AI wars: Google tries to catch up to OpenAI
Gemini could be Google’s most important algorithm in years, said Will Knight in Wired. The search giant released its latest answer to OpenAI’s ChatGPT last week, hoping to use Gemini “to reestablish itself as the world leader in artificial intelligence.” Google describes Gemini as a fundamentally new kind of AI model, one that is “natively multimodal” because “it was trained on images, video, and audio rather than just text.” Starting right away, Google’s Bard chatbot will be powered by Gemini Pro, but a more capable version, Gemini Ultra, will be folded into Bard later in 2024. In a demonstration, Gemini “responded to a video, created simple puzzles, and asked for ideas involving a map of the world.”
The only problem is that parts of the demonstration were faked, said Devin Coldewey in TechCrunch. A Google-produced video called “Hands On With Gemini” purported to show that the new system “can be flexible” and respond easily to images and video. But the demo was “exaggerated for effect” and edited to make it appear as if Gemini worked faster and more intuitively than it does in reality. That’s not an auspicious start.
In any case, Gemini’s capabilities don’t change the AI equation, said Parmy Olson in Bloomberg. The company says its Gemini Ultra outperforms OpenAI’s most advanced algorithm, GPT-4,
“on most standard benchmarks.” But it’s only by a tiny margin. “In other words, Google’s top AI model has made only narrow improvements on something OpenAI completed work on at least a year ago.” It’s not clear where generative AI goes next, said Melissa Heikkilä and Will Douglas Heaven in MIT Technology Review. “For the average user,” Gemini’s incremental improvements over GPT-4 “might not make much difference.” Google also hasn’t solved the consistent problem with large language models: that the chatbots “regularly make things up.”
OpenAI may have already hit another level, said Timothy B.
Lee in Ars Technica. A few days before OpenAI fired (and then re-hired) its chief executive, Sam Altman, word got out that the company “had made a technical breakthrough,” a new model it was dubbing Q* (pronounced cue-STAR). Not much is known about the secretive model other than that it can “solve gradeschool math problems.” But an important clue is OpenAI’s decision to hire computer scientist Noam Brown earlier this year. “Brown developed the first AI that could play poker at a superhuman level.” He built another AI that could play Diplomacy, a strategic game of alliances. “This seems like a good background for someone trying to improve the reasoning ability of large language models,” a step toward AI that can think like a human.