AI research takes a backseat to profits as Silicon Valley prioritizes products over safety, experts say

CNBC

14-05-2025

Not long ago, Silicon Valley was where the world's leading artificial intelligence experts went to perform cutting-edge research.
Meta, Google and OpenAI opened their wallets for top talent, giving researchers staff, computing power and plenty of flexibility. With the support of their employers, the researchers published high-quality academic papers, openly sharing their breakthroughs with peers in academia and at rival companies.
But that era has ended. Now, experts say, AI is all about the product.
Since OpenAI released ChatGPT in late 2022, the tech industry has shifted its focus to building consumer-ready AI services, in many cases prioritizing commercialization over research, AI researchers and experts in the field told CNBC. The profit potential is massive — some analysts predict $1 trillion in annual revenue by 2028. The prospective repercussions terrify the corner of the AI universe concerned about safety, industry experts said, particularly as leading players pursue artificial general intelligence, or AGI, which is technology that rivals or exceeds human intelligence.
In the race to stay competitive, tech companies are taking an increasing number of shortcuts when it comes to the rigorous safety testing of their AI models before they are released to the public, industry experts told CNBC.
James White, chief technology officer at cybersecurity startup CalypsoAI, said newer models are sacrificing security for quality, that is, better responses by the AI chatbots. That means they're less likely to reject malicious kinds of prompts that could cause them to reveal ways to build bombs or sensitive information that hackers could exploit, White said.
"The models are getting better, but they're also more likely to be good at bad stuff," said White, whose company performs safety and security audits of popular models from Meta, Google, OpenAI and other companies. "It's easier to trick them to do bad stuff."
The changes are readily apparent at Meta and Alphabet, which have deprioritized their AI research labs, experts say. At Facebook's parent company, the Fundamental Artificial Intelligence Research, or FAIR, unit has been sidelined by Meta GenAI, according to current and former employees. And at Alphabet, the research group Google Brain is now part of DeepMind, the division that leads development of AI products at the tech company.
CNBC spoke with more than a dozen AI professionals in Silicon Valley who collectively tell the story of a dramatic shift in the industry away from research and toward revenue-generating products. Some are former employees at the companies with direct knowledge of what they say is the prioritization of building new AI products at the expense of research and safety checks. They say employees face intensifying development timelines, reinforcing the idea that they can't afford to fall behind when it comes to getting new models and products to market. Some of the people asked not to be named because they weren't authorized to speak publicly on the matter.
When Joelle Pineau, a Meta vice president and the head of the company's FAIR division, announced in April that she would be leaving her post, many former employees said they weren't surprised. They said they viewed it as solidifying the company's move away from AI research and toward prioritizing developing practical products.
"Today, as the world undergoes significant change, as the race for AI accelerates, and as Meta prepares for its next chapter, it is time to create space for others to pursue the work," Pineau wrote on LinkedIn, adding that she will formally leave the company May 30.
Pineau began leading FAIR in 2023. The unit was established a decade earlier to work on difficult computer science problems typically tackled by academia. Yann LeCun, one of the godfathers of modern AI, initially oversaw the project, and instilled the research methodologies he learned from his time at the pioneering AT&T Bell Laboratories, according to several former employees at Meta. Small research teams could work on a variety of bleeding-edge projects that may or may not pan out.
The shift began when Meta laid off 21,000 employees, or nearly a quarter of its workforce, starting in late 2022. CEO Mark Zuckerberg kicked off 2023 by calling it the "year of efficiency." FAIR researchers, as part of the cost-cutting measures, were directed to work more closely with product teams, several former employees said.
Two months before Pineau's announcement, one of FAIR's directors, Kim Hazelwood, left the company, two people familiar with the matter said. Hazelwood helped oversee FAIR's NextSys unit, which manages computing resources for FAIR researchers. Her role was eliminated as part of Meta's plan to cut 5% of its workforce, the people said.
OpenAI's 2022 launch of ChatGPT caught Meta off guard, creating a sense of urgency to pour more resources into large language models, or LLMs, that were captivating the tech industry, the people said.
In 2023, Meta began heavily pushing its freely available and open-source Llama family of AI models to compete with OpenAI, Google and others.
With Zuckerberg and other executives convinced that LLMs were game-changing technologies, management had less incentive to let FAIR researchers work on far-flung projects, several former employees said. That meant deprioritizing research that could be viewed as having no impact on Meta's core business, such as FAIR's previous health care-related research into using AI to improve drug therapies.
Since 2024, Meta Chief Product Officer Chris Cox has been overseeing FAIR as a way to bridge the gap between research and the product-focused GenAI group, people familiar with the matter said. The GenAI unit oversees the Llama family of AI models and the Meta AI digital assistant, the two most important pillars of Meta's AI strategy.
