Inflection AI has achieved a remarkable breakthrough with the completion of Inflection-2, now the world’s best model in its compute class and the second most capable large language model (LLM) globally. This new model surpasses its predecessor, Inflection-1, exhibiting enhanced factual knowledge, stylistic control, and reasoning abilities​​.

Inflection-2’s training utilized 5,000 NVIDIA H100 GPUs, operating in fp8 mixed precision, amassing approximately 10²⁵ FLOPs. This substantial computing power places it in the same class as Google’s PaLM 2 Large model, yet Inflection-2 outshines its competitor in most standard AI performance benchmarks​​.

Notably, Inflection-2 is designed for serving efficiency, soon to power Pi at pi.ai. A switch from A100 to H100 GPUs, coupled with optimized inference implementation, has remarkably reduced costs and increased serving speed, despite the model’s significantly larger size​​.

Emphasizing safety, security, and trustworthiness, Inflection takes pride in rigorous model evaluations and alignment approaches. The team’s commitment to global alignment and governance mechanisms for AI technology is evident in their participation in the White House’s July 2023 voluntary commitments​​.

Benchmarking against leading models like LLaMA-2, Grok-1, PaLM-2, Claude-2, and GPT-4, Inflection-2 demonstrates superior performance in a variety of tasks, from common sense to scientific question-answering, even outperforming Claude 2 in chain-of-thought reasoning​​​​. Interestingly, although not the primary focus, Inflection-2 also excels in code and mathematical reasoning benchmarks, showing potential for further enhancement through fine-tuning on code-heavy datasets​​.

As Inflection AI continues to expand, they invite collaboration and contribution towards their mission of creating personal AI for everyone, pushing the limits of scale and safety in AI development​​.