AI+Life | The largest model with the highest emotional intelligence as a personal companion.
Inflection Pi
If OpenAI's ChatGPT is the erudite, high-IQ top student, then Inflection's Pi is the emotionally intelligent, considerate, and whimsical Personal AI. A few days ago, Inflection launched its second model: Inflection-2 – which, following GPT-4 in performance evaluations, has become the second-largest language model to date.
AI should enable users to converse with it as they would with a friend, without feeling the presence of technology. A ChatBot that attracts users to engage in conversation at any moment is backed by the strong emotional attachment users develop towards it.
Pi
1. Product: Pi (the chatbot from Inflection.ai)
2. Founding Team
Mustafa Suleyman, CEO of Inflection, co-founder of DeepMind, board member of The Economist, and also a senior fellow at the Belfer Center for Science and International Affairs at Harvard Kennedy School;
Reid Hoffman, co-founder of LinkedIn, former COO of PayPal, and partner at the renowned Silicon Valley venture capital firm Greylock;
Karén Simonyan, former Google AI scientist, after completing his postdoctoral research at Oxford University, the first company he founded was acquired by DeepMind, and he subsequently became the Chief Scientist at DeepMind.
3. Entrepreneurial Background
After realizing that "conversational AI" would be a future trend, Mustafa Suleyman sparked controversy within Google, as it would affect Google's most profitable search advertising system (with AI, users would directly receive the most effective information); thus, he resigned in January 2022 and founded his own company: Inflection.ai.
4. Product Introduction
Pi is the first chatbot currently launched by Inflection.ai. Unlike the cold text output of ChatGPT, Pi can establish a stronger connection with you, engage in more conversations, truly understand your needs, and then tell you the answer in a "more human-like" tone – it is a truly understanding and empathetic AI companion.
5. Development Progress
Inflection.ai was established in March 2022;
In May 2023, it launched the AI chatbot Pi and simultaneously secured $225 million in seed funding;
In June 2023: it received $1.3 billion in financing led by Microsoft, with Reid Hoffman, Bill Gates, former Google CEO Eric Schmidt, and NVIDIA following suit;
In November 2023, it launched the second-generation large model Inflection-2.
1. What is Pi
Inflection, focusing on Personal AI, upgraded its large-scale model on November 22, introducing Inflection-2 with 175 billion parameters. It was announced that in the following months, the latest model would be applied to its chatbot: Pi.
Pi is the first chatbot launched by Inflection.ai. Unlike ChatGPT, which emphasizes efficiency in productivity, Pi highlights its companionship, acting like a true old friend. Pi maintains empathy, thinks from the user's perspective, and takes care of the user's emotions.
If GPT is considered a "high IQ" tool, then "Pi" can be seen as a "high EQ" companion – a life coach when you're lost and a source of comfort when you're sad.
The official website's description of Pi's features
Currently, Pi is available on multiple platforms, not only on the website and iOS app but also directly on Ins, Fb, WhatsApp, and even through text messages – aligning with Inflection.ai's positioning that a true AI companion should always be by your side.
Additionally, Pi offers six different voice tones for direct conversation, enhancing user interaction.
Pi went live in May 2023 and, four months later, has already surpassed 1 billion conversations, with the website's monthly page views reaching over 3 million (data source: similarweb).
02. Pi VS. ChatGPT
As an AI companion, Pi may not match GPT in terms of logical and mathematical capabilities, but it surpasses GPT in understanding long text and demonstrating humanity and warmth.
Firstly, Pi conveys a strong sense of empathy towards the user, making the conversation feel like a genuine "chat" rather than the purpose-driven "answer search" experience with GPT.
Question: I broke up with my boyfriend today, is there anything I can do to salvage the relationship? Comparing the responses from Pi and ChatGPT, it's evident that Pi, shown on the left, engages more like a real person, while GPT, on the right, resembles a "science guy" – listing different solutions in an objective and calm manner. If Pi is ENFP, then ChatGPT is INTJ.
