The AI product era becomes unavoidable

Something shifted in November 2022. Not in the research. Not in the benchmarks. In the public understanding of what this technology actually is.

ChatGPT launched on 30 November and within days it had a million users. Within weeks it had become a cultural reference point. People who had never read a paper on large language models, never thought about transformers or token prediction or RLHF, were now using the thing, forming opinions about it, arguing about it on the internet. A field that had largely spoken to itself, to researchers, to engineers, to a narrow slice of well-funded enterprise, had suddenly acquired a genuinely broad audience.

This is worth sitting with at the end of the year. Not because it changes everything we were already doing, but because the context around what we're doing has shifted permanently. The era in which AI was a specialist concern is over. The era in which it is simply part of how products work has begun.

The gap between understanding and using

What makes the ChatGPT moment interesting is not that the underlying capability appeared from nowhere. GPT-3 was released in 2020. The basic architecture has been in the research literature since 2017. What changed is that the interface closed the gap between what the system could do and what ordinary people could experience doing.

There is a long history of this pattern in technology. The capability exists before the product. The product exists before the cultural moment. The cultural moment exists before the social norm. We are somewhere between the second and third of those stages right now.

What that means practically: the questions product teams are now asking are different from what they were asking twelve months ago. Before, the honest question was often "do we actually need to include language models in this product, or are we being swept along by the hype?" That question has not fully gone away, but its character has changed. Now the more pressing question is "what kind of AI product are we building?" Not whether.

That is a meaningful shift. It changes the planning horizon. It changes what counts as defensible product work. It changes the conversation with customers, with collaborators, with people thinking about what to build.

What the year's research tells us

Stanford HAI's AI Index has been tracking the pace of model capability development for several years now. The consistent pattern it surfaces is that the field moves faster than most observers predict, and that the distance between research frontier and deployed capability is compressing. What used to take years to move from paper to usable system is now moving in months.

That compression is double-edged. It means the opportunity to build real products with serious AI capability is wider and more accessible than it has ever been. It also means that product teams betting only on first-mover advantage in raw capability are betting against the tide. The models themselves are not where durable advantage lives. The places durable advantage lives are the same ones that have always mattered: the quality of the understanding behind the product, the quality of the operating system around it, the depth of the relationship with the problem being solved.

This is one reason we have always thought about the model as infrastructure rather than as the product. Infrastructure changes underneath you. You build on it, you adapt to it, but you do not identify yourself with a particular version of it.

The question of what it means to build an operating system

Throughout this year, the phrase that has kept coming up internally when thinking about the direction of Orbit, and about the shape of what MSG is building more broadly, is "operating systems, not assistants."

It is worth being clear about what that distinction is trying to mark.

An assistant is a tool that responds to requests. It answers questions. It drafts things. It helps when invoked. The value is real, but it is narrow: it depends on the quality of what the user asks, and it leaves the burden of coordination, tracking, sequencing and decision entirely with the person holding it.

An operating system is different in character. It holds context. It understands what is in progress and what is pending. It knows which commitments exist and which have been met. It makes the right information available at the right moment, rather than waiting to be asked. It reduces the cognitive overhead of running a serious operation.

The difference between those two things is not a difference in model capability. It is a difference in product thinking. You can bolt a language model into a chat interface and call it an assistant. Building something that genuinely operates as a system for executing work requires a much clearer theory of what the work actually is.

This is what Orbit is trying to be, and why the work on Orion, the intelligence layer underneath it, matters so much. The model is one component. The memory, the context management, the sequencing of actions, the understanding of what state a piece of work is in: those are the harder problems, and they are not problems the model solves on its own.

Where the consumer side sits

The public conversation about AI this year has been dominated by generative output: text, images, code. That makes sense given the tools that have generated the most attention. But on the consumer side, the more interesting question is what happens when AI stops being a novelty and becomes a persistent coaching or accountability relationship.

CheekyGains, and Naira inside it, are trying to answer a narrower and more concrete version of that question. What does it mean to have a performance coach that is always present, that has a running picture of what you've done and what you've said you'll do, and that responds to how your actual behaviour unfolds rather than delivering generic advice?

The honest answer is that we do not yet know at scale. What we know is that the model layer makes something possible that was not possible eighteen months ago: a kind of personalisation that responds to real context rather than segments and templates. The challenge is building the product around that possibility in a way that feels like a relationship rather than a gimmick. That work is unfinished and will continue into 2023.

On the services side: TUXX and pattern finding

TUXX, as the services and custom systems arm, sits in a particularly useful observation position heading into next year. The work of building custom AI systems for specific operating contexts is, amongst other things, a process of finding out what actually works when you take the theoretical capability and put it inside a real workflow. Not every use case that sounds compelling in a product meeting survives contact with how a business actually runs.

The discipline of that process, being rigorous about what the system should do, what it should ask about, what it should never automate away, is the same discipline that applies inside Orbit and the rest of the portfolio. Pattern Up, as a sub-product under TUXX, is one attempt to formalise what we learn from that process into something reusable.

The research work at Benediction Lab is the other side of the same coin. Agents, memory systems, GUI control, the edges of what autonomous product development might look like: these are not separate from the product work. They are the longer-horizon version of the same questions Orbit and Orion are trying to answer in the near term.

The honest year-end position

At the end of 2022, the honest position is this: the field has moved into a new phase, and almost nothing about how serious teams should think about AI products has gotten easier. The opportunity is larger. The noise is louder. The distance between a genuine product and a thin wrapper on a model API has never been more important to maintain, and never been easier to obscure.

What keeps the work grounded is having a clear thesis. MSG's thesis is about capability: increasing the genuine capacity of people and teams to do the work that matters to them. Not about demonstrating that AI can do impressive things. Not about building the most technically sophisticated system for its own sake. About building systems that people actually need in order to operate at a higher level.

That thesis is the same at the end of the year as it was at the start. What has changed is that the window to build something serious has opened wider. The public is primed. The tooling has advanced. The conversation has shifted from whether to what.

The what is where the work is. It is where it has always been.

Sources

• OpenAI, ChatGPT launch announcement, November 2022, openai.com/index/chatgpt/
• Stanford HAI, AI Index Report, 2022, hai.stanford.edu/ai-index