All about the revised Green India Mission to increase forest cover, address climate change
International Yoga Day: Why is it hard to pinpoint the origin of yoga?
QA::Why Madhya Pradesh CM's wish to count snakes and rear king cobras is unfeasible
What new Registration Bill says, why it was introduced
In less than a week, two Large Language Models (LLMs) have been launched that are closer to or even better at tasks than OpenAI’s GPT-4. After Anthropic’s Claude 3 captured the attention of the AI community, Inflection AI launched its latest LLM, Inflection 2.5, an upgrade to its model that powers its friendly chatbot Pi personal assistant.
Pi has been designed to be “empathetic, helpful, and safe”, according to Inflection AI. Following the success of the chatbot, the company introduced its new major foundation model, Inflection 2, in November 2023. At the time of the launch of Inflection 2, the company claimed it to be the best LLM in the world.
Now, the company is back with an upgraded version that is claimed to be as good as GPT-4. Launched on Thursday (March 7), Inflection 2.5 is available to all Pi’s users at pi.ai, iOS, and Android.
According to the makers, Inflection-2.5 is an “upgraded in-house model that is competitive with all the world’s leading LLMs like GPT-4 and Gemini.”
The company claims that the newly upgraded LLM comes with its signature personality and uniquely empathetic fine-tuning. The company claims that its latest model achieved GPT-4’s performance with only 40 per cent of the OpenAI model’s computation power for training.
Besides, it seems Inflection 2.5 has made some stellar strides in areas of IQ such as coding and mathematics. This means that the model has made substantial improvements on key benchmarks. With the new upgrade, Pi has now been endowed with world-class real-time web search capabilities to ensure that users get access to high-quality and up-to-date information in real-time.
Pi is an AI chatbot with which one can have deep and meaningful conversations. To access the chatbot, one needs to log on to Inflection.AI, click on Meet Pi, and simply start talking to the chatbot right away.
Pi was launched at a time when the world was marvelling at OpenAI’s sensational ChatGPT. While ChatGPT offered human-like responses, Pi came with a radically different vigour. Pi is more humane and has been promoted as a chatbot that has a personality. In other words, Inflection AI dubbed it as a chatbot that is “supportive, smart, and there for you anytime”.
While ChatGPT and Gemini were mostly personal assistants that were useful at work, Pi is more like a companion to humans and is free to use. The chatbot comes with a voice, in six distinct voices, to choose from adding life to conversations.
On the technical side, for training purposes, the chatbot has been shown billions of lines of text on the open web. This allows Pi to have conversations with users and answer a wide variety of questions, according to the company. Incidentally, the name Pi stands for personal intelligence and it offers infinite knowledge according to a user’s need.
Upon its launch, Pi was powered by Inflection’s proprietary LLM Inflection-1 which was trained on thousands of NVIDIA H100 GPUs on a very large data set.
Based on the evaluations, Inflection-1 was ranked the best model in its compute class outdoing the likes of GPT-3.5, LLaMA, and PaLM-540B on a wide range of benchmarks. In 2022, the company launched its new model Inflection-2, which was trained on 5,000 NVIDIA H100 GPUs.
Inflection-2 outclassed Google’s flagship LLM PaLM 2 on various benchmarks including MMLU (massive multitask language understanding), TriviaQA (a realistic text-based question-answering dataset), HellaSwag & GSM8k.
HellaSwag is a challenging dataset for evaluating common sense NLI (natural language interpretation) that is typically hard for state-of-the-art AI models. GSM8K stands for Grade School Math 8K and is a dataset of 8,500 high-quality grade school mathematical word problems that require multi-step reasoning.
With the upgraded language model Inflection-2.5, the company has enhanced the emotional quotient of the Pi chatbot. The company claims that with the new LLM, users are talking to the Pi chatbot about a wider range of topics including discussing current events, getting local restaurant recommendations, studying for a biology exam, drafting business plans, coding, and even fun discussions on hobbies.
According to the company, Inflection-2.5 is competitive with GPT-4 in many areas but uses 40 per cent less computing power for training. It achieves 94 per cent of the average performance of GPT-4 on various IQ-oriented tasks. Inflection-2.5 also outperforms its predecessor Inflection-1 on the MMLU benchmark and performs at the 85th percentile of human test-takers on the Physics GRE.
In essence, Inflection-2.5 retains the Pi chatbot’s unique persona along with maintaining extraordinary safety standards. The model achieves this all the while making the chatbot a more helpful model across the board.
Inflection AI based in Palo Alto, California was founded by Mustafa Suleyman, Reid Hoffman, and Karen Simonyan in 2022 to create personal AI for everyone. Suleyman is a British AI researcher who was the co-founder and former head of applied AI at DeepMind, while Hoffman is the co-founder of LinkedIn. In June last year, the company raised $1.3 billion in a funding round led by Microsoft, Reid Hoffman, Eric Schmidt, Bill Gates, and NVIDIA.