Under Cox, the GenAI unit has been siphoning more computing resources and team members from FAIR due to its elevated status at Meta, the people said. Many researchers have transferred to GenAI or left the company entirely to launch their own research-focused startups or join rivals, several of the former employees said.
While Zuckerberg has some internal support for pushing the GenAI group to rapidly develop real-world products, there's also concern among some staffers that Meta is now less able to develop industry-leading breakthroughs that can be derived from experimental work, former employees said. That leaves Meta to chase its rivals.
A high-profile example landed in January, when Chinese lab DeepSeek released its R1 model, catching Meta off guard. The startup claimed it was able to develop a model as capable as its American counterparts but with training at a fraction of the cost.
Meta quickly implemented some of DeepSeek's innovative techniques for its Llama 4 family of AI models that were released in April, former employees said. The AI research community had a mixed reaction to the smaller versions of Llama 4, but Meta said the biggest and most powerful Llama 4 variant is still being trained.
The company in April also released security and safety tools for developers to use when building apps with Meta's Llama 4 AI models. These tools help mitigate the chances of Llama 4 unintentionally leaking sensitive information or producing harmful content, Meta said.
"Our commitment to FAIR remains strong," a Meta spokesperson told CNBC. "Our strategy and plans will not change as a result of recent developments."
In a statement to CNBC, Pineau said she is enthusiastic about Meta's overall AI work and strategy.
"There continues to be strong support for exploratory research and FAIR as a distinct organization in Meta," Pineau said. "The time was simply right for me personally to re-focus my energy before jumping into a new adventure."
Meta on Thursday named FAIR co-founder Rob Fergus as Pineau's replacement. Fergus will return to the company to serve as a director at Meta and head of FAIR, according to his LinkedIn profile. He was most recently a research director at Google DeepMind.
"Meta's commitment to FAIR and long term research remains unwavering," Fergus said in a LinkedIn post. "We're working towards building human-level experiences that transform the way we interact with technology and are dedicated to leading and advancing AI research."
Google released its latest and most powerful AI model, Gemini 2.5, in March. The company described it as "our most intelligent AI model," and wrote in a March 25 blog post that its new models are "capable of reasoning through their thoughts before responding, resulting in enhanced performance and improved accuracy."
For weeks, Gemini 2.5 was missing a model card, meaning Google did not share information about how the AI model worked or its limitations and potential dangers upon its release.
Model cards are a common tool for AI transparency.
A Google website compares model cards to food nutrition labels: They outline "the key facts about a model in a clear, digestible format," the website says.
"By making this information easy to access, model cards support responsible AI development and the adoption of robust, industry-wide standards for broad transparency and evaluation practices," the website says.
Google wrote in an April 2 blog post that it evaluates its "most advanced models, such as Gemini, for potential dangerous capabilities prior to their release." Google later updated the blog to remove the words "prior to their release."
Without a model card for Gemini 2.5, the public had no way of knowing which safety evaluations were conducted or whether DeepMind checked for dangerous capabilities at all.
In response to CNBC's inquiry on April 2 about Gemini 2.5's missing model card, a Google spokesperson said that a "tech report with additional safety information and model cards are forthcoming." Google published an incomplete model card on April 16 and updated it on April 28, more than a month after the AI model's release, to include information about Gemini 2.5's "dangerous capability evaluations."
Those assessments are important for gauging the safety of a model — whether people can use the models to learn how to build chemical or nuclear weapons or hack into important systems. These checks also determine whether a model is capable of autonomously replicating itself, which could lead to a company losing control of it. Running tests for those capabilities requires more time and resources than simple, automated safety evaluations, according to industry experts.
The Financial Times in March reported that Google DeepMind CEO Demis Hassabis had installed a more rigorous vetting process for internal research papers to be published. The clampdown at Google is particularly notable because the company's "Transformers" technology gained recognition across Silicon Valley through that type of shared research. Transformers were critical to OpenAI's development of ChatGPT and the rise of generative AI.
Google co-founder Sergey Brin told staffers at DeepMind and Gemini in February that competition has accelerated and "the final race to AGI is afoot," according to a memo viewed by CNBC. "We have all the ingredients to win this race but we are going to have to turbocharge our efforts," he said in the memo.
Brin said in the memo that Google has to speed up the process of testing AI models, as the company needs "lots of ideas that we can test quickly."
"We need real wins that scale," Brin wrote.
In his memo, Brin also wrote that the company's methods have "a habit of minor tweaking and overfitting" products for evaluations and "sniping" the products at checkpoints. He said employees need to build "capable products" and to "trust our users" more.
