We first wrote about AI in early 2022, when AI systems were largely being applied to classification and data science problems. Early LLMs weren't generally available, and, for most, the use of AI in finance was something of a dark art, limited to quantitative hedge funds with proprietary 'algos' spoken of in reverential hushed tones. 

With the launch of ChatGPT in November 2022, language models suddenly became available to everyone. It was a seminal moment - in the The Industrialisation of AI (March 2023), we wrote "Large language models are going to be everywhere, embedded in the software we use and the workflows we undertake on a daily basis. More fundamentally, they will change the way we interface with computers and information". But these models weren't immediately useful in our day to day work - they hallucinated information, couldn't perform simple arithmetic, and lacked up to date information (original ChatGPT 3.5's knowledge cut off was September 2021). They were also limited by how much information they could process in a single query - ChatGPT's context length was just 4,096 tokens (or about 3,000 words), and there was no means to upload files.

Two and half years on, there has been rapid progress across all of these dimensions. With each update, models have become more capable and more useful, and AI is now deeply integrated across all aspects of our business. This process has been incremental, and we thought it might be interesting to write about the journey and where we think things are going next. 

Hallucination rates have dramatically improved over the last couple of years

The traditional academic benchmarks have basically been solved, and new models are now pitted against the hardest problems that domain experts can devise (e.g. Frontier Math, Humanity's Last Exam, GPQA Diamond)

Context length has grown exponentially and can now encompass a huge amount of information

Investment Research

As we made the AI infrastructure theme our main focus across equity allocations early on, we spent time testing each iteration of LLM released by the frontier labs to keep up to date with their rate of progress. It was only with Anthropic's Claude 2 in July 2023 that we started to to see genuine utility in our investment workflow, primarily in the area of summarisation. One of our equity portfolios might comprise 50-60 stocks, so 200-240 quarterly earnings releases and management calls per year. Then there are transcripts from sell-side conferences, industry events and long-form podcast interviews, all of which offer important insights into a company's business. It's a lot for a portfolio manager to keep abreast of, and furthermore, with each new announcement, the PM composes some thoughts to circulate internally among the wider investment team. 

Our first LLM prompt for earnings analysis (July 2023)

An excerpt from a later iteration of our earnings prompt, written by AI, and with the use of XML tags to delineate different aspects of the task.

Fast forward to today, and LLMs have become an indispensable component of our research function. We have an ever growing library of prompts for different use cases, from equity to credit analysis, to initiations on new companies or deep dives into new themes. Our prompts are far longer, comprising many more steps to output regularised, structured formats. Models have become generally more intelligent and better at instruction following over time, as well as developing a number of important new abilities:

An example of o3 going on a side quest to search the internet for additional information and build a simple model to perform a sensitivity analysis on Amazon's tariff exposure - this all took place within its reasoning traces, and only the single data point of -100bps made it in to its final (~2000 word) summary)

The last six months of AI research and NVIDIA keynotes have been all about reasoning models and scaling 'test time compute'. This has led to much smarter models, as per the above, but also paved the way for longer horizon tasks. One example we can all try is the 'Deep Research' feature, which is available in various forms via Gemini, ChatGPT and Grok. For us, this has been incredibly valuable for getting up to speed with new investment themes - for example, we used it to get quickly up to speed with the European Defence investment theme that we added to our multi-asset strategies early this year (here is an example of a Deep Research report on EU Defence from OpenAI's model).

Another recent unlock has been in the area of portfolio analytics. We wouldn't have dreamed of using LLMs for this kind of quantitative analysis even a year ago, but reasoning models have world-class data science capabilities, and this is starting to replace a lot of our arduous work on time series data for multi-strategy portfolio construction, performance benchmarking and risk analytics that we would have previously built in Excel. 

In this example, we uploaded several years' worth of time series data to perform risk analysis on one of our funds

Gemini generated these risk metrics for periods in history helping us to quickly and accurately prepare detailed information on the performance of our funds vs. the benchmark and peers.  We can now systematically re-run this analysis with very little human intervention. This is an example of a task that would have taken a senior analyst or PM a lot of time to run each iteration.

What's Next?

We follow AI research closely, and and see a clear path for the jagged frontier of artificial intelligence to continue to expand outwards, coupled with steadily improving reliability enabling longer and longer horizon tasks.This will usher in the Age of Agents, which we wrote about in March.

The best area to watch this unfold is in software engineering. We at Green Ash are a small investment team - we have no software engineers or coding experience, and nor would most of our peers, unless they were much larger organisations. But this is no longer an obstacle, given LLMs like o3 now rank in the 99.7th percentile of competitive coders globally. We have started using AI to write custom code for workflow automation, drawing across disparate data sources and using LLMs to process the information according to structured templates. There are countless use cases for this, from position monitoring, to cash reconciliation and trade execution workflows. Outside of investment activity, we also envisage considerable productivity gains will be possible on the operational side of the business, as well as in our regulatory and compliance obligations. We are very early on this journey and expect many more APIs to become available over time (such as Bloomberg). As the number of available tools and APIs expands, we will be able to build much more sophisticated custom scaffolding using frameworks like model context protocal (MCP).

n8n is a workflow automation platform that allows us to combine business automation processes (expressed in code) with API calls to data sources and LLM outputs

In the early days of ChatGPT, many used the analogy of AI chatbots being like infinite interns. This has since been upgraded to Dario Amodei's 'millions of geniuses in a datacentre' or Sam Altman's 'intelligence too cheap to meter'. The rapidly advancing frontier is enormously empowering to boutique asset managers like Green Ash. We don't have infinite tasks for interns (we struggle to find any), but have a much larger requirement for polymathic digital analysts and data scientists, that can work shoulder to shoulder with us and contribute to high-value research and speed up our investment process. Small teams, augmented by AI, can now have the productive output of much larger organisations, without the institutional inertia imposed by layers of bureaucracy, entrenched business processes and legacy IT systems.