Microsoft + OpenAI = ?

As debates swirl around Open AI and its potential impact on Microsoft's strategy, it’s important to take a step back.

Large language models (LLMs) can already, and certainly could, do many things that can benefit Microsoft. But innovation’s strategy sometimes falls into a subtle trap: using the wrong market segmentation. In other words: is today's market segmentation (the boxes in which we put products and services - such as chat, media generation, or even search) the right one to help us think about the future?

There is reason to believe it isn't.

Quietly and over the years, Microsoft has built a very large portfolio that could be synergistic, not just complementary, with OpenAI's generative AI, and could do so in a way that transcends current market definitions.

To explore what the future might have in store, we will classify the possibilities in the classic three "horizons" but also use one, uncommon, lens: the AI-enabled, augmentation of collective intelligence which leads to the creation of so-called Superminds (1) where the emergent intelligence from networks of people and intelligent machines generates disproportionate value for the individuals, companies, and ecosystems that harness it. Microsoft included.

Below are some ideas but the new framing is possibly more important than the ideas themselves, and once you have understood it, you will be able to fill in the blanks with more and better ideas.

Horizon 1 - today’s tech, today’s use cases

The web is awash with relevant examples, as most people are thinking about this, so let's keep this part to a minimum: helping people write and create content, summarize it, and embed it into Microsoft Office 365 (Word, PowerPoint, Outlook) or Teams. Generative AI can help remove writer’s blocks and TL;DR things. It can do a better job of coming up with Clip Art (Microsoft recently announced its Designer product, for instance). All this could help marketing teams come up with better copy for their content, for instance. OpenAI could make Microsoft search engine evolve, which would be a good competitive threat to Google's dominance - and users will benefit.

All very interesting - but yesterday's "aha" moment. If we stop here, we would be missing the forest for the tree. Strategy is a tool for seeing that forest, so we will take a brief digression before going back to predicting future possibilities.

A new lens for strategy

A modern AI-first strategy requires us to step away from thinking of OpenAI as a chatbot or image generation software, or as a souped-up search or knowledge manager. A better market segmentation could be based on this question: if monetizable value is generated not by a single person, but rather by an organization, or an ecosystem, how could the synergy between Microsoft and OpenAI unlock it? How would AI support it?

A new type of "why?" question

Let's start with the "Why" - the job to be done, writ large. A supermind - any type of it, from an enterprise to an industrial ecosystem, from an individual brain to a flock of birds, does 6 things.

1. Senses the environment.
2. Remembers past conditions that resemble today’s and pattern them.
3. Recombines the knowledge to help create new ideas.
4. Brings information, knowledge, and people’s opinions and emotions together to help decision-making.
5. Acts, in order to maintain its ideal state, and to acquire new knowledge
6. And last, helps the entire system learn from those loops - from the inception of the idea to the execution, and the results.

(Repeat - and note that this is how our human brain, among others, does it.)

A new "what?" question

The four modules listed below help structure the organization of collective intelligence, and augment people's networks with machine’s – and vice versa.

Each of the modules is a lever to steer the system’s dynamics towards higher collective cognitive performance. They allow us to take a more active management role in that complex system.

And then AI's "how"

Now for the "how": what can large language models do for these collective cognitive processes? Generally, AI can augment superminds in four ways - i.e., AI’s “four C’s”:

1. Connect?entities (people, and machines – by, for example, helping pinpoint the right nodes in the network and making them discoverable to each other through search);
2. Curate?knowledge (for instance, semantics-based searches and computer vision that identify the most relevant content, and cluster it for people to process it more easily);
3. Collaborate?across those entities now enhanced by the new knowledge (for instance, natural language processing that automatically summarizes and structures content to make it easier for humans to process);
4. ...and?Compute?any other prediction (for instance, to determine which participant in the network is worth rewarding, or other machine learning algorithms to detect spam and fakes).

LLP and generative AI will power some of these in new ways.

Looking at Microsoft + OpenAI with a new lens

In the case of Microsoft, the analysis using these lenses is summarized below and could inform the strategy of Microsoft, its developer and service-provider ecosystem, and its clients.

To understand the possible, and real, value unlock, remember that OpenAI’s current main problem is a big one and not one that is easy to solve by just throwing more parameters at a big model: the lack of reliable, consistent veracity and logic of its answers. But even before that problem is fixed, extraordinary value can be generated, as some of the use cases may not need perfect reliability in that area as long as humans - and their networks - are in the loop. That's an important design principle we can explore: we don't need currently-unfeasible artificial intelligence perfection if we can harness collective (human+AI) intelligence.

