New models are released by OpenAI, Google, and Mistral, increasing conflict
3 min read
Launches occur within 12 hours of each other, with more industry activity expected in the summer
As the industry prepares for higher activity during the summer, OpenAI, Google, and the French AI firm Mistral all released updated versions of their state-of-the-art AI models within a 12-hour period of one another.
This extraordinary rush of releases coincides with the industry getting ready for the next major release of GPT, the platform behind OpenAI’s well-known chatbot Chat-GPT.
The first release occurred shortly after Nick Clegg’s appearance at an event in London, where he confirmed reports that Meta’s third version of its AI model, Llama, would be published within weeks.
The first release occurred shortly after Nick Clegg’s appearance at an event in London, where he confirmed reports that Meta’s third version of its AI model, Llama, would be published within weeks.
Google revealed Gemini Pro 1.5, its most powerful large language model and a rival to Meta’s Llama, seven hours after Clegg left the stage. With this release, there was a 50 request daily limit on the free tier.
An hour later, OpenAI released their most recent frontier model, GPT-4 Turbo. Gemini Pro 1.5 and GPT-4 Turbo are “multimodal” systems that can handle data other than text. They can process visual input; Gemini can process both video and audio.
Early in the morning in France, Mistral, an AI firm formed by a few of Clegg’s old colleagues from the AI team at Meta, released Mixtral 8x22B, their frontier model. In contrast to its American substitutes, Mixtral was distributed via a straightforward download link for a 281GB file. Like Meta, Mistral takes a “open source” stance, allowing anybody to download and improve its AI systems without restriction.
This method has drawn criticism for its possible risk because it stops developers from taking action to stop their systems from being used illegally or to take models out of circulation if biases or problems are found and need to be fixed. Meta and others contend that in the end, this method produces better results than systems that are “controlled by a small number of very large, well-funded companies in California.”
A report from The Information states that the initial distribution of Meta’s Llama 3 is anticipated to be in smaller, less powerful variants. This summer, the business intends to progressively introduce its most powerful frontier model. But there could be fierce competition for it: According to reports, OpenAI plans to release its next GPT model, GPT-5, in a similar amount of time. Brad Lightcap, the company’s chief operating officer, told the Financial Times that it would be available “soon.”
Experts have raised concerns about whether the “large language model” approach, which is common among all frontier AI systems, may be reaching its limits. Meta’s chief AI scientist, Yann LeCun, responded to a claim from xAI founder Elon Musk, saying, “We hear a lot of people saying: ‘Oh my God, we’re going to get [artificial general intelligence] within the next year.’ It’s just not happening. We have AI systems that can pass the bar exam, but they can’t clear up your dinner table and fill up the dishwasher. We have systems that manipulate language and trick us into thinking they are intelligent, but they cannot understand the world.”
LeCun proposed that researchers focus on developing what he termed “objective-driven” AI, which has the capacity to reason and plan in the real world, rather than solely focusing on language processing.
This approach could lead to AI systems with genuinely superhuman capabilities, according to LeCun. He described this as “more of a vision than anything else,” but noted that it is progressing exponentially, expressing confidence in its eventual realization.
