Category Archives: AI

How will AI impact teachers?


ChatGPT has recently triggered tremendous excitement about AI and its potential impact on education. Much interest has focused on the learner experience, including the ability to personalise learning. There have of course also been concerns around cheating and plagiarism.

However, AI also has the potential to super-charge teachers.

According to McKinsey, 20-40% of current teacher time comprises tasks that could be automated. They estimate that teachers could re-direct approximately 13 hours per week towards activities that raise student outcomes and increase teacher satisfaction.  The tasks of preparing lessons, administration, evaluation and feedback are flagged as high potential for AI.

Love’s Labour’s Lost

These results echo those of Sanoma’s Learning Impact Survey, in which teachers indicate a desire to go digital in those areas which were most labour intensive, flagging essentially the same areas.  This suggests both that the opportunity is in these tasks and that the profession is looking for solutions.


Teaching profession under pressure

The teacher is arguably the most positive intervention in education.  However the teaching profession faces significant challenges.  UNESCO estimates an additional 69m teachers need to enter the profession by 2030 to fulfil global demand.  In some parts of the world, teacher turnover is high, for example in parts of the USA annual teacher turnover exceeds 15%.  In the UK more than 80% of teachers are considering leaving the profession due to dis-satisfaction.

Higher impact & happier teachers needed!

Furthermore, on average teachers spend only half of their time actually teaching.  This represents not only lost productivity from the core task but is also demotivating for many teachers whose passion is to teach rather than the ancillary tasks around it.  Enabling teacher workflow could therefore not only increase productivity but also make the profession more attractive.


$400bn impact & opportunity

The opportunity to solve this productivity gap is huge.  Measured in terms of financials, assuming global spending on education to be some $6trn, of which 45% is on K-12 education,  and of which 75% is spent on staff salaries, this implies a global spend on teaching/staff salaries of some $2trn per year.  A 20-40% uplift in productivity through AI could arguably be worth some $400-800bn per year in terms of paid and unpaid output!  Which is not to say that this is a saving governments could make or a revenue that education companies could earn, because a significant slice of that value should rightfully return to teachers through higher salaries and quality of life, and another part would rightfully get re-directed to teacher-student interaction to increase outcomes and professional satisfaction.


Help the teacher to focus on teaching!

It’s my belief that the teacher will continue to be the killer app in education, and that the biggest opportunity to make not only a positive impact on learning and teaching in K-12 but also to build a successful business, is to enable the workflow of the teacher.  Probably by combining it with the other side of the same coin: the learn-flow (learning experience) of the student.  

Looking forward >>

It will be exciting to see how we deploy AI in the coming years for a positive impact on learning. Looking forward >>.

DeepMind uses AI to understand life.

Life at the molecular level that is.

Last week saw the breakthrough news that Google has essentially solved the protein folding problem with AlphaFold from DeepMind. I was especially interested in this since this was the area of my PhD.

Function follows structure

Proteins carry out a variety of functions from DNA replication to catalysis to structuring the cytoskeleton.  Each protein is built up from a unique sequence formed from 20 different amino acids. Some 200M sequences are currently known, growing by about 30M per year. The chain of amino acids folds into a unique 3D structure.  This structure determines its functionality.

Prediction: the shape of things to come

Some 170,000 protein structures have been determined to date, and DeepMind has used this dataset to create an algorithm which can predict the 3D structure of a protein based only on its sequence of amino acids, to the same level of accuracy as if actually measured using a technique such as X-ray crystallography.  A reasonably sized protein might take as many as 10300 different shapes, so that’s quite a prediction!

This is relevant because understanding the 3D structure of a protein can inform its function and arguably mis-function, thereby potentially accelerating the rational design of interventions such as drugs against disease states for example.  With 200M proteins in scope, the potential for scientific discovery is massive.

Now we can look to Google not only in search of pizza, but also for the elixir of life.

Determined structures

25 years ago I calculated the 3D structure of a protein essentially by hand (serine proteinase human stefin A, see below) – with a simulated annealing protocol using distance and angle constraints obtained from high-resolution Nuclear Magnetic Resonance spectroscopy.  This took 2.5 years! Multiplied by 200M proteins, it would take quite some effort to map the universe of proteins. The task has now been reduced from years to hours!

Family of 17 solution structures showing the backbone atoms of serine proteinase human stefin A. The protein has a well-defined global fold consisting of five anti-parallel β-strands wrapped around a central five-turn α-helix. There are two flexible regions in this structure which are two of the components of the “tripartite wedge” that docks into the active site of the target proteinase. These regions, which are shown to be mobile in solution, are the five N-terminal residues and the second binding loop. In the bound conformation they form a turn and a short helix, respectively.