Now, with this frame of reference, let's go back to the horizons.

Horizon 2 - within a year or two

Now let's change either of the two horizon parameters, e.g., existing use cases but with better tech, or new use cases with existing tech. Think about the implications for Microsoft's portfolio.

Microsoft Teams

Collaboration that happens on Teams means something to an LLM. One could summarize threads and entire channels, giving overwhelmed executives a concise pulse of what happens. Generative AI could support collaborative writing: co-editing documents now can be done not only with multiple people on a document but also with machines and people. It could help with strategy formation and deeper creativity, in groups: LLMs can do concept matrices (2X2 anyone?), for instance, and they can be prompted to help structure problems, fast. Those can be injected into networks of people, so humans do things with them.

Microsoft Viva

Viva Topics: knowledge graphs within enterprises are increasingly discoverable through MSFT Viva topics, with the promise of changing how knowledge management is done.

Viva Connections: employee portals of yore can now become not just a repository of threads, but a natural-language-queryable space. Where you could interact with some (very creative) digital twin of the company, capturing its zeitgeist and conversation.

Viva Engage: LLMs are really good at detecting intent and tone in conversations, and making interactions engaging. Today, the experience for employees on employee-engagement tools doesn't use much AI. All of that might change.

Viva Learning: generative AI can help learners engage with content more naturally, and learn by practicing questions and answers - especially for skills where "one exactly right" answer often doesn't exist (think: executive presence, personal effectiveness).

Viva Insights: the entire social graph side of the story is conspicuously absent from the current discourse on generative AI. That is a massive miss because that is half of a knowledge graph - and that gap may be plugged by Microsoft and OpenAI.

Viva Sales: just launched, it provides salespeople with a more insightful and frictionless experience when planning and executing client interactions. LLMs can be of great help here because they can detect tone and intent, and recommend content and ideas - improving sales productivity. This works at an individual level, but also at an aggregate one. Just take the following example: the dreaded and virtually indispensable Outlook application.

Outlook

Interaction between people is a signal that language models, especially when combined with Viva Insights network analysis, could mine. Think about understanding the tone, intent, and structure of the kind of conversations your sales force is having (and not having) with clients, guide them in near real-time to provide the best ideas, and reduce the variance between the best and the worst salespeople. And pre-sale. And solution architects, etc, etc.

Sharepoint

The Viva Topics and Office 365 points largely apply here. It is an opportunity for knowledge management to embrace some non-linear innovation lift.

Bing

LLMs in isolation are not yet very effective as search engines, and they're also very costly compared to today's search. And yet...Knowledge graphs is what we are after - or rather, the content side of knowledge graphs: Bing search is relatively unencumbered by the advertising revenue dependence that Google has, and could try to build a completely diffrent search experience. LLMs understand language well and can improve the quality of search results. And separately, search can be associated with generative AI to provide immediate solutions to its accuracy challenge (see Elicit.org for instance, or even what Google started doing with its search snippets).

Linkedin

The people (and people's network, both internal and external) side of knowledge graphs: Linkedin knows what we, individually and collectively (companies) know: our skills and our network structures. Among other things, that signal could be used for capturing the zeitgeist of trends, the discourse within companies and industries, and even for assessing the credibility of social media claims, or fighting social media bubbles.

PowerBI

LLMs can offer expressions, equations, and other logical structures that can make PowerBI's configuration easier.

GitHub

Many developers use GitHub as their code knowledge hub. LLMs can be trained with code, and GPT-3 has, resulting in a remarkable improvement in its logical abilities, despite lacking symbolic reasoning. What can this do to improve OpenAI's Codex? Could individual companies fine-tune their models based on their own GitHub repositories, to get to better - and more strategically differentiated and defensible - coding?

Azure

Generative AI will drive demand for cloud computing services - such as out-of-the-box models that can be customized and deployments in enterprise-client environments. Microsoft's Azure will be there to serve enterprise clients through new contracts, but also help them think about security, privacy, ethics, etc. Satya Nadella released a public statement emphasizing how much of OpenAI's power (GPT 3.5, Codex, and Dall-e) was now available as part of Azure OpenAI services. Azure could run the machine-learning backbone, both as raw processing-power infrastructure, as well as by providing pre-trained models for specific uses (say, sustainable procurement, weather forecast, and related supply chain optimization, employee attrition models) that clients build on. This is possibly the most important vector for direct monetization that Microsoft has.

