Your Company Already Wrote Its Best AI Training Examples
The Dartmouth workshop, a 67 year wait for a computer that could talk, and the human examples that finally made one work
You've sat through an AI plan that promised results by the end of the quarter.
The slides looked clean, the demo worked on someone's laptop, and everyone in the room nodded along.
Then the quarter ended, and the answers coming out of the new tool still sounded nothing like your team.
You're in good company, because some of the smartest people of 1955 made a bigger promise.
In 1955, four scientists believed they could teach a computer to talk in one summer.
Their names were John McCarthy, Marvin Minsky, Nathaniel Rochester and Claude Shannon.
McCarthy taught at Dartmouth, Minsky was at Harvard, Rochester worked at International Business Machines, and Shannon was at Bell Telephone Laboratories.
On August 31, 1955, the four of them signed a short proposal.
The first sentence read: "We propose that a 2 month, 10 man study of artificial intelligence be carried out during the summer of 1956 at Dartmouth College in Hanover, New Hampshire."
That proposal is the first known use of the phrase artificial intelligence.
They asked the Rockefeller Foundation for 13,500 dollars to pay for the summer.
The first goal on their list was figuring out how to make machines use language.
They also wanted machines that could form ideas, solve problems that only people could solve, and improve themselves.
The four scientists felt sure about the timeline.
They wrote that "a significant advance can be made in one or more of these problems if a carefully selected group of scientists work on it together for a summer."
They figured the computers were ready, and all they needed was clever code.
In the proposal, they said the machines of 1955 already had enough capacity, and the hard part was writing programs smart enough to use that capacity.
The summer of 1956 did happen, and Dartmouth College still calls that project the birth of the field.
Decades later, the proposal went up on McCarthy's Stanford website, and a note from 1996 sat beside the plan for a final report.
The note said, "There was no report."
Teaching a computer to talk took 67 years, which is a long way past one summer.
On November 30, 2022, OpenAI introduced ChatGPT, and anyone could finally hold a real conversation with a computer.
OpenAI described ChatGPT as a model that "interacts in a conversational way."
What finally made ChatGPT work was people.
OpenAI's trainers wrote example conversations by hand, and each trainer played both sides, the person asking and the assistant answering.
Then the trainers looked at two or more answers from the model and ranked those answers from best to worst.
They repeated that ranking over several rounds.
Those trainers showed the machine what a good answer looks like, one example at a time.
The 1955 group had the right goal, and they put their hope in clever code.
The missing piece turned out to be careful examples, written by people who knew what a good answer sounds like.
Your company has been writing those examples for years.
They live in every quote that won a deal.
They live in every customer email that solved a problem on the first try.
They live in every answer your best person gives without looking anything up.
Nobody on your team thought of that work as training material.
Your people were just doing good work, one customer at a time.
Right now those examples are scattered across sent folders, old quote files and the memory of the one person everyone asks.
The quote that won the deal sits in a folder next to the drafts that lost.
The email that fixed the problem is buried in a long thread nobody wants to reread.
An AI tool pointed at that pile gives you an answer built on whichever file the tool found first.
Your team reads a few answers that sound wrong, and then they stop opening the tool.
So the first job is cleaning up the examples, and that is where Elution starts.
We call that first step discovery.
We work beside the people who run the work, find where the best examples live, and clean them into a form a model can read.
The clean examples go into a plain folder of files and instructions, and you own that folder.
Clean them up, and your AI starts from your best work.
A new hire can answer a pricing question the way your best salesperson would, and the reply reads like the quote that won the deal.
A customer with a problem gets the email that solved the same problem last spring, written in your voice.
Your best person gets fewer interruptions, because the answer she gives without looking anything up is finally written down where everyone can find it.
The slowest part of the whole job is writing the examples, and your team finished most of that work years ago.
Your people keep working the way they work today, and nobody has to write a training manual from scratch.
You own those files and everything we build on them.
When the models get better, your workflows get better with them, because your examples go with you to whichever model comes next.
We give businesses back to the people who built them.
Follow us for more AI history lessons.
Sources
- John McCarthy, Marvin L. Minsky, Nathaniel Rochester and Claude E. Shannon, "A Proposal for the Dartmouth Summer Research Project on Artificial Intelligence," August 31, 1955. Source of the opening sentence, the list of goals, the "significant advance" sentence, the point about machine capacity and writing programs, the 13,500 dollar Rockefeller Foundation budget and the 1996 note "There was no report." http://www-formal.stanford.edu/jmc/history/dartmouth/dartmouth.html
- The same proposal reprinted in AI Magazine, volume 27, number 4 (2006): https://doi.org/10.1609/aimag.v27i4.1904
- John McCarthy, note on the Dartmouth proposal, describing the proposal as, so far as he knew, the first use of the phrase artificial intelligence: http://jmc.stanford.edu/articles/dartmouth.html
- Dartmouth College, "Artificial Intelligence (AI) Coined at Dartmouth," describing the 1956 project as the birth of the field: https://home.dartmouth.edu/about/artificial-intelligence-ai-coined-dartmouth
- OpenAI, "Introducing ChatGPT," November 30, 2022. Source of "interacts in a conversational way," the human trainers who wrote conversations playing both the user and the assistant, and the ranking of model answers by quality over several rounds: https://openai.com/index/chatgpt/
Original: https://elutionlabs.substack.com/p/your-company-already-wrote-its-best