In the vibrant landscape of artificial intelligence, a significant stride has been taken with the development of Inflection-2—a monumental achievement that marks a notable leap forward in the realm of personal AI.
Inflection, on a mission to make personal AI accessible to everyone, had previously introduced Inflection-1, a language model powering Pi. Now, the spotlight turns to Inflection-2, surpassing its predecessor in both capacity and capability. Inflection-2 showcases remarkable improvements in factual knowledge, stylistic control, and reasoning, positioning itself as the second most capable Large Language Model (LLM) globally.
Figure 1: Comparison of Inflection-1, Google’s PaLM 2-Large, and Inflection-2 across a range of commonly used academic benchmarks. (N-shots in parentheses)
The journey of Inflection-2's development is illuminated by its training process on 5,000 NVIDIA H100 GPUs, utilizing fp8 mixed precision for an astounding ~10²⁵ FLOPs. This places it in the same training compute class as Google's flagship PaLM 2 Large model, outperforming it in various standard AI benchmarks such as MMLU, TriviaQA, HellaSwag, and GSM8k.
(The above figure shows results on a wide set of benchmarks ranging from common sense to scientific question answering)
Results on TriviaQA and NaturalQuestions,two question answering tasks:
Efficiency takes center stage as Inflection-2 gears up to power Pi. The transition from A100 to H100 GPUs, coupled with optimized inference implementation, not only reduces costs but also enhances serving speed despite the model's substantial size.
The narrative doesn't stop at technical prowess; safety and responsibility weave through the core of Inflection's ethos. The safety team rigorously evaluates and aligns the models with the best-in-class approaches, ensuring they adhere to the White House's voluntary commitments for safety and security in AI.
Partnerships with industry leaders—NVIDIA, Microsoft, and CoreWeave—have played a pivotal role in making Inflection-1 and Inflection-2 a reality, underlining the collaborative spirit driving advancements in AI.
As the results unfold, the performance of Inflection-2 is laid bare across benchmarks, comparing favorably against Inflection-1 and formidable external models like LLaMA-2, Grok-1, PaLM-2, Claude-2, and GPT-4. From the MMLU benchmark to question answering tasks, Inflection-2 exhibits its prowess, even outperforming competitors in chain-of-thought reasoning.
Ref Table 4: Results on math and coding benchmarks
Mathematical reasoning and coding, not the explicit focus of Inflection-2's training, reveal surprising strengths, showcasing the model's versatility. The journey doesn't end here; plans for scaling to even larger models on the full capacity of a 22,000 GPU cluster are already in motion.
The implications of Inflection-2's advancements ripple across the industry, signaling a transformative era for enterprises. With its enhanced capabilities, cost-efficiency, and rigorous safety measures, Inflection-2 not only promises a new era for personal AI but also presents a blueprint for responsible AI integration in enterprise settings.
Enterprises looking to harness the power of AI must take note of the collaborative approach adopted by Inflection, forming strategic partnerships with industry leaders. This collaborative spirit, coupled with a commitment to safety and ethical considerations, serves as a guiding light for enterprises venturing into the realm of advanced AI.
As the journey of Inflection continues to unfold, scaling to unprecedented heights, the call is extended to those in the industry ready to embrace the future of AI responsibly. The narrative of Inflection-2 is not just a technological breakthrough; it's a beacon illuminating the path towards a new era of responsible and impactful AI in industry and enterprises.
We research, curate and publish daily updates from the field of AI. Paid subscription gives you access to paid articles, a platform to build your own generative AI tools, invitations to closed events and open-source tools.
Consider becoming a paying subscriber to get the latest!