Minecraft

LLMs can already help users build custom commands, for instance. But as we will see in the next section, this is just the beginning of what can be done in gaming.

Microsoft ERP (Dynamics 365)

What can you do when you help people predict things in workflows? For instance, summarizing an underwriting case for a financial institution (say commercial lending or property insurance), so that their employee takes a call on it?

Horizon 3 - change both: new tech, and new use cases. By 2030

These are speculations and long shots, but they're worth taking.

LLM can fold proteins by reading their latent structures as if they were language, so why wouldn’t they read the latent structure of career paths (from LinkedIn) and give people more creative options, as a career coach? Using the same capabilities, can generative AI identify the sources of the competitiveness of companies, by looking at their people’s skills, and the networks between them?

LLM models can be fine-tuned. For instance, MedPalm recently showed promise in the medical space. What could Microsoft and its clients do with fine-tuning? That may be a path toward increased and more consistent veracity in specific domains and industries.

Now think AR, which today is incarnated in the fledging Microsoft HoloLens. 2023 will be the year in which augmented reality starts becoming real, especially in a business-to-business context. The feed of information overlaid on the physical reality, and how users engage with it, can benefit from LLMs. The current limitation with augmented reality (think google Maps Live) is that people can only absorb so much visual input through a small screen. But humans are very good at using visual and audio input at the same time, bidirectionally - that is, by querying their environment with their eyes, their ears, and their mouth. (For more on this, if you’re so inclined, I suggest digging into neuroscientist’s Karl Friston concept of active inference, and see what it makes you think of.)

But also think gaming - Xbox. Generative AI can enrich or even reinvent the gaming experience with both natural language creations and imagery - customizing the player experience, infinitely.

More generally, generative AI can add visual layers on top of reality – augmenting it with additional visual information and creativity, for instance. That could also be beneficial in gaming, as well as in serious applications like healthcare, construction, and defense.

And things like Minecraft's user-generated videos have been used to train AI to learn about the world. What happens if Microsoft deploys LLMs natively on Minecraft, and uses the "natural experiments", the crowdsourced myriad activities that players do, and their sequence, to detect latent patterns akin to language structures that yielded the likes of ChatGPT?

And, just possibly, there is that GitHub connection. LLMs think semantically, not symbolically, and that's a limitation on the path of artificial general intelligence. However, training GPT-3 on code has improved GPT's logical structuring of problems and answers. Is there something more there, using some part of the largest code repository on earth, that could take AI to the next level?

LLMs could help build a “superintelligence” based on all of those language, imagery, and network signals that people, and the world's instrumentation, provide to them. Microsoft could use some of the data, and certainly the metadata to train its models, resulting in further improvement. This would add value back to OpenAI. Perhaps even start a flywheel, where Microsoft's touchpoints into the real world embody OpenAI's algorithms, and make it smarter thanks to machine learning registering the latent patterns of trillions of feedback loops.

Make no mistake - I am not sure (and I am incompetent in assessing if) that could result in general artificial intelligence, but all of this could generate a significant “augmented collective intelligence” allowing the emergence of the collective intelligence of networks of people and intelligent machines. Which is a superintelligence of sorts, where machines can think semantics at scale, and humans - as well as the world's instrumentation, such as IoT - add the symbolic abstractions and the embodiment (meaning, the sensorial layer) to the collective "brain". A supermind.

Where can this go?

Time will tell what Microsoft and OpenAI can do together.

But more generally, what if we systematically used the above Augmented Collective Intelligence Why, What, and How lenses to inform our AI technology strategy? What would we envision? What gaps could we fill? How would that inform our product and service design? How would that make us think about the risks and their mitigation?

Let’s not miss this ride: it might build some of tomorrow’s intelligence, that we badly need to solve tomorrow’s problems.

(1) For more on Superminds, refer to MIT's Center for Collective Intelligence and the work of Prof. Thomas Malone and his team.

More on how to design superminds at Supermind [dot] design, in particular the 2030 Positive Futures scenarios where some of the above ideas are already articulated. And why not, some of those could just be 2025 stories.