From Hype To Hyperscale In AI | AdExchanger

"The vendor landscape is turning into a sea of sameness. So what's worthwhile and what's worth chucking in the bin? The best way to separate AI hype from reality is to roll up your sleeves and try out the tech for yourself, says Ikkjin Ahn, CEO and co-founder of machine learning-based ad tech startup Moloco, on this week's episode of AdExchanger Talks. It's like watching a movie, he says. How do you know if it's good before you even try it?"

"You can build "toy examples" very easily with generative AI, Ahn says, but "the key differentiator is how much you can feed the volume, how much scale you can achieve." That's where AI hyperscale comes in, which involves running AI on huge, cloud-style infrastructure. It's a mix of ultra-fast hardware and optimized software that's been trained to power AI models faster and more efficiently than regular computers."

"Moloco uses AI Hypercomputer, which is Google's branded end-to-end supercomputing stack. It runs on Google Cloud and allows Moloco to process billions of requests a day at 10x speed and relatively low cost. Online advertising is evolving from a reliance on carefully engineered features and audience segments to harnessing the power of these "foundational models" that can learn from huge volumes of raw data, Ahn says."

"It's a process, though. While generative AI models are currently pretty good at answering questions, most are way too slow and expensive to meet the needs of real-time advertising, he says, which demands more powerful models that can operate instantly and affordably. "Think about how long it takes to get an answer from ChatGPT," Ahn says."

Many ad tech vendors claim AI capabilities, producing a homogeneous vendor landscape. Practical, hands-on testing is the most reliable way to distinguish useful AI from hype. True contenders demonstrate the ability to operate at massive scale rather than rely on toy examples. AI hyperscale combines ultra-fast hardware and optimized software on cloud-style infrastructure to run large models faster and more efficiently. Cloud supercomputing stacks can process billions of requests daily at far greater speed and lower cost. Real-time advertising requires models that operate instantly and affordably, which many current generative models cannot yet provide.

Read at AdExchanger