Welcome to Nural's newsletter where you will find a compilation of articles, news and cool companies, all focusing on how AI is being used to tackle global grand challenges.

Our aim is to make sure that you are always up to date with the most important developments in this fast-moving field.

Packed inside we have

  • DeepMind's generalist agent tackling 600 tasks
  • Google's AI Test Kitchen
  • plus, Tractor kill-switches, a double-edged sword?

If you would like to support our continued work from £1 then click here!

Graham Lane & Marcel Hedman

Key Recent Developments

DeepMind unveils a generalist AI agent: tackling 600 tasks

A generalist agent
A Generalist Agent
Inspired by progress in large-scale language modelling, we apply a similar approach towards building a single generalist agent beyond the realm of text outputs. The agent, which we refer to as Gato, works as a multi-modal, multi-task, multi-embodiment generalist policy. The same network with the sam…

What: Taking large-scale language models as a starting point, DeepMind researchers used a similar approach to build a single, generalist agent called Gato that can carry out over 600 tasks. This single agent, with the same weights, can play Atari, caption images, chat, stack blocks with a real robot arm and more. The agent decides based on context whether to output text, joint torques, button presses, or other tokens.

Key Takeaways: The announcement prompted a debate about whether this is a step towards more powerful Artificial General Intelligence. The consensus seems to be that this generalist agent shares the same problems as the large language models on which it is based, namely a lack of reasoning and the fact that new learning requires retraining. A research scientist on the technical AGI safety team at DeepMind joined the discussion describing a “Pareto frontier” in which specialist models will continue to perform better on complex tasks but these will be subsumed by generalist models over time.

Google is beta testing its AI future

Google is beta testing its AI future
A new app will give select users access to Google’s latest AI language models.

What: Google announced a new app called “AI Test Kitchen” at its recent I/O conference. This will enable selected individuals to probe Google’s latest language models for bias and errors. The app provides structured access to the language models and a clear path for user feedback.

Key Takeaways: Google had a torrid time recently suffering reputational damage from mishandling internal complaints and sacking staff who raised issues about bias in Google language models. It is now engaging a broader community in testing and identifying problems. It is, however, one thing to identify the problems but another to address them. There is increasing research into mitigation strategies. However, new models are still being launched with clear statements about shortcomings but without any clear solutions.

About those kill-switched Ukrainian tractors

Article: About those kill-switched Ukrainian tractors

What: John Dere tractors worth millions of dollars were reportedly looted from Ukraine by Russian troops and transported to Chechnya. However, the high-tech tractors were disabled with a remote kill-switch and are currently unusable.

Key Takeaways: Behind this news is a back-story about the power that John Dere holds over farmers. John Dere owns all the telemetry and location data generated by the farmers using the tractors. This was monetized by selling it not only to the seed company Monsanto (now Bayer) but also to private equity firms trading in the agricultural futures market. Furthermore, all repairs and modifications need to be carried out by a John Dere technician using a software unlock code. Finally, John Dere controls the licence to the tractor software without which the tractor is unusable. The kill-switch may be a welcome feature in some circumstances, but it could also be a serious security vulnerability. And it is a way of enforcing stringent commercial terms on the tractor “owner”.

AI Ethics

🚀 Google is using a new way to measure skin tones to make search results more inclusive

🚀 Let’s create AI with integrity: AI for Good webinar on Tuesday 24 May

Other interesting reads

🚀 Using AI to improve the health and wellbeing of people with learning disabilities

🚀 Why AI is vital in the race to meet the Sustainble Development Goals

🚀 Aston University develops AI traffic lights in bid to cut queues

Cool companies found this week

Machine learning

Hugging Face - is a community-oriented, machine learning platform hosting 100,000 pre-trained models and 10,000 datasets for NLP, computer vision, speech, time-series, biology, reinforcement learning, and chemistry. The company has raised $100 million in round C funding with the aim of becoming the “GitHub of machine learning”.

Human-computer interaction

Inflection - is an AI-first company aiming to redefine human-computer interaction. It is led by LinkedIn and DeepMind co-founders and was referenced in our Newsletter #68. The company has now raised $225 million in venture funding to use AI to help humans “talk” to computers.

Clinical trials

Unlearn - aims to accelerate clinical trials by using AI, digital twins, and novel statistical methods to “enable smaller control groups while maintaining power and generating evidence suitable for supporting regulatory decisions”. The company has raised $50 million in round B funding.

And finally ...

How might this swarm be used

AI/ML must knows

Foundation Models - any model trained on broad data at scale that can be fine-tuned to a wide range of downstream tasks. Examples include BERT and GPT-3. (See also Transfer Learning)
Few shot learning - Supervised learning using only a small dataset to master the task.
Transfer Learning - Reusing parts or all of a model designed for one task on a new task with the aim of reducing training time and improving performance.
Generative adversarial network - Generative models that create new data instances that resemble your training data. They can be used to generate fake images.
Deep Learning - Deep learning is a form of machine learning based on artificial neural networks.


Marcel Hedman
Nural Research Founder

If this has been interesting, share it with a friend who will find it equally valuable. If you are not already a subscriber, then subscribe here.

If you are enjoying this content and would like to support the work financially then you can amend your plan here from £1/month!