LLMs have the power to change the world of work.
For the first time, we’ve created a system that allows non-experts to communicate with computers using natural language.
Let’s explore that idea.
Behind the digital interfaces that power today’s world of work lies a complex array of machine languages – also known as programming languages.
(One could even argue that digital interfaces themselves form a language: a language of buttons, shapes, images, colors, and text.)
These machine languages are operated by a technically literate cadre – programmers, developers, coders – who instruct computer systems and their connected machinery to do things.
Non-technical folks are expected to bridge the gap: to communicate with the technically literate, who in turn communicate with the machine. These folks are the product managers and business owners – translators of business needs into digital build-outs. (Using a dialect of natural language all their own, you might say – the kind learned in B-school.)
Then along comes the Large Language Model.
A language model so vast, it can accommodate (or perhaps swallow) all languages – especially those that are digitized and cross-annotated. And there’s a rich reservoir of those: websites, web apps, documentation, and code repositories, all containing traces of humans talking to machines, and vice versa.
The Large Language Model or LLM, is prodded in natural language (largely English) to regurgitate everything it has swallowed in training.
The regurgitation is far from perfect, but is improving at breath-taking speed. Human annotation and synthetic data are fast filling the gaps in inference and output.
The technically illiterate can now directly plug in to the Matrix.
What’s the right stance for this scenario?
How should we collectively react to the current context?
The market seems to be pulling towards workflow automation – a front-and-center use case for the pre-LLM crop of machine intelligence.
The argument for automation basically treats the human as a cost center, to be snipped from the work tree*.
But not quite.
The human worker stays plugged in as a stop gap against poor LLM decisions. The awfully-labelled (and just awful) “human-in-the-loop”.
*(100% valid in a some (many) cases).
Why LLM-in-the-loop is a better lens (“morally” and naturally).
- Tools are things we make to serve human goals
- Humans are great at using tools to do amazing things
- Institutions, in general, and Corporations in particular are set up to serve human aims
- Helping humans work better is of greater service to humans than is replacing them
- Technology is deflationary and we risk shooting ourselves in the economic foot by bypassing human users for short-term gain.
- LLMs are probabilistic models built on top of old fashioned deep machine learning, only now pre-loaded with all the brilliance and follies of a small part of humanity.
- LLM use is expensive at scale. This simplest way to bring costs down is to invoke llm use restrictively, “in-the-loop” as it were.
What is LLM-in-the-loop
LLM-in-the-loop is an inversion of the HITL paradigm that says – keep user workflows fixed as they are and “inject” inference only at the point of need in the workflow.
Take, for example, the need to research a market and share the findings with a third party (this is essentially all of the consulting and market research industries).
This work is almost a 100 percent executed in PowerPoint, Excel and Word – which can almost be considered a “digital primitive” of sorts. These editors enable full document control and some amount of authorship tracing.
Step1: Starting the research
The starting point of most research is locating the right sources of information and extracting it to useable form – the bane of every analyst, for it eats into the time spent in real analysis-related work.
The HITL solution to this is – a one-shot LLM run that scans the web and extracts all the information it can find, perhaps references it and outputs an excel or table. The analyst must now trawl through the data, verifying sources and checking the data.
The LITL solution to this – a multi-turn, multistep process to get to the final answer and required output, that takes longer, is less automated, but gracefully traces the curves of the analyst’s existing workflow.
Here’s how it goes down in Simplesurface:
Coming soon!
