Inflection AI launches Inflection-2, outperforming Llama 2 and PaLM 2

Inflection AI, a startup that created the conversational chatbot Pi, has unveiled its most user-friendly chatbot ever.

The company has launched a new AI model called Inflection-2, which it claims outperforms two well-known models developed by Google and Meta, following closely behind OpenAI's flagship large-scale model, GPT-4.

Inflection-2 was trained on 5000 NVIDIA H100 GPUs using fp8 mixed precision, equivalent to approximately 10²⁵ FLOPs. This newly released model will be integrated into Inflection's chatbot Pi, which was launched in May. Inflection works closely with Microsoft, Nvidia, and CoreWeave to manage its computing cluster.

Inflection-2 surpasses the largest 700 billion parameter version of LLaMA 2, Elon Musk's startup xAI's Grok-1, Google's PaLM 2 Large, and Anthropic's Claude 2, only falling behind GPT-4 in the MMLU task.

The company states that while mathematical and code benchmark tests are not the focus, Inflection-2 performs well in four of them. However, it lags significantly behind GPT-4 in the two benchmark tests where OpenAI shared its results.

It also performs the best in two out of three question-answering task benchmark tests, losing to PaLM 2 Large in one of them.

Compared to Inflection-1, Inflection-2 offers higher cost-effectiveness and speed despite its larger scale. This achievement is attributed to the switch from A100 to H100 GPUs and highly optimized inference implementation. Additionally, Inflection AI plans to leverage the full capacity of its 22,000 GPU cluster to train even larger models.