Thank you for keeping us company as we have tried to stay abreast of the full-throttle development of AI in 2021.
We’ve witnessed the massive hype of large language and image models, followed by doubts regarding bias, sustainability and the growing power of big tech. Then finally moves towards smaller and more efficient architectures. And we’ve watched with a mixture of hope, cynicism and trepidation as world leaders sought to take tackle the climate crisis.
At the end of the year, we send season’s greetings and a relaxing holiday so that we may return refreshed and re-energised to address the challenges that await us in 2022!
Marcel and Graham
P.S. Our next issue will be on Tuesday 4 January.
Packed inside we have
- OpenAI augments their language model with web surfing
- A single DeepMind AI model that plays chess, Go and poker
- The human face of management by algorithm
If you would like to support our continued work from £1 then click here!
Graham Lane & Marcel Hedman
Key Recent Developments
Improving the factual accuracy
of language models through web browsing
What: OpenAI has developed a new technique to more accurately answer open-ended questions based on their GPT-3 large language model. The prototype copies how humans operate on the web, searching, following links, and scrolling through web pages. The prototype cites its sources, thereby making it easier to address factual accuracy. Humans assess the truthfulness of the AI answers against reliable sources. Furthermore, the quality is evaluated by humans comparing the AI-generated answers to human-generated answers.
Key Takeaways: The enhanced system performs better than the underlying GPT-3 system. The approach is part of a trend of seeking engineering solutions to make large language models more efficient and accurate. There has been much development in this area in the past year, summarised in this VentureBeat article.
DeepMind makes AI system that can play poker, chess, Go, and more
What: DeepMind has launched a new single system that can play both complete and incomplete games. Complete games such as chess represent a finite space in which each move can, theoretically, be known in advance. Incomplete games, such as poker, are infinite in the sense that players might bluff or collude with one another, and the AI must account for this.
Key Takeaways: Incomplete games resemble many areas of human activity such as negotiations or volatile problems such as logistics planning. However, complex incomplete games such as NetHack (a roguelike dungeon game) remain a significant challenge for AI systems. In the recent NetHack Challenge, symbolic AI agents (incorporating rules and human experience) significantly outperformed deep neural network agents. And humans out-performed them both.
Paper: Player of Games
How a facial recognition system potentially failed to recognise a driver of colour and may have cost him his job
What: Uber requires drivers to submit selfies whilst working in order to prevent fraudulent shift swapping with unlicenced drivers. A dark-skinned driver, called Pa, claims that his account was deactivated without warning. He believes it may be because the facial recognition system struggled with his dark skin in poor lighting. He has been unable to get the account reinstated. In response, Uber states that all decisions regarding deactivation are undertaken by three trained human reviewers.
Key Takeaways: The article provides a human face to the growing concern both in Europe and the UK about management by algorithm. Campaigning organisations are calling for transparency, accountability, redress, respect for privacy and an end to the exploitation of workers’ data.
A controversial pilot using AI, facial recognition and thousands of CCTV cameras tracking the movement of people infected with Covid gives rise to privacy concerns.
Private companies, research institutions and government agencies are now obliged to report sensors deployed in public spaces.
A campaigner warns that "the pace of technology is really beginning to outpace the rate of diplomatic talks”.
Other interesting reads
News of a unique, high-quality labelled dataset supporting AI research by tracking progress against Sustainable Development Goals.
Almost instantaneous structural changes in a protein excited by light have been reconstructed using machine learning and quantum mechanics.
Report of a “major milestone” in AI recognising information that is not present in its training data.
An example of the problem to be solved is a military targeting system fell from 90% accuracy to 25% following a subtle tweak in the system configuration.
Data scientist - AxionRay
Axion are looking to hire a talented NLP DS lead as they enter hypergrowth. Axion is a stealth AI decision intelligence platform start-up working with electric vehicle engineering leaders to accelerate development, funded by top VCs.
Comp: $100k – $180k, meaningful equity!
If interested contact: email@example.com
Cool companies found this week
Immunai - uses single-cell genomics and machine learning to discover and develop immuno-therapies. The company raised $215 million in round B funding to develop its data atlas of annotated single-cell immune data
Serve Robotics - recently announced a robotic delivery service for Uber Eats in San Francisco and raised $13 million in seed funding led by strategic investor Uber Technologies
And Finally ...
Stuck for Christmas present ideas - how about an electric flying car?
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.
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!