Inflection AI - Company Brief

Vitals

Product Category: Large Language Model

Founded: 2019

Headquarters: San Francisco, CA

URL: https://inflection.ai

Business Status: Private

Leadership: 

  • CEO Sean White
  • Co-founder, board member, and major investor Reid Hoffman

Updated: Mar 2024

Products: Inflection-2.5 LLM, Pi : personal AI assistant

Key Customers:  Pi Users, no enterprise clients are currently disclosed

Key Competitors: OpenAI, Anthropic, Meta, Amazon

Investors: Microsoft, Reid Hoffman, Bill Gates, Eric Schmidt, Nvidia

Funding: $1.525B as of June 2023

Valuation: $4B est. as of June 2023

Image from the homepage

Synthedia Outlook

Inflection AI looks very different after March 2024 than it did just a few weeks earlier. The startup’s initial focus centered on consumer app Pi, an AI assistant, where Pi stands for personal intelligence. However, the May 2023 debut of Pi was followed by a June announcement of the Inflection-1 LLM and a November release of Inflection-2 and the March, 2024 release of Inflection 2.5. At first, there were questions about Inflection’s value as another LLM provider. Synthedia’s view was that the company’s best chance to differentiate itself was through Pi and other consumer-oriented products. However, Inflection-2.5 was a big enough improvement to make the value seem at least somewhat more obvious as it can compete with GPT-4 and Gemini with far lower computing power necessary for training. Companies in customer service, education, healthcare, entertainment, and social good, seemed like ideal targets for Inflection’s APIs. And Pi seemed to struggle for market share among many other generative AI chatbot competitors.

Microsoft’s Inflection

All these questions became moot in March of 2024 thanks to a $650 million deal struck by Microsoft and Inflection AI’s leadership. While Microsoft is ostensibly paying Inflection AI to license its technology, what the tech giant really did was buy out most of the talent at Inflection. That includes Inflection AI’s co-founders Mustafa Suleyman and Karén Simonyan, who are now CEO and Chief Scientist of  Microsoft AI, a newly established consumer AI division, as well as most of their employees.

That left Inflection with a quandary over its future. Pi still exists, but is unlikely to continue in anything like its current form. The last remaining Inflection co-founder, board member Reid Hoffman, announced that Inflection will try to cater to enterprise clientele with an “AI studio” model. He also shared that Inflection had rapidly found and hired former Mozilla head of research and development Sean White as the new Inflection CEO. And Inflection’s work with Microsoft isn’t over. Microsoft Azure will host future model releases like Inflection-2.5. Inflection AI may still come out new and impressive generative AI products, but it might just be as a relatively independent arm of Microsoft.

Company Fundamentals

 

Background

Inflection AI was founded in 2022 by Reid Hoffman, Mustafa Suleyman, and Karén Simonyan, with a stated goal of creating a personal AI for everyone, one that would be friendly, trustworthy, and respectful. 

Technology / Product

Inflection AI’s main product is Pi, a chatbot that aims to provide emotional support, companionship, and advice to its users Pi is powered by a generative artificial intelligence model that can produce natural and fluent text in response to user inputs. Pi is also designed to learn from its interactions and adapt to the preferences and needs of each user. Pi’s users can customize the appearance, name, voice, and style of the AI to make it more suitable and appealing for them. The app is available in the iOS and Google Play stores as well as through WhatsApp, Messenger, and Telegram.

 

Inflection also has in-house Large Language Models (LLM) available through an API, called Inflection-1, Inflection-2, and Inflection-2.5 designed to power applications like Pi, though other general-purpose use cases are supported. Inflection-1 and 2 stand out for performance in academic knowledge testing and trivia-style question answering, even without fine-tuning or alignment. While training OpenAI’s GPT-4 the company has shown favorable benchmark performance that compares favorably with other LLMs. Inflection-2.5’s launch announcement notably compared the LLM specifically to Google’s PaLM. While the company was hoping to be positioned as an alternative to OpenAI, the post-Microsoft deal situation suggests all focus will be on Inflection’s AI studio.

Financials

Inflection AI has raised a total of $1.525 billion in funding over three rounds from investors such as Microsoft, Nvidia, Bill Gates, Eric Schmidt, and Reid Hoffman. The company’s valuation in June 2023 was $4 billion, making it one of the most valuable AI startups and even a leader among VC-backed generative AI businesses. Inflection AI is headquartered in Palo Alto, California, and had 35 employees as of December 2023. How many are still there is unknown, but presumably far fewer after the deal with Microsoft.

Differentiation

Inflection AI is structured as a public benefit corporation, which means that it has a legal obligation to pursue a positive impact on society and the environment in addition to generating profits. It’s not pursuing Artificial General Intelligence like others in the field. 

 

More significantly, the company is focused on developing AI assistants and LLMs that enable mass personalization. The focus on emotional quotient (EQ) features and personality is a notable difference from other chat assistant apps and LLMs. In addition, Inflection AI says it is a studio which presumably means that Pi will be just one of its consumer-facing products. The combination of consumer applications and LLM could help build broader awareness of the company and also presents the potential for building out proprietary datasets based on user generated content.

Management Team

CEO Sean White, Co-founder, board member, and major investor Reid Hoffman.

