AI’s $344bn single-track bet on LLMs faces growing doubts: Parmy Olson

Key topics:Silicon Valley bets heavily on large language models for AI growthLLMs face high costs, reliability issues, and diminishing returnsStartups explore alternative AI methods beyond text-based models.Sign up for your early morning brew of the BizNews Insider to keep you up to speed with the content that matters. The newsletter will land in your inbox at 5:30am weekdays. Register here.Support South Africa’s bastion of independent journalism, offering balanced insights on investments, business, and the political economy, by joining BizNews Premium. Register here.If you prefer WhatsApp for updates, sign up to the BizNews channel here..By Parmy Olson.Every investor knows not to put all your eggs in one basket. So why is Silicon Valley betting on just one way to build artificial intelligence? This year the world’s four largest tech firms will spend $344 billion on AI, mostly on data centers used to train and run so-called large language models (LLMs) like ChatGPT that can process text, audio and visual content. The technology is largely underpinned by the same technique of predicting tokens that appear next in a sequence.Their spending isn’t all in vain of course. Personal-use chatbots are already growing quickly, with some AI startups starting to break even and businesses still in the early stages of boosting themselves with generative AI. Large language models represent the first AI technique to achieve mainstream adoption at enormous scale: More than 700 million people use ChatGPT each week, for instance. But history is full of people who got fixated on a single “winning” approach to tech, only to fall behind when the landscape suddenly shifted. Think of BlackBerry’s devotion to the physical keyboard before Apple Inc. crushed it with touchscreens, or Yahoo’s big bet on portals while Google quietly dominated search. Could a novel approach to AI suddenly upend all the capital being spent on chatbot technology? Perhaps. China’s DeepSeek offered a glimpse of how unconventional approaches can surprise the market when it released a smaller, more efficient model in January and posted its blueprints on the web. DeepSeek’s model was an LLM, but its method signaled that all the resources being poured into AI research today drove a tide that could raise other boats. Through its history, AI has moved forward by blending past insights with new ideas, and the pursuit of super-intelligent machines may demand no less..Read more:.Chinese AI-tech breakthrough shatters silicon valley’s illusions: Catherine Thorbecke.Much of that exploration now happens at places like Covariant, a Bay Area startup that’s building software to help machines perceive their surrounding space rather than sift through patterns in data. Companies focused on robotics and drones, drug discovery or climate modeling, tend to be those who have naturally stayed away from the language-model obsession because their tech needs to react to physical world conditions in real time. Atman Labs, a British startup, is mining ideas from before the advent of deep learning, which “were also important and have been forgotten,” according to co-founder Sumon Sadhu. Their path has echoes of Google DeepMind’s years-long effort to build super-intelligent AI via different tracks — from game-playing AI systems like AlphaGo to a simulation-based technique called reinforcement learning — before the release of ChatGPT shifted the company’s entire focus onto large language models. Now, some cracks are appearing in the large-language model thesis, from the eye-watering costs to the prospect of diminishing returns. The latest models from OpenAI or Google are only slightly better than the older ones, even as more money is poured into their development. Hallucinations haven’t gone away, muddying the path to adoption for companies in healthcare or legal analysis.A recent study in Nature also shows that the social reasoning abilities of language models — being able to figure out what people really mean in conversation — depend on an extremely small set of model features, and that tiny tweaks can break them. That raises fundamental questions about reliability. Somewhat relatedly, OpenAI admitted last month that ChatGPT’s safeguards for vulnerable people could break down during long conversations. That disclosure came after the bot gave self-harm instructions to a teenager. Some of tech’s more outspoken figures have pointed to the flaws. “Silicon Valley totally effed up in overhyping LLMs,” Palantir Technologies Inc. Chief Executive Officer Alex Karp said at its AI conference last week. Yann LeCun, Meta Platforms Inc.'s chief AI scientist, has long argued that large language models are a "dead end" for smarter machines because they don’t understand their physical surroundings or plan ahead. They’re just “token generators,” he warns. LLMs aren’t going away, but the history of markets shows the dangers of becoming infatuated with a single solution. Investors and businesses should stay alert for technical breakthroughs and be ready for the ground to shift. In technology it can — and often does — before anyone expects..© 2025 Bloomberg L.P.

AI’s $344bn single-track bet on LLMs faces growing doubts: Parmy Olson

Key topics:

Related Stories