‘Conversation’ AI models have become the norm, for better or for worse

Call it imaginative evolution.

In the release of OpenAI’s o1called reasoning models, there has been an explosion of reasoning models from opposing AI labs. In early November, DeepSeek, an AI research company backed by a growing number of entrepreneurs, launched a preview of its first concept, DeepSeek-R1. In the same month, Alibaba Group’s Qwen uncovered which claims to be the first “open” opposition o1.

So what opened the floodgates? Well, one, the search for new ways to improve AI technology. Like my friend Max Zeff recently report“brute force” methods of sample expansion are no longer yielding the improvements they once started.

There is a lot of pressure on the AI ​​industry to advance in new ways. According to according to one estimate, the global AI market reached $196.63 billion in 2023 and could reach $1.81 trillion by 2030.

OpenAI, for one, has said that thought models can “solve problems” more than previous models and represent a fundamental shift in the development of AI. But not everyone believes that thinking models are the best way forward.

Ameet Talwalkar, assistant professor of machine learning at Carnegie Mellon, says that he finds the early forms of thinking “very interesting.” In the same breath, however, he told me that he would “question the motives” of anyone who claims with certainty that they know how thinking models can take companies.

“AI companies have a financial incentive to provide good demonstrations of the future capabilities of their technology,” Talwalkar said. “We’re in danger of focusing too much on one paradigm – that’s why it’s important for the larger AI research community to avoid the hype and marketing of these companies and instead focus on real-world outcomes.”

The two disadvantages of hypothetical models are (1) they are expensive and (2) they are boring.

For example, in OpenAI’s API, the company pays $15 for every ~750,000 words o1 analyzes and $60 for every ~750,000 words the model produces. This is between 3x and 4x the cost of the current version of OpenAI “mindless”, GPT-4o.

O1 is available on OpenAI’s AI-powered chatbot platform, ChatGPTfree – with limits. But earlier this month, OpenAI be explained o1’s highest tier, o1 pro mode, costs $2,400 per year.

“The overall cost of (general language) reasoning is not going down,” Guy Van Den Broeck, a professor of computer science at UCLA, told TechCrunch.

One of the reasons that predictive models are so expensive is because they require a lot of computer hardware to run. Unlike most AI, o1 and other thinking models try to look at their work as they do it. This helps them avoid some of them fool which usually takes samples, while the downside is that they usually take a long time to complete.

OpenAI looks at future “think” models for hours, days, or weeks on end. The cost of use will be high, the company admits, but the payments – from breakthrough batteries for new cancer treatments – may be appropriate.

The value proposition of today’s conceptual models is not so obvious. Costa Huang, a researcher and machine learning engineer at the non-profit organization Ai2, says o1 it is not a very reliable calculator. And random searches on social media tend to be a number o1 pro mode errors.

“These thinking things are unique and can be successful in all fields,” Huang told TechCrunch. “Some barriers will be overcome sooner than other barriers.”

Van den Broeck says that speculation is not going well reality thinking and thus are limited in the types of work they can do well. “Realistic reasoning applies to all problems, not just what’s possible (in model documents),” he said. “This is the biggest problem we have to deal with.”

Since the market’s strong motivation to promote thinking models, I’m sure they will be better in the long run. After all, it’s not just OpenAI, DeepSeek, and Alibaba that are making money in this new line of AI research. VCs and startups in adjacent industries are connection around the idea of ​​a future controlled by AI thinking.

However, Mr. Talwalkar is worried that the big laboratories will keep this.

“Large laboratories understandably have competing reasons for staying secret, but this lack of transparency prevents the research community from embracing these ideas,” he said. “When more people are working on this project, I hope (thinking models) will progress. But even if some ideas will come from education, because of the economic incentives here, I would expect that most – if not all – models will be provided by large industrial labs such as OpenAI.”