The Problem with Predictable LLMs
Large language models (LLMs) like ChatGPT and Claude have become ubiquitous in modern AI applications, from coding assistance to creative brainstorming. However, these models often suffer from a lack of creativity and variety in their responses. This predictability, while acceptable for structured tasks like coding or research, becomes a hindrance when users need innovative ideas or diverse suggestions.
For instance, ask any mainstream LLM for a random number between 1 and 10, and you'll likely get 7. Repeat the request, and the sequence of numbers becomes eerily predictable. This phenomenon, known as 'groupthink,' limits the models' ability to generate fresh and unexpected ideas.
Springboards and the Birth of Flint
Enter Springboards, an Australian startup aiming to break the groupthink rut with its LLM, Flint. Unlike traditional models, Flint is designed to produce a wider range of responses to open-ended questions, such as 'Where should I go in Europe?' While other LLMs might converge on similar answers, Flint strives to offer something different.
'Most language models are fighting hallucinations,' says Pip Bingemann, cofounder and CEO of Springboards. 'We welcome them.' This approach allows Flint to explore uncharted territory, even if it means sacrificing some coherence for the sake of creativity.
The Random Number Game
Bingemann demonstrated Flint's capabilities with a simple experiment: asking for a random number between 1 and 10. While ChatGPT and Claude predictably returned 7, Flint offered unexpected values like 3.7916. This 'sales trick' highlights Flint's ability to break away from the predictable patterns of mainstream LLMs.
But the differences go beyond numbers. When asked to name a type of car, ChatGPT and Claude predictably suggested Toyota or Honda, while Flint came up with a Ford F-150. This variation underscores the 'lost information' that traditional models fail to surface due to their biased training.
The Creative Catapult
Springboards' tool, which integrates Flint alongside other LLMs, is designed for creative professionals in advertising and marketing. Users can drag and combine text generated by different models to brainstorm ideas. Flint serves as an alternative for those seeking more variety in their outputs.
Zoe Scaman, a business strategist, found Flint particularly useful for its ability to 'catapult' her thinking in different directions. In a test involving a finance company reinvention, Flint suggested rebranding the concept of wealth accumulation, while other models offered more conventional ideas.
The Challenges of Creativity
While Flint shows promise, it is still a prototype and faces challenges. Increasing the 'temperature'—a setting that controls the randomness of outputs—can make responses incoherent. Springboards addresses this by training Flint to selectively boost randomness at specific points in its output, such as when suggesting destinations in response to a travel query.
'Flint is programmed to throw an oddball in,' says Maximilian Weigl, cofounder of marketing firm Uncommon. 'It’s more of an invitation to think wider.' However, Weigl cautions against over-reliance on AI, emphasizing the importance of human creativity and judgment.
The Future of LLMs
Springboards' innovation with Flint highlights a broader issue in the AI community: the need for more diverse and creative outputs from LLMs. While models like ChatGPT and Claude excel in structured tasks, their predictability limits their usefulness in brainstorming and ideation.
Flint offers a glimpse into a future where LLMs can serve as true creative partners, generating fresh ideas and sparking innovation. However, achieving this vision requires balancing novelty and coherence, a challenge that Springboards and other startups are actively tackling.
Conclusion
Springboards' Flint model represents a significant step toward breaking the groupthink rut in LLMs. By embracing hallucinations and selectively boosting randomness, Flint offers users a wider range of creative outputs. While the model is still in its early stages, its potential to revolutionize how we use LLMs for brainstorming and innovation is undeniable. As AI continues to evolve, startups like Springboards will play a crucial role in pushing the boundaries of what LLMs can achieve.