One of the lesser known, but no less impressive, LLMs is called Inflection and it powers Pi, a personal AI app available across a wide range of platforms. Before Thursday’s release of Inflection-2.5, Inflection AI was focused on making its LLM empathetic but the new update also makes it smart, almost as smart as GPT-4 and it has only been trained on 40% of the amount of compute as GPT-4.
Inflection AI said that in this update it has made strides in areas like coding and mathematics that have led it to do better in industry benchmarks. The update also gives Pi the ability to do real-time web searches so it has the latest information available to it.
The company said that the update has already been rolled out to users and this has led to improvements in user sentiment, engagement, and retention. For anyone who hasn’t tried it, it’s a lot like ChatGPT and even supports voice calls; anyone who doesn’t want to use the app can also chat to the AI via WhatsApp, Instagram, Messenger, and Telegram.
Inflection-2.5 is slightly behind GPT-4 in the MMLU, GPQA, BIG-Bench-Hard, Mathematics, Coding, and Common Sense benchmarks but only by a little bit while making significant strides over Inflection-1. You can see the results below:
Despite the IQ improvements in this update, Inflection AI said that its LLM is still high in EQ making it friendly and approachable. The company also said that it incorporates extraordinary safety standards.
The release of Inflection-2.5 comes hot on the heels of the release of the Claude 3 family of models, Haiku, Sonnet, and Opus. Those models are also on par with, or better than, OpenAI’s GPT-4. Google’s latest Gemini models also offer a similar sort of performance.
Meta is allegedly preparing its Llama 3 LLM for release around July and it too will probably be on par with GPT-4. With most of the competition catching up to GPT-4, OpenAI could feel pressured to release an update to its model; GPT-4 was announced on March 14, 2023.
Source: Inflection AI