Is Your AI Smarter Than an Intern?
Abroad is the notion that Artificial Intelligence (AI) is here to replace the human mind.
Nonsense. AI cannot substitute for a human mind. What it can do—especially in business applications—is substitute for the mind-numbing. The best, most practical business of AI is to operate on the level of one of your slower interns.
Consider One AI, an Israeli company, still private so, alas, out of reach for most investors. The One AI platform, called Studio, allows a non-expert user to create an AI “agent” and then endow it with a selection of some 40 individual skills, each operating independently of the others, but passing results down the line. If you think of creating a Dungeons and Dragons character and buying him a selection of powers, you won’t be far off.
Suppose your business has accumulated dozens of hours of recorded sales calls with potential customers. One “skill” will transcribe a call, and other will “outline” it, sorting the conversation into main headings. Another will extract next steps from a transcript and forward them to the right staff person for action. At the high-end, another skill extracts cues that indicate a customer’s inclination to buy, so salesmen can focus on the most promising leads.
In short, intern work.
The common denominator of such tasks is that humans learn them easily. Hand an intern a transcript of 10 customer service calls. Tell him to identify next steps and make sure the info goes to the right staff person. There’s a good chance he will get it right the first time. Hand him 10,000 transcripts and there’s a good chance he’ll quit.
The One AI Studio is powered not by a Large Language Model (LLM) such as the one made famous by ChatGPT, but by a small language model.
Have you ever had an intern who insists on making his work “meaningful” and even creative. The results can be worse than annoying-spewing inaccuracies.
LLMs create the same problems, but at a much greater scale. LLMs like ChatGPT turn out to be much worse than useless for most business applications. LLMs are trained on huge amounts of data, billions upon billions of instances, evaluated across tens of thousands of parameters. Their unnerving goal is to answer questions in ways similar to how such questions have been answered before: thus, all that data is used in training. This training, and the enabled “inference” runs, are what’s causing the world’s data centers to glow in the dark.
The goal of these models is not truth but plausibility, to sound convincing on any conceivable topic. Plausible errors are disasters because they are so darn… plausible.
Amit Ben, co-founder of One AI, explains the idea of a large language model was to “take the models and blow them up exponentially and push exponentially more information into their training, so they’re able to do much more, and to do many different things.”
The “amazing” result is that an LLM can do an extraordinary number of tasks for which it has not been explicitly trained and can generate responses “that might be perceived (falsely) as reflecting a deep understanding.” The downside, in addition to the staggering costs, is that these models are black boxes “notoriously hard to control, predict, or verify” The user’s “ability to wrangle it to do something specific is diminished” because we don’t really know how it works in the first place.
Consider you are not-the-sharpest-tool-in-the-shed intern; would you complicate the young man’s training by insisting that before he digs into those transcripts he also learns to write poetry or draw his pet goldfish in the manner of Van Gogh, tasks LLMs do happily? What if, to record data from sales calls, your intern first had to master iambic pentameter? What if it were impossible for an employee to proofread a document without first mastering Van Gogh’s brush strokes?
No one would ever hire an intern again. Just as no enterprise will replace an intern with a large language model inclined to flights of fancy.
Emerging is a simple principle:
The training of an AI should not be broader than the training needed by a human to perform the same task.
Because a small language model is trained at human scale, it is not only more effective but vastly cheaper to build and use than an LLM.
It is just not very bright. Go for very bright, try to make AI into a mind, and that way madness lies. AI can’t do mind, but it excels at the mind-numbing.