Tristan was last on the show some three years ago in LFP063 which has been solidly in the Top5 most downloaded episodes of all time even since. Today he joins us to update us on what is hot in AI/ML – and there’s a lot 😉 – along with predicting commodity prices.
Tristan is now CEO of ChAi who are creating insurance products for the commodity markets and at the forefront of some of the most interesting developments in AI/ML today.
In this episode we cover some of the key important developments and in particular dive into some amazing examples of how Alt Data is crucial to the developments. The focus in AI/ML is moving away from algorithms and onto data.
Topics discussed include:
- recap of some of my takeaways from LFP063 (if you haven’t heard it check it out – its a must-listen on AI/ML):
- a dive into the slipperiness of intelligence even when it comes to people (judging by twitter and all too many politicians in the past three years intelligence is disappearing from human beings) ML not AI (its not “intelligent” its a pattern matching algorithm in a computer arrived at by a system programmed to be able to find a non-linear function that maps input sets to output sets)
- ML techniques been around for a long time just massive increase in computing power
- “interpretability” its importance and implications – your algorithm might say outperform on credit by excluding lending to podcasters but that may not be legally permitted
- the folly of taking lots of data and throwing lots of freely available tools at it – the vital nature of domain knowledge (ie can apply wisdom)
- a whole host of case studies that Tristan had worked on from predicting cardiac events to position ambulances
- Persepolis, Persia and Iran – ancient and recent history
- what Aryans used to be (nothing like blonde Germans)
- oil was first discovered in Iran
- the removal of democracy in Iran at the behest of the US/UK
- the origins of ChAI
- the lead that China has in AI/ML – I wonder why… 😉
- predicting the price of Copper
- predicting other metal prices
- what “prediction” means in this context – a spectrum of outcomes
- industrial metals can be well predicted; oil is far far harder due to the geo-politics around it
- if a black box predicts prices well does this not change – ie become – the market?
- different types of predictions and how the use case affects them
- when models are useful and when they are not – the underling discontinuous nature of the actual phenomena which lead to the price (so never believe the “n standard deviations” line – there is “normal” behaviour” and “abnormal behaviours” and there may be no relationship between them
- “Machine Learning and AI have their limits and we’ve got to be humble about where those limits are.”
- “People are seeing AI as a panacea to all the world’s problems, crediting it with far too much intelligence when it isn’t actually there.”
- the complexity of factors in the oil market
- human wisdom + AI/ML
- “We don’t have a black box system but a glass box system – you can go in and unpick all the assumptions that are in the model.”
- to increase the confidence in AI the models need to be much more explainable
- regulatory moves towards explainability
- the algorithms in AI have become really good off the shelf – it has become harder to differentiate at that level
- hence the importance of alt. datasets
- examples of satellite data and shipping data – how they are used (super cool!)
- computer vision is now a commodity in the cloud
- “Combining all this alt data gets you leaps and bounds ahead”
- can one in extremis calculate the price if one knows enough about supply and demand?
- “There has been a big democratisation in access to alt. data.” – even startups can access it now…
- the real reasons ML is used these days…
- Evolutionary/self-learning systems – notably Deep Mind’s AlphaGo Zero, AlphaChess Zero, Starcraft
- Deep Mind’s practical application in reducing power usage by 17% at Google’s server farms
- are these techniques evolutionary techniques? Formal and informal answers to this.
- what is transfer learning? Examples.
- a one-year old can be shown one dog and recognise all dogs thereafter – how come a computer could never do this? How does the child generalise so well?
- generalised AI vs domain-specific at present
- the importance of deep domain knowledge
- how insurance’s needs differ
- are we approaching Peak AI?
- what the future holds…
- ChAI’s mission and target clients
- how they have grown so fast
And much much more 🙂
Share and enjoy!