The cheap new AI model taking aim at OpenAI and Anthropic

{"text":[[{"start":8.42,"text":"A start-up recently valued at $200mn is drawing multibillion-dollar valuation offers with promises of a new generation of tools that are more efficient than products from OpenAI and Anthropic."}],[{"start":20.8,"text":"Founded by former OpenAI researcher Diogo Almeida, TypeSafe AI released its model “Jev” only last week, pitching it as a cheaper alternative to traditional large language models like ChatGPT and Claude for certain tasks."}],[{"start":35.56,"text":"Jev, aimed almost exclusively at software developers, has drawn intense interest and gone viral on social media since it emerged from so-called stealth mode. Its launch video on X generated 40mn views in less than a week."}],[{"start":49.52,"text":"The company has been approached by potential investors with funding offers that would value it at $10bn or more, according to people familiar with the matter."}],[{"start":59.44,"text":"The enthusiasm reflects a growing concern across companies about the cost of deploying AI. TypeSafe is betting that many routine tasks now handed to expensive general-purpose models can instead be performed by cheaper, more specialised systems. If the idea gains traction, it could put pressure on the business models of frontier AI companies such as OpenAI and Anthropic."}],[{"start":82.28,"text":"Almeida told the FT he came up with the idea for Jev while working at OpenAI four years ago, wondering if chatbots were the best tool for all the tasks AI could be asked to perform."}],[{"start":92.84,"text":"“Let’s assume an AI-based economic revolution actually occurred: what percentage of those calls to AI would be for human consumption . . . versus for computer consumption?” he recalled."}],[{"start":103.24,"text":"TypeSafe’s pitch is that LLMs, designed for interacting with human beings, are ill-suited for the sort of programmatic, repetitive and high-volume tasks many global businesses are trying to automate using AI."}],[{"start":null,"text":"

"}],[{"start":115.88,"text":"“ChatGPT is way smarter than me. Yet, for some reason, work in areas where there is an incredible financial incentive to automate is not being automated right now. Because currently AI sucks at it,” Almeida said."}],[{"start":126.58,"text":"Jev does not produce sentences, explanations or images, but instead promises fast, cheap decisions within software applications. The “classifying” tasks it can be used for include, for example, deciding whether to approve, block or review a request, routing a support ticket or underwriting insurance or credit risk."}],[{"start":147.18,"text":"The model is designed for developers to build into the back end of their software, and does not have a consumer-facing interface. One developer plugged it into a joke website, “AskJev”, a tribute to the defunct search engine Ask Jeeves, which went viral on social media."}],[{"start":162.4,"text":"James Hardiman, general partner at DCVC, the Silicon Valley venture firm that led TypeSafe’s most recent funding round, said the start-up was already profitable given the drastically lower computing costs for its product, which makes its model “orders of magnitude” cheaper to use than leading LLMs."}],[{"start":179.6,"text":"He noted the company’s popular AI tool for fundamentally mundane tasks hit the market at a time of spiralling concern about the dangers of the most advanced models."}],[{"start":188.6,"text":"“Literally a week ago everyone was talking about the AI apocalypse . . . In some ways, Jev introduced some sobriety into that conversation.”"}],[{"start":197.32,"text":"Lowering the cost of an AI response is crucial to its appeal. The name Jev is a reference to Jevons paradox, the economic observation that making a resource cheaper or more efficient can ultimately increase its total consumption as new uses emerge."}],[{"start":212.12,"text":"While large language models churn through long chains of reasoning that can consume a lot of computing power, TypeSafe says their model quickly selects from a limited range of possible answers based on a probability calculation."}],[{"start":224.04,"text":"The underlying methods are more similar to older machine learning systems that were deterministic, producing fixed responses to queries and therefore do not hallucinate."}],[{"start":233.2,"text":"This in turn brings down the number of “tokens” it burns, reducing the cost. The company claims each query is about a hundred times cheaper and faster to process than LLMs, charging around 4.2 cents per million tokens compared with LLMs that can cost several dollars per million."}],[{"start":249.32,"text":"Hosting platform Vercel said Jev drew more than twice as much interest from paid developer accounts in its first 24 hours than any previous model launch, eclipsing ones from leading frontier companies OpenAI and Anthropic."}],[{"start":262.76,"text":"OpenRouter, another online gateway for AI models, reported that over the weekend the number of tokens fed to Jev — a measure of how many requests it processed — more than tripled."}],[{"start":272.92,"text":"Andrej Karpathy, an OpenAI co-founder recently hired by Anthropic, wrote on X that Jev appeared to have tapped into “latent demand” for models that provide simple, cheap and fast decisions, in an area that has been “underinvested into because of a race to higher intelligence” by the frontier AI model companies."}],[{"start":290.84,"text":"However, TypeSafe’s secrecy about how it trained its system and similarity to existing tools have led some to question how revolutionary Jev will prove to be."}],[{"start":300.04,"text":"Anastasios Angelopoulos, co-founder and chief executive at Arena, an AI model evaluation platform, said: “It’s unclear to me what makes these models different from standard ‘zero-shot classifiers’, which are relatively well-known technology.”"}],[{"start":313.72,"text":"Meta, Google and Hugging Face already offer classification tools that can categorise material they have not been explicitly trained on. TypeSafe has kept details of how it trained Jev, using open-weight models and computer-generated “synthetic” data, tightly under wraps."}],[{"start":329.2,"text":"The start-up concluded its $40mn seed funding round at a $200mn valuation more than a year ago, Almeida added. He declined to discuss any further funding round or specific new investors but confirmed that potential investors were “battering down our door”."}],[{"start":344.52,"text":"“We are a troop of missionaries, the spark of a revolution . . . so that is something to really think about in growing the company and not losing our way,” he said."}],[{"start":353.16,"text":"Additional reporting by Tim Bradshaw in London"}],[{"start":357.96,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1790316613_1882.mp3"}