The OpenAI Dev day seems to have been a genuinely important milestone in the current wave of AI innovation and exploration – see the Verge’s coverage here – offering three major enhancements to previous functionality:

  • GPT-4 Turbo with a 128k context window, which allows for much richer prompts and has knowledge of the world up to April 2023;
  • An Assistants API, which allows developers to develop their GPT apps that can also call multiple models and do Retrieval Augmented Generation (RAG); plus,
  • Multi-modal capabilities combining text, visual and text-to-speech agents.


The demo that Sam Altman performed, creating a custom GPT agent for coaching startup founders using only natural language inputs, provided a fascinating preview of what looks like becoming a Cambrian explosion of small, single purpose agents. And to continue the evolutionary metaphor for a second, these announcements instantly killed off a number of startups that had created a small niche overcoming the lack of RAG and multi-modal capabilities in the previous version of the platform. OpenAI’s Andrej Karpathy described this as a “new (still a bit primordial) layer of abstraction in computing.”

AI coaching bots and other low-hanging fruit

Small specialised agents might be the mainstay of popular use cases for a while, and in the immediate aftermath of the OpenAI announcements, a multitude of role-based coaching GPTs emerged. In fact, coaching bots are not a bad area of marginal utility, because a lot of coaching is essentially dialogic, playing back what somebody says and offering other inputs or ways of thinking, or just advice that has worked for others. In executive education, you can’t throw a stone these days without hitting an HR-exec-turned-life-coach, and whilst coaching has a role to play in educating business leaders, I confess a degree of scepticism that this is more of a feel-good intervention than one that will take us from the old world of leaders to the new world of distributed leadership in digital organisations. So too-cheap-to-meter coaching bots could save companies some time and money (and hopefully encourage them to spend more money on crabby, demanding teachers like me who want you to learn the models and eat your vegetables). Of course, a truly great coach considers ‘how’ people say things, and in what order, and also what they don’t say. In fact, they probably also pick up a bunch of non-verbal clues and use their experience to assess the person’s written output and CV in the context of what they are hearing to build a more holistic picture, and these are very human skills that cannot be replicated by a bot.


popular AI meme – this version shared by Rosita



In this context, I would also recommend Matt Webb’s recent musing on what too-cheap-to-meter intelligence means for the future of ubiquitous computing. In education more broadly, Ethan Mollick continues to experiment, and recently made a GPT to help students research and write a paper, with results that were interesting enough to prompt him to use this in all his future lectures.

Diminishing returns from recycling internet ‘knowledge’?

This early phase of Generative AI has produced some incredible new abilities, but also raised some questions about the potentially diminishing returns on recycling online social content through AIs, which risks AI-generated content bouncing around between models and being treated as correct, compounded by a tendency to mistake popularity for insight. Google’s experiments with AI-enhanced search results are of potential concern, as covered recently by the Atlantic. To be fair, the page rank model had some measure of human feedback to validate results, at least before they started poisoning their own well with advertising results. As David Galbraith put it:

At the moment generative AI takes seconds to do what takes hours and hours for low value, unoriginal content … the stuff in the middle of the bell curve of quality. No amount of time given to it allows it to produce the stuff at the top end where money is made.

A lot of general intelligence can be built on scraping the public internet, but some of the highest value and most impactful use cases will be in business, where we have the potential to create intelligent management software to replace the dreadful, bloated bureaucratic systems that bedevil large organisations today. This will require far more specialised, accurate and often proprietary information than the outpourings of content farms and keyboard warriors.


Towards GPT for managing complex organisations

Even within the protected walls of the enterprise, it is likely that many of the same challenges and pitfalls will exist. Ivan Vendrov wrote recently about the potential downsides of data-driven decision making, and how this is sometimes used as a cop-out when people don’t want to do the work and make hard decisions. Another risk is that we build this emerging agent layer around already divided organisational structures, repeating a mistake that many digital pioneers have made in the last century – wrapping tech around very old structures and ways of working to create marginal improvements at the expense of real change and embracing new possibilities and potential. Or worse still, we allow enlightened leaders to tweak their algorithms to produce results that feel right… to them, as appears to have happened in this hilarious story about the always-objective, radically-transparent hedge fund founder Ray Dalio. At least for now, I guess we mostly want to use this emerging enterprise agent layer in the background to do the boring automation and repetitive water-carrying to create more space for human agency and innovation, before we become more ambitious. But we won’t be able to go beyond this and create truly connected intelligence, unless we do the organisational transformation prep needed to create a service-oriented, modular architecture with connected platforms and data. That is far harder today than creating a demo GPT using the amazing tech that companies like OpenAI are making available. Even the path to adoption of a slightly better Clippy in our stone age office apps seems hard, because of the lack of connected data and insights most organisations have to work with. Here’s a talk I gave from 2015 on this topic that barely scratched the surface of the organisational architecture changes required to embrace these new superpowers:


We have been urging orgs to do this work for a decade or more, but short-term thinking has so far driven the agenda. Let’s hope that the coming automation of low-value, routine tasks will afford leaders more time to implement the promise of the connected enterprise.