Original source: Deep Reflection SenseAI
![]()
Image source: Generated by Unbounded AI
If OpenAI’s ChatGPT is a knowledgeable and highly intelligent scholar, Inflection’s Pi is a thoughtful and playful Personal AI with full emotional intelligence. A few days ago, Inflection launched its second model, Inflection-2, which is currently the second largest language model after GPT-4.
With a team of 50 people and the world’s largest GPU cluster: 22,000 H100s and a total funding of $1.5 billion, Inflection, founded in March last year, has brought warmth to large language models and wishes one-year-old ChatGPT a happy birthday.
AI Native Product Analysis
Pi
![]()
**1. Product: **Pi (Inflection.ai conversational bot)
2. Founding Team
- Mustafa Suleyman, CEO of Inflection, co-founder of DeepMind, member of the board of directors of The Economist, and 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, partner of Greylock, a well-known venture capital institution in Silicon Valley;
Karén Simonyan, a former Google AI scientist who founded his first company after completing his postdoctoral research at the University of Oxford, was acquired by DeepMind and subsequently became DeepMind’s Chief Scientist.
3. Entrepreneurial Background
Mustafa Suleyman quit his job in January 2022 to start his own company, Inflection.ai, after realizing that “conversational AI” would be the future, sparked controversy within Google because it would affect Google’s most profitable search advertising system (because AI gives users the most effective information directly).
4. Product Introduction
Pi is the first chatbot launched by Inflection.ai at present, compared to ChatGPT’s cold text output, Pi can generate stronger links with you, generate more conversations with you, truly understand your needs, and then tell you the answer in a “more human-like” tone - is an AI companion who really understands you and has empathy.
5. Development process
- Inflection.ai Founded in March 2022;
- In May 2023, the AI conversationbot Pi was launched, which received a $225 million seed round of funding in the same period.
- June 2023: $1.3 billion in funding led by Microsoft, followed by Reid Hoffman, Bill Gates, former Google CEO Eric Schmidt, and Nvidia;
- In November 2023, the second-generation large model Inflection-2 will be launched.
01. What is Pi
Inflection, which focuses on Personal AI, upgraded its large model on November 22 to launch Inflection-2 with 175 billion parameters, and announced that it will apply the latest model to its own chatbot: Pi in the next few months.
Pi is the first chatbot launched by Inflection.ai, and unlike ChatGPT, Pi does not emphasize how efficient it can help people improve in terms of productivity, but highlights its companionship attributes, just like a real old friend, Pi will always maintain empathy, think from the user’s point of view, and take care of the user’s emotions.
If GPT is a “high IQ” tool, then “Pi” can be seen as a “high EQ” companion - it can be your Life Coach when you are confused, and it can give you some comfort when you are sad…
![]()
A description of the features of the Pi on the official website
At present, Pi provides the use of multiple terminals, not only websites and iOS apps, but also supports direct use in Ins, Fb, Whatapp, and even you can directly use your mobile phone number to have a text message conversation with them - this is very much in line with the positioning of Inflection.ai, after all, a real AI companion should be by your side at all times.
In addition, in addition to typing and communicating, the Pi also offers 6 different voices that allow users to have a direct conversation with them.
![]()
Pi went live in May 2023, and four months later, Pi has surpassed 1 billion conversations, and the website now has an average of more than 3 million monthly PVs. (Source: similarweb)
![]()
02. Pi VS ChatGPT
As an AI companion, Pi’s logical and mathematical abilities may not be as good as GPT’s, but its understanding of long texts, as well as the humanity and warmth it embodies, are far better than GPT’s.
First of all, Pi will express a strong sense of identification with the user, and the conversation state makes you feel like you’re really “chatting” with it, rather than the purposeful “answer search” that you would do with GPT.
Question: **I broke up with my boyfriend today, is there any way to get it back?**Comparing the answers of Pi and ChatGPT, it can be clearly seen that the Pi in the picture on the left is more like a real person talking to you, while GPT in the picture on the right is like a “science man” - objectively and calmly listing different solutions. If Pi is an ENFP, ChatGPT is an INTJ.
![]()
Pi vs. ChatGPT’s different answers
Second, unlike ChatGPT, Pi is not in a hurry to give an answer, but will ask you more questions in context, just to understand the user better, and then give the most appropriate answer - this is what Inflection calls “empathy”, and it also reflects the outstanding ability of Inflection models to have long conversations. **
To illustrate this, we ran a test where we asked the two bots 2 consecutive questions:
1. Do you know Notion and what is it?
**2. I’m having some issues. **
As you can see, Pi’s answer is gradual, and when I ask the second question, Pi can continue to ask “What problems are you having with Notion?”;
ChatGPT, on the other hand, after giving a long explanation of the first question, did not connect the second question with context, but refreshed the “memory” and restarted the conversation with the user.
