Introduction
There often seems to be a mismatch between AI-themed discussions and the decisions that each of us are making day to day. Is work as we know it over? Will AIs take over the world? Is AI a Very Bad Thing? No doubt, these are some of the most pressing philosophical and political questions of our age. However, if the domain that you control is smaller than the future of humanity then these discussions risk moving you further from a solution, not closer to one.
This newsletter aims to digest what is happening, to enable you to make decisions. We focus on: What matters? Each month we read and discuss the latest academic articles, product announcements, policy papers, cool demos & VC think pieces. Expect summaries, reactions and links to the originals.
This newsletter is for any decision maker that is considering:
Should we use AI tools? If yes, what for?
What does this mean for my workforce? How can I manage this change?
What are other people going to do with AI that we should be prepared for?
If that sounds like you then read on!
Each month’s newsletter will start with two sections most directly applicable to those grappling with AI-caused disruption. Later sections will zoom out & include topics we are working on, wider context & summaries of articles that are being hotly discussed in the AI world. If there’s something more that you’d like covered please let us know.
Contents
What Is GenAI Good For?
Evidence Of Impact - Productivity & Quality
Writing Code - But Software Engineers Do Much Else Besides
Agents & App Stores
How To Successfully Integrate GenAI With Existing Organisations?
What To Buy, What To Build
Existing Tools
Case Studies
Resources: FAQs, Handbooks/Playbooks
Our Recent Work
Organisational Policies & Workforce Changes
Future Of Work - A Webinar for HR Directors
Zooming Out
Policy & Regulation
How Widespread Is Adoption
Venture Capitalist Views
Learning More
Newsletters
Groups & Meetups
The Lighter Side
What Is GenAI Good For?
Knowledge sector jobs see an AI supported boost in both efficiency and productivity with the evidence base growing both broader and deeper. Distributional effects are dominating (in jobs varying from call centre agents to consultants) as GenAI tools help to bring newer & lower performing workers up to the level of the best performers, who still benefit but not as much. Nevertheless, employees in a study at BCG quickly became overly reliant on GPT4 and performed worse than unassisted employees on tasks such as re-evaluating datasets on the basis of interview responses, where GenAI tools were weak. Employees will need to become skilled editors, assessing and curating the work of AI tools, even as much of the first pass work is being automated. Call centre productivity gains of 14% on average. BCG study. BCG write up.
Adoption has been swift and successful amongst coders, reportedly quashing traffic to coder-support websites such as Stack Overflow. As with other knowledge work, the biggest gains are seen from taking people with low (or no) coding skills and enabling them to do work which previously required specialist training. Professional software engineers, however, spend a lot of time doing higher order skills such as integration and system design which currently are beyond even the best GenAI tools. An evaluation framework has been developed - at present, even state-of-the-art AI models resolve only the simplest real-world issues. Senior Software Engineer - Can AI Do My Job?. Evaluation Framework. No Code Data Science.
AI powered tools & agents have become radically easier to build and share. Just as we were getting used to the web interface (like ChatGPT) and API-based development (the way many teams have been accessing GenAI models to make usable products), OpenAI have released shareable agents to further disrupt the market. The agents, which OpenAI have (helpfully) called GPTs, work a bit like the apps for phones. One developer can build a tool, using the underlying model framework and some integrations, and share these tools as a finished product with other users. The days of “copy my prompt to get the results you want” are on the way out. Instead, we have tools which can give interactive feedback on student essays, carry out and summarise research or even automatically commentate on video footage. This area is moving very fast and will continue to improve. Intro to GPTs. Video commentary.
How To Successfully Integrate GenAI With Existing Organisations?
In many cases “wait and buy” is the right strategy, particularly for features which are going to be common to many businesses. The primary trend has been for big, existing software providers to bundle in GenAI features as fast as possible, squashing many would-be startups in the process. Consequently “buy” might rapidly become “continue to use” as Microsoft Office and Google Suite are rapidly becoming AI-infused. Nevertheless, there will still be fortunes made building new vertical products that do one thing really well and we will gradually change our workflows to incorporate these new products. To my mind, this is the most interesting commercial space at present. Bundling and Unbundling. AI Disruption. AI Opportunities.