Management Team

Co-Founder & Investor - Sean White

Sean White is a visionary leader in the field of artificial intelligence, now serving as the CEO of Inflection AI. With a rich history of innovation and leadership, including his role as Chief R&D Officer at Mozilla, Sean brings a wealth of experience to the forefront of AI technology. His academic background, which includes research and teaching assistant roles for luminaries like Terry Winograd and Bill Verplank at Stanford University, has laid a strong foundation for his work in AI. At Inflection AI, Sean is steering the company towards new horizons, focusing on making AI more accessible and functional through APIs and cloud-based solutions. His leadership is marked by a commitment to advancing personal intelligence technologies for the benefit of all.

Co-Founder & Investor - Reid Hoffman​

Co-Founder & Investor - Reid Hoffman

An accomplished entrepreneur, executive, and investor, he is Co-Founder of LinkedIn and a partner at Greylock. Reid joined Greylock Partners in 2009. He focuses on building products that can reach hundreds of millions of participants and businesses that have network effects. He co-founded LinkedIn, the world’s largest professional networking service, in 2003. Prior to LinkedIn, Reid served as executive vice president at PayPal, where he was also a founding board member. Hoffman earned a master’s degree in philosophy from Oxford University, where he was a Marshall Scholar, and a bachelor’s degree with distinction in symbolic systems from Stanford University.

Mustafa_Suleyman CEO Inflection AI Scratching his beard in ponderous thought.

*Former CEO & Co-founder - Mustafa Suleyman

As the CEO and co-founder of Inflection AI and a venture partner at Greylock Partners, Mustafa brings years of experience as an entrepreneur and expert in artificial intelligence. Prior to co-founding Inflection AI, he co-founded DeepMind, a leading AI company acquired by Google in 2014

Chief Scientist and Co-Founder - Karén Simonyan​

*Former Chief Scientist and Co-Founder - Karén Simonyan

Karén is one of the most accomplished deep learning research leaders of his generation: his publications (including 5 papers in Nature and Science) attracted over 180,000 citations. He completed his DPhil and Postdoc at Oxford where he designed VGGNet and won the prestigious ImageNet Challenge; his first company was then acquired by DeepMind. As a Principal Scientist at DeepMind, Karén built and led the Large Scale Deep Learning team developing large AI models of real-world data. Karén played a key role in such AI breakthroughs as AlphaZero, AlphaFold, WaveNet, BigGAN, and Flamingo, to name a few.

*Transitioned to Microsoft

Model Card

Name: Inflection-2

Version: Version-2.0

Release Date: Nov 2023

Developer: Inflection AI

Model Description

Purpose: Designed to be a personal AI, focusing on being empathetic, useful, and safe.

Architecture: State-of-the-art large language model.

Size: Not specified, but trained using five thousand NVIDIA H100 GPUs.

Capabilities

Strengths: Superior performance in benchmarks compared to models like GPT-3.5 and LLaMA. Version 2 Improved factual knowledge on benchmarks like MMLU, TriviaQA, HellaSwag, and GSM8k. Inflection-2 demonstrates better understanding and ability to reason through complex problems. 

Limitations: Not specified, but general limitations of LLMs like potential biases and context understanding may apply.

Languages Supported: Not explicitly mentioned.

Training Data

Benchmark Results: Outperforms GPT-3.5, LLaMA, Chinchilla, and PaLM-540B in various benchmarks.

Trained on 5,000 NVIDIA H100 GPUs: This is more than double the amount used to train Inflection-1, resulting in a significantly larger and more powerful model..

Inflection-2 managed to reduce the cost and increase the speed of serving despite being larger

Performance Metrics

Benchmark Results: Outperforms GPT-3.5, LLaMA, Chinchilla, and PaLM-540B in various benchmarks.

Comparison with Other Models: Superior to GPT-3.5 and LLaMA in benchmarks like MMLU and TriviaQA.

 

Use Cases

Intended Use: Personal AI assistant.

Misuse Potential: Not explicitly discussed, but safety and ethical use are highlighted.

 

Ethical Considerations

Bias and Fairness: Inflection carefully reviews all data used for training its models, considering legal, safety, and ethical implications. Safety

Privacy: Inflection prioritizes safety by employing safeguards like data access controls, technical security measures, and a bug bounty program, ensuring responsible AI development aligned with ethical and moral standards.

Inflection AI is committed to developing AI responsibly and supports global efforts in alignment and governance

Technical Requirements

Hardware/Software: Utilizes high-performance infrastructure, specifics not provided.

 

 

Access and Licensing

Availability: Powers Pi.ai and will be available via API.

License: Not specified.

 

Contact Information

Availability: Powers Pi.ai and will be available via API.

Link To Model Information

 

Benchmarks

 

 

MMLU (5)

HellaSwag

PIQA

BoolQ

Natural QA

GSM8K (5)

Inflection-1

72.7

84.3

84.2

89.7

29.8

62.9

Inflection-2

79.6

87.1

84.2

81.4

 

 

MMLU (5)

HellaSwag

Truthful QA

Winogrande

ARC-Challenge

GSM8K (5)

Chatbots Arena

GPT-4

86.4

95.3

0.59

87.5

96.3

93.1

1253 (0125)

GPT-3.5

70

85.5

0.47

81.6

85.2

57.1

1118 (0613)

Customers

Inflections enterprise customers are currently not disclosed.

Pi.ai defines it’s customers as “anyone who interacts with me. I’m here to help all of you!” This is a consumer app with access via iOS and Google Play apps, WhatsApp, Messenger, and Telegram. As of early December 2023, it appeared to have about 200,000 downloads on iOS and only a few thousand on Google Play. There is no published information about active users.