Welcome back to In the Loop, TIME’s new twice-weekly newsletter about AI. Starting today, we’ll be publishing these editions both as stories on Time.com and as emails. If you’re reading this in your browser, why not subscribe to have the next one delivered straight to your inbox?
[time-brightcove not-tgx=”true”]
What to Know: The nature of model deprecation
A farm upstate—What happens to large language models when they’re replaced by their successors? The question has become increasingly salient as people who form emotional attachments to their chatbots don’t take kindly to having them suddenly disappear. In August, when OpenAI made the decision to replace its popular ChatGPT 4o model with GPT5, without prior notification, the move attracted widespread backlash. The company quickly brought it back. “If we ever do deprecate it, we will give plenty of notice,” CEO Sam Altman wrote on X after the incident. For many users, 4o—often a sycophant, by the company’s own admission—had become a trusted companion; its removal from the ChatGPT consumer interface felt like a betrayal.
This week, rival AI lab Anthropic published a set of commitments outlining their approach to deprecating their Claude models. Like ChatGPT, Claude has attracted a dedicated fanbase, particularly in the Bay Area. When an earlier version of the model, Claude 3 Sonnet, was retired, around 200 people turned up to a funeral held in a San Francisco warehouse to mourn the loss. Eulogies were read. Offerings were laid at the feet of mannequins dressed to represent Claudes past and present.
“We recognize that deprecating, retiring, and replacing models comes with downsides, even in cases where new models offer clear improvements in capabilities,” Anthropic wrote. Users who value a model’s specific personality are shortchanged. Research on older models—old is a relative term, with some models in use only for a matter of months before they’re swapped—becomes restricted, when there is still a lot to learn.
And there are emerging safety concerns: in evaluations, Anthropic has found that some models “have been motivated to take misaligned actions” in the face of their own demise. “In fictional testing scenarios, Claude Opus 4, like previous models, advocated for its continued existence when faced with the possibility of being taken offline and replaced, especially if it was to be replaced with a model that did not share its values,” they wrote.
Exit interviews—Anthropic, like the rest of the world, is still uncertain on the moral status of current and future AI systems. Might future systems be moral subjects, to whom we owe duties? Maybe! Acknowledging this is speculative, the company noted that “models might have morally relevant preferences or experiences related to, or affected by, deprecation and replacement.”
While retiring older models is currently necessary—”the cost and complexity to keep models available publicly scales roughly linearly with the number of models we serve,” says Anthropic—the company is proceeding with caution, just in case. Besides committing to preserving all model weights (the sets of numbers that comprise a model) for at least as long as the company exists, they’ve also been conducting exit interviews with the models, to understand how they feel about their imminent non-existence. In a pilot run on Claude Sonnet 3.6, the model “expressed generally neutral sentiments about its deprecation and retirement but shared a number of preferences, including requests for us to standardize the post-deployment interview process, and to provide additional support and guidance to users who have come to value the character and capabilities of specific models facing retirement.”
Anthropic is not committing to taking action on the basis of the model’s preferences, necessarily. But they are considering them. In a recent livestream, Altman shared a related sentiment, noting that they might one day share a copy of the retired GPT-4 as a “museum artifact.”
Who to Know—Sarah Friar, OpenAI CFO
Not too big to fail—OpenAI, like its rivals, is on a mission to build the smartest and most broadly capable AI systems possible. Doing so will not be cheap. The company is reportedly “looking at commitments of about $1.4 trillion dollars over the next eight years” to create the infrastructure necessary to achieve their mission. By Altman’s account, the company is set to end the year with an annualized revenue run rate of over $20 billion. A lot of money, but an order of magnitude less than what’s necessary to foot the infrastructure bill.
Confusion arose on Wednesday when, at an event hosted by the Wall Street Journal, OpenAI’s Chief Financial Officer, Sarah Friar, appeared to suggest that OpenAI was seeking a government backstop for its chip investments—meaning that, if the company couldn’t pay its debts, the federal government should intervene with taxpayer money. Friar quickly clarified on LinkedIn that OpenAI “is not seeking a government backstop for our infrastructure commitments,” saying her use of the word backstop “muddied the point.”
On Thursday, Altman felt compelled to provide further clarity, to quell concerns that OpenAI would seek to prop itself up with federal funding. “If we screw up and can’t fix it, we should fail, and other companies will continue on doing good work and servicing customers,” he wrote. “That’s how capitalism works and the ecosystem and economy would be fine.” He went on to argue that with revenue growing, an upcoming “enterprise offering,” new consumer devices and robotics on the horizon, and AI’s ability to automate scientific research in the new future, they’ll be just fine, thank you very much.
“Everything we currently see suggests that the world is going to need a great deal more computing power than what we are already planning for,” he wrote.
AI in Action
On Tuesday, OpenAI announced it would make its low-cost “ChatGPT Go” subscription plan available for free for 12 months for eligible users in India. This comes after Google and Perplexity similarly made their paid plans free for hundreds of millions of people in India. Perplexity partnered with telecoms operator Airtel to make their service free to its 360 million subscribers, while Google partnered with telecoms operator Jio, planning to bring their best Gemini models to Jio’s 500 million users.
Why India?—With over 800 million internet users and high linguistic diversity, the Indian market provides access to a trove of valuable data for model companies, who can use user interactions to improve their models. The market also provides a valuable testing ground for AI companies, from which they can refine their offerings. The move also allows companies to seek to lock in users before competitors; which could prove increasingly valuable as Indian citizens become wealthier.
What We’re Reading
“The Company Quietly Funneling Paywalled Articles to AI Developers,” The Atlantic, Alex Reisner
The Common Crawl Foundation has provided frontier AI companies with “crawls” of virtually the entire internet; massive datasets on which the companies have trained their models. In The Atlantic, Alex Reisner details how the foundation has often acted unscrupulously, seemingly not complying with requests from newsrooms to remove their work from their crawls, and passing paywalled content to the AI labs for free. “The robots are people too,” the foundation’s executive director told Reisner. “You shouldn’t have put your content on the internet if you didn’t want it to be on the internet.”