"We can't keep building nanny products," Brin wrote. "Our products are overrun with filters and punts of various kinds."
A Google spokesperson told CNBC that the company has always been committed to advancing AI responsibly.
"We continue to do that through the safe development and deployment of our technology, and research contributions to the broader ecosystem," the spokesperson said.
The debate of product versus research is at the center of OpenAI's existence. The company was founded as a nonprofit research lab in 2015 and is now in the midst of a contentious effort to transform into a for-profit entity.
That's the direction co-founder and CEO Sam Altman has been pushing toward for years. On May 5, though, OpenAI bowed to pressure from civic leaders and former employees, announcing that its nonprofit would retain control of the company even as it restructures into a public benefit corporation.
Nisan Stiennon worked at OpenAI from 2018 to 2020 and was among a group of former employees urging California and Delaware not to approve OpenAI's restructuring effort. "OpenAI may one day build technology that could get us all killed," Stiennon wrote in a statement in April. "It is to OpenAI's credit that it's controlled by a nonprofit with a duty to humanity."
But even with the nonprofit maintaining control and majority ownership, OpenAI is speedily working to commercialize products as competition heats up in generative AI. And it may have rushed the rollout of its o1 reasoning model last year, according to some portions of its model card.
Results of the model's "preparedness evaluations," the tests OpenAI runs to assess an AI model's dangerous capabilities and other risks, were based on earlier versions of o1. They had not been run on the final version of the model, according to its model card, which is publicly available.
Johannes Heidecke, OpenAI's head of safety systems, told CNBC in an interview that the company ran its preparedness evaluations on near-final versions of the o1 model. Minor variations to the model that took place after those tests wouldn't have contributed to significant jumps in its intelligence or reasoning and thus wouldn't require additional evaluations, he said. Still, Heidecke acknowledged that OpenAI missed an opportunity to more clearly explain the difference.
OpenAI's newest reasoning model, o3, released in April, seems to hallucinate more than twice as often as o1, according to the model card. When an AI model hallucinates, it produces falsehoods or illogical information.
OpenAI has also been criticized for reportedly slashing safety testing times from months to days and for omitting the requirement to safety test fine-tuned models in its latest "Preparedness Framework."
Heidecke said OpenAI has decreased the time needed for safety testing because the company has improved its testing effectiveness and efficiency. A company spokesperson said OpenAI has allocated more AI infrastructure and personnel to its safety testing, and has increased resources for paying experts and growing its network of external testers.
In April, the company shipped GPT-4.1, one of its new models, without a safety report, as the model was not designated by OpenAI as a "frontier model," which is a term used by the tech industry to refer to a bleeding-edge, large-scale AI model.
But one of those small revisions caused a big wave in April. Within days of updating its GPT-4o model, OpenAI rolled back the changes after screenshots of overly flattering responses to ChatGPT users went viral online. OpenAI said in a blog post explaining its decision that those types of responses to user inquiries "raise safety concerns — including around issues like mental health, emotional over-reliance, or risky behavior."
OpenAI said in the blogpost that it opted to release the model even after some expert testers flagged that its behavior "'felt' slightly off."
"In the end, we decided to launch the model due to the positive signals from the users who tried out the model. Unfortunately, this was the wrong call," OpenAI wrote. "Looking back, the qualitative assessments were hinting at something important, and we should've paid closer attention. They were picking up on a blind spot in our other evals and metrics."
Metr, a company OpenAI partners with to test and evaluate its models for safety, said in a recent blog post that it was given less time to test the o3 and o4-mini models than predecessors.
"Limitations in this evaluation prevent us from making robust capability assessments," Metr wrote, adding that the tests it did were "conducted in a relatively short time."
Metr also wrote that it had insufficient access to data that would be important in determining the potential dangers of the two models.
The company said it wasn't able to access the OpenAI models' internal reasoning, which is "likely to contain important information for interpreting our results." However, Metr said, "OpenAI shared helpful information on some of their own evaluation results."
OpenAI's spokesperson said the company is piloting secure ways of sharing chains of thought for Metr's research as well as for other third-party organizations.
Steven Adler, a former safety researcher at OpenAI, told CNBC that safety testing a model before it's rolled out is no longer enough to safeguard against potential dangers.
"You need to be vigilant before and during training to reduce the chance of creating a very capable, misaligned model in the first place," Adler said.
He warned that companies such as OpenAI are backed into a corner when they create capable but misaligned models with goals that are different from the ones they intended to build.
"Unfortunately, we don't yet have strong scientific knowledge for fixing these models — just ways of papering over the behavior," Adler said.