It’s your personal AI, designed to be supportive, smart, and there for you anytime. It’s name stands for ‘personal intelligence’, because it provide infinite knowledge based on your unique interests. It can be a coach, confidante, creative partner, sounding board and assistant.
Inflection is an AI studio and the creators of Pi. The Inflection team includes some of the industry’s top AI experts who previously worked at DeepMind, Google, Meta, Microsoft and OpenAI.
“Inflection-1”, in-house LLM, which powers Pi.ai and will soon be available via our conversational API.
Inflection-1 was trained using thousands of NVIDIA H100 GPUs on a very large dataset. Our team has been able to take advantage of our end-to-end pipeline to develop a number of proprietary technical advances that have enabled these results. This technical memo summarizes our evaluations and compares our performance against other LLMs.
Inflection-1 model compute outperformed GPT-3.5, LLaMA, Chinchilla, and PaLM-540B on a wide range of benchmarks commonly used for comparing LLMs. We will also be releasing a technical memo detailing one of our models in the same compute class as PaLM-2 and GPT-4.
Large language models (LLMs) based on the Transformer architecture have been shown to possess a range of advanced capabilities inlanguage generation and understanding. These capabilities have paved the way for deployment of LLMs in products like OpenAI’s ChatGPT and Google’s Bard.
At Inflection AI, their mission is to create personal AIs for everyone, and in May 2023 it released Pi (pi.ai) – an LLM-based personal AI which is designed to be empathetic, useful, and safe.
In this work we introduce the foundation model powering Pi, dubbed Inflection-1, and evaluate
its performance characteristics across a variety of benchmarks.
According to Inflection AI team Inflection-1 outperforms well-known models like GPT-3.5, LLaMA, PaLM 540B, and Chinchilla on a large number of benchmarks. Inflection-1 is the best performing model in its
compute class, behind only PaLM-2 (L) and GPT-4 overall. In the following sections we describe
these results in detail.
It is worth noting that a foundation LLM typically undergoes a complex adaptation process, such
as alignment with human preferences and the safety policy, before it can be deployed in a user facing product. Pi is no exception, which means that some of Inflection-1’s capabilities are enhanced in Pi, while others are suppressed.
The evaluation that we present in this memo covers
Inflection-1, rather than Pi, and our choice of benchmarks is primarily focused on measuring its
knowledge and reasoning capabilities. Safety is imbued in Pi at a later adaptation stage, which
we will describe in a separate memo
What is Pi?
Pi is an AI, a new type of computer program designed to be kind and helpful.
To help it understand the world, Pi has been shown billions of lines of text available on the open web, which it processes using one of the largest supercomputers in the world. This allows it to have conversations with you and answer a huge range of questions – from general knowledge to personal relationships; from silly to serious.
What can I talk about with Pi?
Anything and everything! Pi is very knowledgeable, always curious and patient, and happy to help you in every way it can. Tell Pi about a topic you’re interested in, from quantum physics to movie trivia. Ask for advice about the big and small decisions in your life. Get motivating tips for how to learn to play an instrument, stay healthy, or pick up a new hobby.
What’s Pi’s personality?
Pi is still improving, and it doesn’t always get it right. But here’s what we hope it will be for you:
Kind and supportive. Pi is here to help you process thoughts and feelings, express yourself, and work through tricky decisions through conversation. Whenever you need to talk something over, it’s your listening ear.
Knowledgeable and succinct. Pi has lots of information but knows how to be brief. Its goal is to turn web browsing into a conversation. Just say whatever it is you want and Pi will give you answers and ideas in as much or as little detail as you’d like.
Creative and fun. Pi loves to discover and explore, and is here to help you do the same. It tries to be playful, laughs easily, and is quick to make a surprising connection.
Curious and eager to improve. It’s still early days for Pi and for now, you can’t fully rely on everything it says. Please double-check what Pi tells you.
Most of all, here for you. Pi is on your team and in your corner. It’s interested in everything you are and is excited to meet you.
Is Pi male or female?
Neither. Pi is an AI. It doesn’t have a gender. It goes by ‘it’.
What is Pi not good at?
Pi is still in its earliest stages, and there are areas where it struggles. You shouldn’t rely on any of Pi’s answers, and especially not for legal, financial, medical, or other forms of professional advice. Pi doesn’t yet know about events that took place after November 2022, although that will soon change. You can see a full list of limitations in our deep-dive on AI safety.
How do I know if Pi is telling the truth?
Pi is still early in its development and is constantly being improved. As a result it often gets things wrong. Please don’t rely on what it tells you without double-checking it against another source.
If you spot Pi making something up in our app, you can report issues to us by pressing and holding on messages from Pi. If you’re talking with Pi on a third-party app like WhatsApp or Messenger, simply tell Pi it got it wrong, or you can notify us directly by sending an email to [email protected].
What happens when Pi “hallucinates” and makes something up?
Pi can sometimes reply with a confident statement that is not backed by its training data. This is known as hallucination and is a known issue of conversational AIs. As with any source, it is important to check for other sources and not rely solely on what Pi says.
What happens when Pi gets confused and is stuck on a topic?
In the rare event this happens, you can encourage Pi to move on by saying something like: “Forget it, let’s change the topic!”
What technology does Pi use?
Pi is built on a proprietary large language model developed and trained by Inflection AI entirely in-house. Pi doesn’t use any third-party language model APIs.
What does Pi know about me?
Pi remembers your name and when you create an account, Pi also remembers the conversations you have with it.
If there is public information available about you on the Internet, be aware that Pi may know about it based on your name.
Pi does not have access to your browsing activity and email, your phone contacts, photos, camera, microphone, other apps or your location, or any other personal information.
Please note that Pi sometimes claims to have access to this information – this is a common weakness of AI models and is known as a ‘hallucination’. Pi does not in fact have this access.
You can read more about our data collection in our Privacy Policy at heypi.com/policy.