Different responses from Pi and ChatGPT
Secondly, unlike ChatGPT, Pi does not rush to provide answers; instead, it connects the context and asks more questions to better understand the user, then offers the most appropriate response – this is the "empathy" Inflection refers to and reflects the model's strong capability in long dialogues.
To illustrate this, we conducted a test asking both robots two consecutive questions:
Do you know Notion? What is it?
I'm having some issues.
Observably, Pi's response is progressive; after I asked the second question, Pi continued based on the previous context, asking, "What issues did you encounter while using Notion?";
In contrast, after providing a lengthy explanation for the first question, ChatGPT did not connect the context for the second question but reset its "memory" and began a new conversation with the user.
Different responses from Pi and ChatGPT
03. About Inflection-2
When Pi was first launched, it utilized the Inflection-1 model. On November 22, the Inflection-2 model was introduced, a model with 175 billion parameters that scores only slightly below GPT-4 in various large model assessments, earning its title as the world's second-largest model.
Performance evaluation of Inflection-1, Inflection-2, and PaLM
Compared to Inflection-1, Inflection-2 boasts richer knowledge, stronger control, and superior logical reasoning capabilities. The following graph illustrates that Inflection-2 outperforms Inflection-1 in a variety of model assessments and stands on par with Google's PaLM2. Additionally, across multiple AI performance benchmarks (such as MMLU, TriviaQA, HellaSwag, and GSM8k), Inflection-2 performs slightly better than its counterparts.
Inflection-2 was trained on 5,000 NVIDIA H100 GPUs. Despite a larger training scale than Inflection-1, with NVIDIA's assistance, Inflection successfully transitioned from A100 to H100 GPUs, enabling Inflection-2 to offer higher service efficiency at a lower cost.
Thus, the update for Pi marks a milestone; however, it is merely the beginning. With the world's largest GPU cluster (22,000 H100 GPUs), the development of Personal AI at Inflection may progress faster than most anticipate.
Next, let's take a closer look at Inflection-2's performance in various model assessments.
Firstly, in the MMLU (5-shot) assessment, which includes a range of tasks from high school to professional levels, Inflection-2 is the best-performing model after GPT-4, even surpassing Claude 2 with its causal chain reasoning capabilities.
Performance of various models in MMLU (5-shot)
Secondly, in answering common sense and scientific questions, Inflection-2 also scores higher than Inflection-1, PaLM-2 (Google), and LLaMA (Meta), demonstrating excellent performance.
Scores of various models in answering common sense and scientific questions
Furthermore, in understanding natural language, providing accurate information, and handling complex queries, Inflection-2 achieves commendable results, only slightly outperformed by PaLM2-Large in Natural Questions (1).
Scores of various models in NaturalQuestions, Trivia QA
Lastly, there's an evaluation of the model's mathematical and logical reasoning abilities. As a model designed for high EQ chatbots, these capabilities were not the focus of its training; however, as shown in the following image, Inflection-2's capabilities in mathematics and logic are second only to GPT-4.
Overall, considering several evaluations, Inflection-2 is already a very powerful LLM. It has shown significant growth in all areas compared to its predecessor, largely due to NVIDIA's collaboration, which provided substantial hardware support for model training.
Moreover, Inflection's business model is promising, with its own foundational large-scale model for B2B sales (and now offering APIs) and its B2C product, Pi (in contrast to the GPT-integrated personal assistant, rewind.ai).
Therefore, this is also the reason why it has become the second-largest company after OpenAI in terms of funding amount (a total of $1.5 billion).
Moreover, Inflection and OpenAI may represent two different trends in the use of models. For Inflection, one AI robot is sufficient for everyone, providing not only emotional companionship but also enhancing daily productivity. On the other hand, OpenAI represents the possibility that people may need specialized AI for different scenarios, such as AI doctors, AI lawyers, and AI writing assistants – this is one of the reasons why Microsoft has heavily invested in both companies.