Astronort: Why now?
What is Astronort?
Astronort is a business situational awareness & decision intelligence platform. Businesses have long had ‘big data’ strategies to making use of their numerical & database information, but in addition to that they have piles of… words. People chat online, write emails & memos, release white papers, report to their board and more.
Historically, getting information out of all these words that is meaningful to computers has been difficult. Tacit information about the mood of the company and information bubbling under the surface has been inaccessible to BI tools.
LLMs can change all this, and Astronort is making it happen.
What technical challenges are there?
At a high-level the technical challenges we face are:
Processing unstructured content into plain-text facts.
Extracting the most important information from the vast number of facts, ignoring irrelevant noise.
Synthesising insights and trends from the information.
Forecasting possible future states based on the insights and trends.
Surfacing critical considerations, providing relevant context for senior leadership to make better, more impactful, and more timely strategic decisions.
What AI technologies are we using
This is a not a complete list, but notable AI technologies we use are:
Large Language Models (LLMs)
ChatGPT and GPT4 are the best known examples of these. Give it a prompt, and it will give a response. They are extremely talented wordsmiths, terrible mathematicians, chronic confabulators, and mediocre encyclopaedias. They are a groundbreaking technology that has been in development since 2017, but has come into their own in 2023.
Embeddings & Vector Databases
These enable the creation of search indexes organised by meaning rather than words. They can be used to find information about a topic, cluster information into related areas, and measure how closely related concepts are. Embeddings are a more established machine learning technology, but a useful companion to applying LLMs at scale.
OCR and text extraction
Tools for turning scanned documents into text, or getting text out of poorly structured information, are largely a commodity these days, such as the decades-old open source project, Tesseract, or AWS Textract.
What can LLMs do? Where do they get stuck?
There’s a hint in the name, that LLMs are language models, not knowledge models. They have an extremely good grasp of language. They also have an unprecedented amount of common sense for a machine.
They can tirelessly read piles of text, and perform tasks such as reformatting, summarising, answering questions or extracting important points.
But they are not knowledge models. LLMs are the guy in the bar who has something convincing to say about everything, regardless of how much they know. They are prone to “confabulation”, where falsehoods are confidently stated as facts.
That said, given the necessary knowledge and direction, they can perform impressive tasks.
So, why now?
Paraphrasing James Carville: it’s the LLMs, stupid! The power of modern LLMs make it much more feasible to reliably extract the full meaning from unstructured content. Although narrow tools such as sentiment analysis have been available for some time, LLMs provide machines with reading comprehension & common sense.
LLMs are also very easy to adapt. “Prompt engineering” — carefully constructing questions (or chains of questions) to direct an LLM’s output — is orders of magnitude easier than training custom models for particular prompts. This is well suited to tailoring Astronort’s analytical work.
The emergence of powerful LLMs makes Astronort possible, and is the catalyst for embarking on this project now.
Why you? Can’t anyone put their documents into ChatGPT?
Sure, but as we said, LLMs are not knowledge models. Simply asking ChatGPT to analyse the document gives results that are at best milquetoast, and at worst confident and completely baseless.
To build Astronort, we are codifying the C-suite experience and knowledge of our cofounders (as long time COO and CP&CO’s) into the system that surrounds the core AI technologies, to draw the most valuable insights out of the source material.
Our CTO has experience as CTO, CEO, and advisor in technology companies, giving him a deep understanding of how technology & business decision-making intersect.
Our team are uniquely suited to build Astronort into a valuable tool for CEOs.
