Hey everyone, it’s November 22, 2023, and today, the startup Inflection AI is making some serious waves in the world of artificial intelligence. Inflection AI just announced the launch new AI model, Inflection-2.
The Inflection AI Journey
To give you some background, Inflection is a relatively young startup, founded back in March of this year. We’ve quickly gained traction, securing an initial investment round of about $200 million. Partnering up with industry giants like Nvidia and Microsoft, we managed to secure a whopping $1.3 billion in investments.
The first model, Inflection-1, was developed in lightning-fast time, and today, they revealed its successor, Inflection-2.
So, what’s so special about Inflection-2, you might wonder? Inflection-2 stands out as one of the top-tier models in the world within its compute class. It proudly holds the title of being the second most capable large language model globally, just after GPT-4.
That’s some serious competition we’re talking about, beating the likes of major players such as Meta, Microsoft, and Google.
If you take a look at the benchmarks (check out Figure 1), you’ll see the dark green representing Inflection-2 outperforming not only its predecessor, Inflection-1, but also Google’s Pal 2 model. With Mustafa Suan, our CEO and former co-founder of Google DeepMind, leading the charge, it’s no surprise Inflection pushing boundaries.
The Power Behind Inflection-2
Now, let’s talk about what fuels this powerhouse. Inflection-2 was trained on a groundbreaking infrastructure using 5,000 Nvidia H100 GPUs. These GPUs are top-of-the-line, the cream of the crop in the realm of data center GPUs, leaving competitors trailing with their A100s.
But hardware is just one part of the equation. What about the software, the precise training that’s essential for optimal performance?
Well, we’re currently fine-tuning Inflection-2 through reinforcement learning and feedback mechanisms. We aim not just for performance but for a model that behaves seamlessly within our AI ecosystem, known as “P”.
P, powered by Inflection-1 currently, is a versatile AI companion. It can brainstorm, journal, learn, summarize, and engage in conversations. Soon, Inflection-2 will join its ranks, enhancing its capabilities further.
Analyzing Benchmark Results
Ah, benchmarks, the tangible proof of a model’s prowess. Inflection-2’s performance across various benchmarks against competitors like Llama 2, Grok 1, PalM 2, Claude 2, and GPT-4 showcases its strength. While benchmarks are valuable, they’re just a part of the larger picture. We believe in our model’s potential but caution against reading too much into marketing materials.
The Road Ahead
Looking into the future, we’re gearing up for even bigger strides. We’re not just stopping at Inflection-2; we’re already training with a staggering 22,000 Nvidia H100 GPUs. Our aim? To build even grander models in the coming months. Think Inflection-3, then Inflection-4, each ten times larger in scale.
While some in the industry are focused on smaller models, we’re stepping into the ring with giants. We’re betting on the demand for these massive models, although it’s too early to predict their market reception.
Our infrastructure is ready; it’s now about intelligent pre-training, fine-tuning, and behavioral optimization through reinforcement learning.
So here we are, a startup with an extraordinary infrastructure, armed with 22,000 Nvidia H100 GPUs and the ambition to create cutting-edge, large-scale language models. It’s a journey into uncharted territory, aiming to rival the best in the industry.
The challenge lies not just in building these models but in ensuring they’re friendly, accurate, and beneficial. Can Inflection outperform the likes of OpenAI’s GPT-4? Only time will tell, but the possibilities with Inflection-2 and beyond are exhilarating.