Minotaur Quarterly

We know we talk about AI a lot. But the pace of acceleration in the final weeks of 2025 surprised even us – enough that this quarter we're dedicating most of our commentary to what's changed, what we've built, and what it means for software stocks.

Two shifts have converged in the last few months to create what feels like a step-change in AI capability. Neither is purely about the models getting smarter (though they have). Both are about how we use them.

The first shift is skills. A large language model can only take so much text (known as the context window), and the more you give it, the worse it tends to perform. If you want it to know how to do a hundred different things, you could give it all those instructions upfront, but that degrades performance. Skills solve this by giving the model a short description of what it could learn upfront, and if it needs that capability, it asks for the full instructions. It keeps the model focused and working better.

When Anthropic launched their Skills system in October, they provided a standardised format for packaging and sharing these instructions. The Skills Marketplace, a collection of skills aggregated from open sources (not vetted, so read any before you use them), now has over 77,000 skills. Your agent doesn't know how to work with Excel? Give it a skill and now it does.

We've built our own library of skills that let the agent access external services and take actions on our behalf. A lot more tasks are now automatable, and because skills are shareable, the ecosystem compounds.

The second shift matters more. We're moving from chatting with AI to running AI in loops.

Geoffrey Huntley, previously Canva's AI Developer Productivity Tech Lead, popularised a deliberately unsophisticated approach under the name "Ralph Wiggum": define what "done" looks like in a specification file, run the agent in a loop, give each iteration a fresh context window but let it see what was built before. As Huntley put it when explaining the name: "It's kind of dumb, kind of lovable, and it never gives up." The contrast with complex multi-agent orchestration systems is striking: the dumb approach, through sheer iteration, often outperforms the clever one.

We experienced this over the Christmas break. Thomas built a personal AI assistant (and a Minotaur Assistant for the team) with skills for email, calendar, Slack, Xero, Confluence, web search, image generation, and our internal fund reporting. He also built automation scripts that run AI agents in loops against task lists, converting unstructured notes into structured plans and executing them overnight while he slept.

Now we can give the assistant tasks like "read this email from our lawyers and update our notes on ETF setup costs" or "why is Lenovo down?" (which triggers a skill that simultaneously searches news, X, and email before responding). It can read an investor's due diligence request, pull the relevant fund data, and generate a customised Excel report tailored to the specific request – a task that could previously take hours of manual work. Each new skill expands what's automatable without polluting unrelated tasks with irrelevant context.

This brings us to a debate we've been having internally. We suspect many investors are having it too.

One dynamic that interests us is where value accrues in this new stack.

If an AI assistant becomes the layer that sits between you and your services – email, calendar, documents, CRM, research, analytics – then the underlying UI becomes less important. The "winner" is whoever owns the permissions, the workflow graph, the audit trail, and the ability to safely take action across systems.

That has second-order implications. Some software becomes more defensible (if it is the control plane). Some becomes less defensible (if it's a thin UI over commoditised functionality). Value may shift toward platforms that can govern agentic workflows rather than those that just serve screens. We're already living this. We don't pay for software unless we can access it via an API. If it can't fit into our automated systems, we don't use it.

On January 12, Anthropic launched Claude Cowork, a desktop agent that manages files, connects to services like Notion and Asana, and executes multi-step tasks autonomously. It's essentially a consumer-friendly version of what we built ourselves over Christmas, though less powerful. Microsoft is integrating Claude into Copilot. OpenAI is pushing ChatGPT as the default interface. The race to own the orchestration layer is already underway.

In terms of what this means for our portfolio, we've reduced our software exposure since quarter end, including cutting Atlassian, despite our bullish commentary on the name and sector late last year. We sold because the goalposts are shifting fast on unit economics and defensibility of seat-based workflow software as agentic tooling spreads. The distribution of outcomes has widened materially, and at current prices we don't think the risk-reward is attractive.

We still believe in the AI supercycle. Our positions in Nvidia and the memory makers (discussed in our December monthly) reflect this. Compute and infrastructure demand remains strong.

We also initiated a position in Hut 8 during the quarter. On December 17, Hut 8 announced a 15-year, 245 MW data centre lease to Anthropic at their River Bend campus in Louisiana, backstopped by Google (AA+ credit), with a total contract value of US$7 billion. We bought the stock that day. We were able to move quickly both because of AI-assisted analysis and because the deal structure was familiar: in a prior role, Thomas was a large shareholder in NEXTDC and Asia Pacific Data Centre, the trust NEXTDC created to fund their land and buildings when capital was scarce. River Bend has a similar feel – contracted, investment-grade-backed infrastructure with visible cash flows. At mid-point build cost estimates and before capitalised interest, the project yields ~15% unlevered in year one, rising with 3% annual escalators over the lease term. We thought the market was underpricing the impact of this deal on the company.

In a world that compounds change, good investing isn't about being right once. It's about staying right as the facts evolve.

AI moves in dog years. The past twelve months feel like seven years of progress. OpenAI's Deep Research isn't even a year old. Skills launched three months ago. Opus 4.5 arrived two months ago. The Ralph Wiggum pattern went viral this month. Each felt like a step change at the time. They're stacking up. The big unknown for software stocks isn't where we are now. It's where we'll be in another twelve months if this pace continues.

We see this in our own work. Since starting Minotaur, AI has become increasingly central to our research process. With this latest wave, it's now thoroughly entrenched. We're building AI systems that are deliberately "opinionated", in that they encode our philosophy, our decision rules, our view of what matters. The problem with using off-the-shelf ChatGPT to research or summarise is that by the time you're looking at the output, you're trying to extract insight from something that's already been compressed. Information has been lost. The signal you needed might have been discarded as noise. Because we've infused our AI with how we actually think, we can apply that lens at the detail level, before the lossy compression, not after.

The progress we've made incorporating AI into the deeper parts of our research process makes us think it won't be long before we can generate full investment recommendations with AI. At that point, we'll be able to measure whether the human judgement of Thomas and Arms adds or detracts value beyond the systems we build.

Diogenes once said to a would-be philosopher about the importance of lived experience vs. textbook: "You do not choose painted figs, but real ones… and yet you pass over the true training and would apply yourself to written rules." In markets right now, there's no shortage of "painted figs", ie. investors opining loudly on AI's impact on software without actually using the tools. Our advantage is simple: we eat the real figs. We build with these systems, run them in loops, and rely on them both in research and operationally. That's why we're not guessing when a step-change matters – we actually feel it, which is why we can change our mind with discipline when the facts move.