What our 2019 FAQ wouldn’t explain

In June 2019 Apres launched its platform for AI training data. The press kit had a release dated June 19, a company overview, an FAQ and our logos. The headline announced a platform “that Creates AI Training Data Using AI.” We also launched on Product Hunt that year. The message I drafted for people who followed our upcoming page ended: “We’re very small and the support of people like you means the world.”

The product was real and the pricing was simple. The FAQ said the platform was free up to 2,500 data points, and after that we charged only for the data we delivered: about $0.08 for a task labeled by a person, down to $0.001 for a data point the machine generated. It handled text tagging, categorization, image annotation, data entry, OCR and summaries.

Then there is this answer, under “How does the AI training work?”:

How it actually works is a secret mix of stacks and intelligence geared towards achieving the highest confidence data labels.

Apres FAQ, June 2019

And under the next question, “Is the AI training configurable?”: “We actually don’t make the training configurable at this stage.”

The best question in the FAQ

One question in that FAQ was better than our answer to it:

If your AI can accurately label my training data, why wouldn’t I just use your AI for my application?

Apres FAQ, June 2019

We answered that our models were built to produce probabilistic labels for training data, backed by human oversight, and weren’t meant to push high-confidence results into a customer’s product. That was true. It also meant we were asking customers to trust labels made by a process we wouldn’t describe, from a model they couldn’t configure, checked by people they never saw.

Buyers who train AI are exactly the people who know how much can go wrong inside a black box. We were selling them one.

What the notes were already saying

The odd part is that my own notes from earlier that year pointed the other way. The AI overview I kept in February 2019 described a system that would show its work. Experts would write labeling rules, the platform would apply them, and then it would “represent the machined results back to the experts,” showing “the coverage, accuracy, and conflict of the various rules” so the expert could fix them. One open question in it was which examples to show the people writing rules, to help them improve.

My January 2019 notes had a sketch of a model that notices it is confused (in the example, that a trolley doesn’t behave like a car) and would then, in the note’s words, “ping the human that it needs more training examples for that situation.”

Both ideas are about a machine explaining itself to a person who can correct it. The FAQ hid exactly that part.

The turn

My plan for January 2020 had five items. The first was “Explainability V1.”

A year after that, Apres announced a $1.5 million seed round as an explainable AI company. Tech.eu described a tool that translates model behavior “into a language everyone else can understand.” Our company overview that January opened with a tagline, “Own your decisions,” and described turning black-box models into “transparent, interpretable conversations.” My quote in the announcement was: “Companies should not have to choose between AI performance and transparency.”

In about a year and a half we went from “a secret mix of stacks and intelligence” to a company whose whole product was showing how a model reached its answer.

What I took from it

A customer who can’t see how a result was made has to trust you instead. Sometimes that is fine. When the customer is the one who will answer for the result, it isn’t, and they know it.

Excellent is designed around the opposite of that FAQ answer. In the workflow we have described, the receipt for a finished task records what changed, what proved it, what could not be checked, and who approved it. If I wrote that FAQ today, the answer to “How does it work?” would be a link to the evidence.