‘Judgmental’ AI models have become a trend, for better or worse

Call it a justification renaissance.

in the year After the release of OpenAI’s o1the so-called reasoning model was an explosion of reasoning models from rival AI labs. In early November, DeepSeek, an AI research company funded by quantitative traders, launched a preview of its first reasoning algorithm. DeepSeek-R1. That same month, Alibaba’s Gwen team opened what it claims is the first “open” competitor to o1.

But what opened the floodgates? First, the search for new approaches to improve generative AI technology. Like my colleague Max Zeff recently informed“Brute force” methods of scaling models no longer yield the improvements they once did.

There is intense competitive pressure on AI companies to keep up the current pace of innovation. according to according to one estimate, the global artificial intelligence market reached $196.63 billion in 2023 and could be worth $1.81 trillion by 2030.

OpenAI, for one, claimed that reasoning models can “solve more difficult problems” than previous models and represent a step change in generative AI development. But not everyone is convinced that thinking models are the best way forward.

Ameet Talwalkar, Associate Professor of Machine Learning Carnegie Mellon says it considers the first crop of reasoning models to be “pretty impressive.” In the same breath, however, he said he would “question the motives” of anyone who confidently claims they know how far reasoning models will take the industry.

“AI companies have financial incentives to offer rosy predictions about the capabilities of future versions of their technology,” Talwalkar said. “We run the risk of myopically focusing on a single paradigm—so it’s critical for the broader AI research community not to blindly believe the hype and marketing efforts of these companies and focus instead on concrete results.”

The two downsides of thought models are that they are (1) expensive and (2) power hungry.

For example, on OpenAI’s API, the company charges $15 per analysis for every ~750,000 words and $60 for every ~750,000 words the model generates. That’s between 3 and 4 times the price of OpenAI’s latest “ungrounded” model. GPT-4o.

O1 is available on OpenAI’s AI-powered chatbot platform. ChatGPTfree — with limits. But earlier this month, OpenAI presented a more advanced o1 level, o1 pro mode, which costs $2,400 per year.

“The overall cost of (large language model) reasoning certainly doesn’t go down,” UCLA computer science professor Guy Van Den Broeck told TechCrunch.

One of the reasons that mental models are so expensive is that they require a lot of computing resources to run. Unlike most AI, o1 and other reasoning models try to verify their work while doing so. It helps them avoid certain things traps This usually crashes the models, and the downside is that solutions often take longer to obtain.

OpenAI envisions “thinking” future thinking models over hours, days, or even weeks. Operating costs will be higher, the company admits, but revenues – from breakthrough batteries to new cancer drugs – it might be worth it.

The value proposition of today’s reasoning models is less clear. Costa Huang, a researcher and machine learning engineer at the nonprofit Ai2, notes that o1 not a very reliable calculator. And implicit searches on social media a number o1 pro mode errors.

“These reasoning models are specialized and may underperform in general areas,” Huang told TechCrunch. “Some restrictions will be removed sooner than others.”

Van den Broeck argues that reasoning models do not perform actual ability to reason, and therefore the types of tasks they can successfully solve are limited. “True inference works on all problems, not just what is likely (in the model’s training data),” he said. “It’s still a major problem that needs to be overcome.”

Given the strong market incentive to improve thinking models, it’s a safe bet that they’ll only get better over time. After all, it’s not just OpenAI, DeepSeek and Alibaba investing in this new line of AI research. There are VCs and founders in adjacent industries combination Around the idea of ​​a future based on artificial intelligence.

However, Talwalkar worries that large labs will maintain these improvements.

“Big labs have competing reasons to remain secretive, but this lack of transparency severely hampers the research community’s ability to engage with these ideas,” he said. “As more people work in this direction, I expect (thinking of models) to progress rapidly. But while some of the ideas may come from academia, given the financial incentives here, I would expect most, if not all, of the models to be offered by large industrial labs like OpenAI.”