this type of learning human-level intelligence is impossible.  He has set up a new company, Advanced Machine Intelligence Labs, to do just that.  Language is used by people to describe the world from their perspective which might be inaccurate, incomplete or distorted and that can mislead LLMs.  However, using AI apps we can also ‘distort’ videos of the world, so that machine intelligence will have to be based on direct observation of the real-world, which after all is the approach that science attempts to use.

Source:

Yann LeCun, Intelligence is really about learning. FT Weekend, 3-4 January 2026

Annie Ernaux, The Years, Fitzcarraldo Editions, London, 2018.

Extra! Extra!

Leave a reply

Extra! Extra! Read all about it!  As newspaper vendors used to shout.  The Pint of Science Festival is happening across the UK in the week beginning Monday 18th May for three evenings in venues in 43 locations.  I am talking on the first evening, May 18th, in Lime Street Social (51 Lime Street, L1 1JQ) on ‘I Sell Here, Sir, What all World Desires to have – POWER’.  My title is a quote from Matthew Boulton, who with James Watt, set up a factory in Birmingham to produce steam engines in the 18th century.  I am going to talk about producing nuclear power units in a factory (see ‘Commoditization of civil nuclear power’ on June 5th, 2024).  If you would like to come to the event and hear three other speakers besides me and have a pint or two then please register at https://pintofscience.co.uk/events/liverpool/.

Face-to-face with a polar bear

Leave a reply

Decorative photograph showing helicopter over glacierWhat would you do if you found yourself face-to-face with a polar bear?  Run?  Probably not a good idea because polar bears regard anything that runs away as a potential meal and they can reach speeds of 40 kph or 25 mph, so they can easily outrun you.  This question is probably theoretical one for most people, even though climate change is reducing the area of sea ice and forcing polar bears to forage on land, as far south as Newfoundland.  However, it was one that we debated when planning our expedition to Greenland where we did expect to encounter them [see ‘Perched blocks and muskoxen‘ on February 4th 2026].  We considered taking firearms; however, to be effective you need a high-powered rifle and it would have been challenging to transport several from the UK to Greenland so that each group had one available when we were not working together.  More importantly, we were visiting their domain and we had no desire to harm any polar bears.  As an alternative, it was suggested to us that, like us, polar bears find it difficult to run downhill at speed and so this was a potential escape route but, of course, only if you are on a steep hill!  In common with most animals, polar bears are frightened of fire so our chosen means of defence was to remove our shirt, douse it in the liquid stove fuel we were carrying, wrap it around one end of our ice-axe and set it alight as a flaming torch.  We did see polar bears while we were in Greenland, always in the coastal regions and never near to our base camp.  We kept our distance and moved cautiously so that our defence plan never had to be implemented.  Nowadays, the internet provides advice for polar expeditions, see for example the guide of the International Polar Guides Association.

Decorative photograph showing basecampImages: helicopter over Bersaerkerbrae glacier after delivering supplies to base camp (top) and base camp on the western lateral moraine of the Bersaerkerbrae glacier (right).  We never got close enough to a polar bear for a good photograph.

Experiencing success vicariously

Leave a reply

Decorative image of a graduation ceremonyThe final PhD student for whom I will act as lead supervisor is scheduled to finish this month.  I have graduated forty PhD students since I was appointed a lecturer in 1985.  I am still in touch with many of them – they are divided between industry and universities with a bias towards industry (about 60%).  For the first twenty years, I was a sole academic supervisor often with an industrial supervisor providing support.  Then I moved to the US where a PhD committee provides supervisory guidance to the student and supervisor.  By the time I returned to the UK, about fifteen years ago, it had become accepted practice to appoint a second supervisor for each PhD student.  So, although I decided a couple of years ago not to accept any new PhD students as lead supervisor, I am acting as second supervisor for five students.  This is a great role since you have less responsibility, but you are engaged with the exciting research.  The topics vary from understanding the nanoscale mechanics of particles interacting with cells (see, for example, ‘Label-free real-time tracking of individual bacterium‘ on January 25, 2023 through to ‘Structural damage assessment using infrared detectors in fusion environments‘ on March 15, 2023), and just starting this year, innovative methods for communicating confidence in computational models.  Although the research is exciting, at a training session for supervisors during the CDT Winter School that I attended in January (see ‘Experiencing success vicariously‘ on January 7, 2026), we discussed our roles as supervisors and in particular that the research project is not the principal outcome of the PhD.  Instead, the development of the PhD student is the principal outcome.  It’s all about nurturing and mentoring people and the reward is experiencing their success vicariously.

Image: still from a video of a graduation ceremony at the University of Liverpool on December 9, 2025.  As Dean of the School of Engineering, I am at the lectern presenting PhD graduates, but I am hidden behind the Vice-Chancellor who has his back to the camera on the extreme left of the image.  You can watch the video at https://www.liverpool.ac.uk/graduation/the-ceremony/watch-graduation/catch-up/school-of-engineering/9-december-2025-10am/ .