Inflection takes on OpenAI with its new Inflection-1 model

Training an AI model from scratch isn't cheap, but startup Inflection is ready to show its cards.

The well-funded startup behind the friendly conversational bot Pi just pulled back the curtain on Inflection-1, the in-house foundation model powering its assistant.

Holding its own against GPT-3.5

In terms of the compute power used to train it, Inflection-1 sits roughly in the same weight class as OpenAI's GPT-3.5.

Inflection released a technical memo comparing its new model against GPT-3.5, LLaMA, Chinchilla, and Google's PaLM-540B. The company claims Inflection-1 holds its own, and in some areas even outperforms, rival models in its class.

On middle- and high school-level exam subjects like biology, as well as everyday common-sense reasoning, the model scored impressively well.

Where it stumbles hard is coding. GPT-3.5 easily beats Inflection-1 in code tasks, and GPT-4 completely smokes the field. Inflection is not denying the gap, noting that an even larger model designed to take on GPT-4 and PaLM-2(L) is currently baking in the oven.

Why it matters

We think Inflection is playing a very deliberate, clever game here.

Instead of burning billions trying to out-code OpenAI on day one, Inflection tuned this model specifically for what its assistant Pi needs: empathy, relatable dialogue, and solid common sense. If you just want a personal companion to talk through your day, Pi is available right now on the web and messaging apps.

However, there are clear trade-offs. If your daily workflow depends on software development or heavy analytical reasoning, Inflection-1 simply won't cut it.

More importantly, these benchmark numbers come entirely from Inflection's internal testing. Until third-party researchers can independently poke around the model, the claims should be taken with a grain of salt. Competing at the GPT-3.5 tier is a solid milestone, but the real test begins when their heavyweight challenger finally steps into the ring against GPT-4.