LLMs are caught in a groupthink rut. This startup is attempting to get them out.
When the researchers requested 25 totally different LLMs (together with fashions from the highest US companies in addition to many open-source fashions from China and elsewhere) 50 occasions every to put in writing a metaphor about time, a lot of the 1,250 responses had been a model of “Time is a river” or “Time is a weaver.”
(I requested a few of my colleagues the identical query and 6 individuals gave me six totally different solutions. My spotlight: “Time is a favourite sweatshirt, formed by a lifetime of damage.”)
Whenever you search for it, you see repetition all over the place, says Kieran Browne, cofounder and CTO at Springboards. “The best way that the majority chat interfaces are designed, it makes it really feel such as you’re having a private dialog,” he says. “I believe most individuals don’t actually notice the extent to which they’re getting the identical stuff as all people else.”
One other instance: If you happen to ask “What ought to I identify my band?” Most fashions will say one thing involving “glass,” “neon,” “velvet,” or “static,” says Browne.
After I tried it, ChatGPT spat out a listing of 56 band names. On the high was “Glass Harbor.” Skimming by, I discovered “Static Empire,” “Neon Hearts,” and “Velvet Echo.” I requested Gemini; it gave me 15 recommendations, together with “Static Horizon.”
Among the recommendations seemed fairly cool, although. ChatGPT’s “Couch Astronauts” caught my eye, so I googled it—and located {that a} band referred to as Couch Astronauts already exists.
(OpenAI says that coaching fashions to provide dependable and coherent solutions can cause them to converge round acquainted, high-probability responses and that pushing tougher for novelty can result in weaker or much less dependable responses. It additionally notes that the “Synthetic Hivemind” paper studied fashions from 2024 which have since been up to date.)
Artistic catapult
Springboards has developed a software backed by a collection of LLMs, together with ChatGPT and Claude, that artistic professionals in promoting or advertising and marketing can use to brainstorm concepts. The software permits you to drag round textual content produced by totally different fashions, choosing the bits that you just like and mixing them into one thing new—in idea. Springboards is pitching Flint in its place mannequin that customers of its software can choose when in search of extra selection.
