[AI Agent Diaries] The One Where China Stood Up
Stargate, Operator, Billionaire spats, DeepSeek and Product Research
I started out AgentsDecoded thinking that in-between opinions, reviews of frameworks and pieces exploring the essence of agency I will include a weekly news column. After three weeks of doing that though I realised two things:
1. There is so much news every week I never feel I am actually doing a proper job of it.
2. There are already enough other newsletters gathering announcements and news, the world does not need another one.
What I think I will do instead that is hopefully more useful to you (and me), is use the weekly column cadence as a way to reflect on the week that has gone by. This will be a very specific point of view of a startup co-founder and CPTO down in the trenches working with people to implement AI solutions in the real world, far away from the dizzy heights of the leading AI labs and research centres with billions in investment.
Here we go.
Looking back at my week, we started off with some user research at work to gather data for how we should shape our product roadmap. It is always so powerful to get unbiased views of how people are looking to use AI Agents and where they are today. Two interesting perspectives are emerging.
1. There is a lot of opportunity in just helping companies upgrade from old Conversational AI tech . There are many mid-sized organisations that have built chatbots using the technologies post 2023, with standard ML intent classifiers and what is now essentially obsolete Conversational AI capability. All these organisations are now trying to figure out how to retool and move to technology stacks with GenAI but it is a slow process (certainly much slower than the rate at which LLMs are progressing). Definitely an opportunity for agencies out there to help people move up.
2. There is an army of consultant AI Agents coming. The vertical AI Agent space is ripe with opportunities. People with specific domain expertise (e.g. a specific niche of contract law) can develop solutions that combine their expertise with LLM capability - effectively turning what they know as an experienced consultant into a product. We will see a flurry of startups that are led by human consultants developing AI Agent Consultants and selling those (together with their standard services).
In the wider world of AI Agents three news stories dominated the agenda.
Let's start with the Stargate Project. Five hundred billion (!) in AI infrastructure investment in the US, confusingly announced from the White House but actually funded entirely from private organisations, of which apparently only a small part is currently available if we are to believe what Elon Musk said, that is not related with the project and is annoyed because his enemy Sam Altman is involved. Confused? Yeah, me too. What it does tell us is that nations are getting competitive. They want to lead the AI race, and win the AI wars - whatever that is supposed to mean.
The second story, which provides an interesting counterpoint to the first one is that DeepSeek, a Chinese lab that was previously mostly unknown to the western world and media, is crushing it at that very AI race. Releasing models that are cheaper, faster, better and open source. Sure they are standing on the shoulders of work done by US-based labs - but they are introducing innovations that are entirely their own. Europe and the US likes to tell itself that China just copies stuff, but that is clearly not the case. Among various geopolitical implications it also means that prices for access to LLM functionality are going to quickly start going down all around - which creates an interesting and challenging dynamic for the much more expensive to run US and European labs.
Finally, OpenAI released Operator. It's first AI agent. Essentially a more complete version of Computer Use similar to what Anthropic released last year. I saw two types of reaction online. "WOW!" and "meh". Ignoring all the hype coming from the increasingly large army of “AI influencers” there are good reasons for both excitement and reserve. As I mention here, while it is impressive stuff the uses cases and user experience are not quite there yet.
Much more has happened over just the past few days (e.g. Anthropic released Citations) but I am already a bit over 700 words and that is the cap I set myself! Now, off for the weekend, for hopefully a couple of days where AI is not dominating what I am doing!