![]()
Pi vs. ChatGPT’s different answers
03. About Inflection 2
When Pi was first launched, it used the Inflection-1 model, and on November 22, the Inflection-2 model, a model with 175 billion parameters,** scored only slightly below GPT-4 in various large model evaluations, making it the second largest model in the world. **
Compared with Inflection-1, Inflection-2 has more knowledge, stronger manipulation, and higher logical reasoning ability, and the following chart shows that Inflection-2 is far better than Inflection-1 in various model evaluations, and is on the same level as Google’s PaLM2 GSM8k), Inflection-2 are slightly better than the latter.
![]()
Inflection-1, Inflection-2, and PaLM performance evaluations
Inflection-2 is trained on 5,000 NVIDIA H100 GPUs, and although the training scale is much larger than Inflection-1, Inflection has completed the transition from A100 to H100 thanks to NVIDIA’s help, so Inflection-2 can provide higher service efficiency at a lower cost.
So, this is a milestone step for the Pi update, but it’s just the beginning, Inflection has the world’s largest GPU cluster (22,000 H100s), so Personal AI is likely to progress a lot faster than most people think.
Next, let’s take a closer look at how Inflection-2 performed in each model evaluation.
First of all, on the MMLU (5-shot), a variety of tasks are set from high school to professional level, and Inflection-2 is the best-performing model besides GPT-4, even surpassing Claude 2 with causal chain reasoning ability.
![]()
How each model performs on MMLU (5-shot).
Secondly, in terms of answering general knowledge and science questions, Inflection-2 also achieved a score that exceeded Inflection-1, PaLM-2 (Google), and LLaMA (Meta), and performed very well.
![]()
Scores for each model on the answers to general knowledge and science questions
In addition, Inflection-2 also achieved good results relative to other models in terms of its ability to understand natural language, provide accurate information, and handle complex queries, although it performed slightly lower than PaLM2-Large on Natural Questions(1).
![]()
Scores for each model on NaturalQuestions, Trivia QA
Finally, there is the evaluation of the model’s mathematical and logical reasoning ability, which is not the focus of its training as a model positioned to serve high-EQ chatbots, but you can still see from the image below that Inflection-2 is second only to GPT-4 in terms of mathematical and logical ability. **
![]()
Therefore, based on several reviews, Inflection-2 is already a very powerful LLM, and compared to its previous generation model, it has achieved a lot of growth in various scores, mainly due to the cooperation of NVIDIA, which has provided great help for model training in hardware.
However, Inflection’s own business model is also very promising, not only with its own basic model, ** for ToB sales (and now also offering APIs), but also its own ToC product, Pi (in contrast, a personal assistant rewind.ai integrated with GPT services). **
So, that’s why it’s the second-largest company after OpenAI in terms of funding ($1.5 billion in total).
In addition, Inflection and OpenAI may also represent 2 different trends in the use of models**, for the former, each person only needs 1 AI robot, which can not only provide emotional companionship, but also improve daily productivity, while the latter, which means that people may need a subdivided direction of AI in different scenarios, such as AI doctors, AI lawyers, writing AI assistants, etc. - which is why Microsoft will bet heavily on these 2 company for one of the reasons.
04. Easter Egg: About Mustafa Suleyman
Finally, a little easter egg about the legend of CEO Suleyman and how Inflection was founded.
Suleyman’s humanist outlook was formed by his mother’s role as a nurse and his childhood witness to “human suffering” as a nurse, and his Syrian father, a taxi driver and an enthusiastic community participant.
He later studied philosophy at Oxford, but dropped out of Oxford to help post-911 Muslim youth – though it was during his own admission, that his attitude towards technology and humanism began to take shape.
In 2010, Suleyman founded DeepMind with Demis Hassabis and ShaneLegg, which was subsequently acquired by Google for $650 million in 2014, and in 2016, Suleyman launched the DeepMind Health project at the Royal Society of Medicine to improve the efficiency of the NHS and improve the quality of first-line care.
Later, Suleyman became the head of applied AI at DeepMind, where he was tasked with applying machine learning techniques to a wide range of Google products and processes, and by 2019, his team had launched about 50 AI projects at Google, and the AI was applied to Google’s data centers, reducing cooling costs by 30%. **
In 2020, Suleyman saw the trend towards “conversational” human-computer interaction and joined Google’s natural language research team to begin working on early versions of the LaMDA model.
Subsequently, LaMDA caught the attention of Google executives, but due to conflicts with the current Google business model (users can bypass search ads and get the best answers directly from AI) and concerns about antitrust laws (bypassing third-party creators in search and only Google providing users with answers directly), Suleyman’s personal vision was not taken seriously on a sustained basis - **, so he chose to quit his job and founded what is now Inflection AI, a “conversational” , a “humanist” AI companion with empathy. **
References
