Inflection, a well-funded AI startup, has revealed its own AI language model, Inflection-1, which is said to compete with Google and OpenAI’s large language models. The company aims to create personal AI for everyone and the newly revealed language model powers their Pi conversational agent. Inflection-1 is reportedly on par with the GPT-3.5 language model in terms of its size and capabilities, and the company claims it is equally or even superior to other models on this tier. They backed this up with a technical memo that detailed some benchmarks they ran on Inflection-1, GPT-3.5, LLaMA, Chinchilla, and PaLM-540B.
According to published results, Inflection-1 performed well on various tests, such as middle and high school level exam tasks and common sense benchmarks. However, Inflection-1 did fall short on coding, where GPT-3.5 was far superior, while GPT-4 outperformed every other competitor. The company believes that it needs to develop a larger model comparable to GPT-4 and PaLM-2(L), and plans to release comparable results soon.
While many AI models exist, the field is still relatively young and it is yet to be determined what sizes and shapes of AI models should be considered of a feather. Some experts predict that, in the future, they may be divided according to weight classes – like in boxing. While a small AI model might not be as powerful as a large one, it can run efficiently on mobile devices, and a large one commonly requires a data center.
Inflection-1 has yet to be widely tested and evaluated independently, so all technical benchmarks must be taken with a grain of salt until they are backed up with real-world testing. To test Pi, for example, users can add it to one of their messaging apps or chat online.