What tools are out there? For supercharging existing workflows we still find GPT4 the best option for quality and value. It’s $24 / month from OpenAI, although at the time of writing signups were paused due to overdemand. Consider giving a small number of employees a subscription as a trial. We use it for: building quick tools to speed up repetitive tasks, research and summarisation, creating glossaries and reformatting existing materials. Beyond this Microsoft recently announced that Excel is soon to have Python (one of the most common coding languages) embedded into workbooks. ChatGPT written code, executed & shareable in the Excel sheets that businesses already use, looks to be a powerful combo and easily accessible with little additional training. Goodbye VBA. Python in Excel. Pause.
Case studies. Our friend Gianni Giacomelli from MIT’s Collective Intelligence Design Lab has released the tool that he developed to enhance creativity & ideation within his teams. We also loved ONS’ Stats Chat, which is being built in the open so other teams can learn from their experience and copy their code. Creativity tool. Stats chat. Public Sector Foundation Models.
Playbooks and Resources. The best of these cover essential background information, hard rules for keeping organisations safe, practical steps for implementation and risk mitigation procedures. Many of these have focused on the Public Sector but the lessons are applicable to a wide range of organisations. FAQs. Civic Organisations Handbook. Local Authorities Guide. Charities Checklist.
Our Recent Work
Three phenomena support a principles based approach to leadership. The sky-high levels of: the rate of change, the risk of disruption from AI underpinned competition & cyber attacks, and the prospect of productivity gains. Set guidelines for employees and encourage them to explore and find solutions, whilst remaining clear which boundaries must not be crossed. Typical top down, instructive methods are likely to be too slow to adapt in periods of such high change. Article.
HR Directors have an unparalleled opportunity to lead the change. People teams control the levers for training as well as many crucial employee policies. Together with organisations’ most senior leaders, they also are responsible for setting company culture. Hays and Paradigm Junction ran a Webinar to explore these issues. Webinar.
Zooming Out
Policy action is mainly focused on tech behemoths and governments themselves, in the near term. In early November both the UK and USA committed to maintaining significant AI Evaluation capacity in government whilst the announcement of two further summits means this won’t fall down the policy agenda in the near term. Action was mostly targeted at government itself (more reports and research) and the largest tech firms (who must now disclose if models get significantly bigger). Application specific regulation seems further away, unless your line of business is bioweapons. China was included, which is a necessary cooperative move, but National Security concerns continue to ring loud in Whitehall & Washington and this necessitates at least some interstate competition. There is rising self-confidence given the way the UK’s Taskforce was able to attract significant expertise by operating differently. How this is used and what lessons can be learned remain to be seen. Pre-Summit Discussion Paper. Post-Summit Declaration. White House Executive Order. Tom Westgarth on Future Questions.
Employees at all levels remain, on balance, sceptical. Workers fear job losses, which partly explains why executives are declaring loudly their intentions not to use AI at many firms, despite the productivity gains. This position is largely nonsensical since AI is already baked into many of the services we consume every day (there is lots of AI in Google search, for instance), with much more to come. Adoption. Executive Attitudes.
Silicon Valley is fiercely divided. Despite the push for much stricter AI regulation, there is a growing group who believe that slow technological progress is the major issue facing the West today. A manifesto from VC Marc Andreessen was briefly the talk of the AI-town. Worth noting, that on many other issues, particularly Nuclear Fusion and Science & Innovation funding, the two sides of this debate are close allies. AI Safety/Regulation is a wedge issue amongst the Silicon Valley community. A more dispassionate, detailed take and set of predictions comes courtesy of Nathan Benaich and the Air Street Capital team, whose annual State of AI Report is as good an overview as any out there. Manifesto. State of AI Report.
Learning More
The best writing on GenAI and its challenges is being done by individuals writing newsletters or on LinkedIn, with high quality analysis taking much longer to feed through into the mainstream.
We always enjoy reading:
Ben Evans for tech analysis & what to think about the latest news
Andrew Ng for a top down view from an industry veteran
Jack Clark for the latest research in AI
If you’d like to connect with others grappling with the same problems then the following communities are great places to start.
James Fox, a good friend, has just joined the London Initiative for Safe AI (LISA) as the Research Director. LISA is a new organisation that supports impactful AI safety organisations, upskilling programs, and independent researchers based in the UK and Europe. He is keen to hear from anyone with an interest in shaping their vision and priorities going forward. Email
Civic AI Observatory. A group for those discussing organisational AI Policies with a number of regular meetups. Launch. WhatsApp.
The Lighter Side
Fine tuning AI to be lazy. Link.
A reminder that there is no clear dividing line between Parody & Deepfakes, courtesy of McKinsey. Link.
The Tokyo subway is mesmerising. Link.
Meanwhile AI is helping John Lennon to show up at a station in Liverpool. Link.
