Can a 30-person start-up build Britain’s answer to OpenAI? - FT中文网
登录×
电子邮件/用户名
密码
记住我
请输入邮箱和密码进行绑定操作:
请输入手机号码,通过短信验证(目前仅支持中国大陆地区的手机号)。如果该手机号尚未注册,将自动创建 FT中文网账号。短信可能需要几分钟送达,验证码15分钟内有效,请耐心等待:
请您阅读我们的用户注册协议和隐私权保护政策,点击下方按钮即视为您接受。
人工智能

Can a 30-person start-up build Britain’s answer to OpenAI?

London-based Cosine has UK government backing to build a ‘sovereign AI’ model but faces far better resourced rivals
00:00

{"text":[[{"start":7.16,"text":"Britain’s bid to build a “sovereign” AI model rests with a London start-up of about 30 employees that has raised just $15mn."}],[{"start":16.12,"text":"Cosine, founded by chief executive Alistair Pullen and Yang Li, is building Lumen Sovereign, a frontier AI model backed by government support and developed alongside companies including BAE Systems, PwC and Lloyds Banking Group."}],[{"start":31.52,"text":"Officials have described it as “Britain’s first fully sovereign frontier AI model”. Others question whether the company has the talent, computing resources and capital to deliver it."}],[{"start":41.72,"text":"“We do have a track record in this field. We have done real large-scale work on large models and genuinely improved their performance,” Pullen told the FT."}],[{"start":50.68,"text":"The project comes as governments place greater emphasis on AI sovereignty and control over critical technology. Days after the UK said in June it would provide Cosine with access to its Isambard supercomputer, the US temporarily barred foreign groups from accessing Anthropic’s Mythos and Fable models, sharpening concerns over national security and dependence on overseas AI providers."}],[{"start":null,"text":"

Several people work at computer desks in an open-plan office filled with monitors, a whiteboard and various office supplies.
"}],[{"start":74.4,"text":"The obstacles are formidable, however. The company has raised just $15mn, compared with the billions invested in rivals such as France’s Mistral and the far greater sums committed to OpenAI and Anthropic."}],[{"start":87.32,"text":"Even so, Cosine is one of only a handful of UK companies attempting to build frontier AI models from scratch — a field dominated by well-funded US and Chinese groups."}],[{"start":97.92,"text":"Rather than trying to match OpenAI and Anthropic across every application, Pullen said Cosine had built its expertise by adapting and training large language models for customers in highly regulated industries, including financial services and defence, where AI systems are required to run on companies’ own servers."}],[{"start":115.84,"text":"Pullen argues that work gave Cosine the technical foundations to take on Lumen Sovereign. “The kind of work that you need to do to build the model from scratch is very similar to all the work we’ve done in the past, at maybe a couple of orders of magnitude increased size,” he said."}],[{"start":129.96,"text":"The company was founded in 2022 after Pullen concluded that large language models would transform software development. Together with Li, previously an executive at several technology start-ups including collapsed payments group Ziglu and cryptocurrency exchange Luno, he set out to build what they described as an “AI software engineer”."}],[{"start":151.2,"text":"After joining the highly competitive start-up incubator Y Combinator, the pair gained early access to OpenAI’s models and fine-tuned them for software engineering tasks."}],[{"start":null,"text":"
A modern brick building with large windows and a glass entrance, housing Cosine's co-working offices.
"}],[{"start":162.16,"text":"But larger rivals established a lead in autonomous coding tools, most notably Anthropic by releasing Claude Code. The experience forced a rethink. Rather than competing directly with the biggest AI labs, Cosine shifted its focus to building AI systems for regulated industries — work that ultimately laid the foundations for Lumen Sovereign."}],[{"start":184,"text":"Cosine is building its model using a “mixture of experts” framework, an approach popularised by China’s DeepSeek. Its architecture is based on Mistral’s Large 3 model, although Pullen said Lumen Sovereign would be roughly twice the size."}],[{"start":199.08,"text":"The biggest constraint is computing power. Pullen estimates that Isambard will provide about 80 per cent of the computing capacity needed to train the model, with the remainder coming from private UK data centres."}],[{"start":211.08,"text":"Unlike OpenAI and Anthropic, however, Cosine does not plan to operate the model itself. Instead, it intends to make Lumen Sovereign available for others to deploy, avoiding the enormous cost of running AI models at scale."}],[{"start":224.72,"text":"“The tech seems great,” said one venture capitalist familiar with Cosine, but added “they need to raise hundreds of millions of pounds of investment to be remotely competitive, and at the moment they are quite early in that journey.”"}],[{"start":237.28,"text":"Pullen acknowledges the challenge. Cosine hopes to begin sharing “checkpoints” — snapshots of Lumen Sovereign during training — with partners by September before a broader public release in October. The company plans to launch a fundraising round later this year to finance the next stage of development."}],[{"start":254.68,"text":"The money would also fund an aggressive hiring push. Cosine has yet to recruit researchers from the world’s leading AI labs and Pullen concedes it cannot match the pay packages on offer at OpenAI and Anthropic."}],[{"start":266.8,"text":"Instead, he hopes the chance to build a frontier model in a small team will appeal to engineers who have become “pigeonholed” inside larger organisations."}],[{"start":275.24,"text":"Sceptics believe the lack of funding available to UK start-ups will constrain Cosine’s grand ambitions. In response, Pullen said he likes his staff to have a “healthy chip on the shoulder”, telling them: “Look, we can show that we can do it. We can definitely pull this off.”"}],[{"start":291.68,"text":"Additional reporting by Tim Bradshaw in London"}],[{"start":297.64,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1786433291_2329.mp3"}

版权声明:本文版权归FT中文网所有,未经允许任何单位或个人不得转载,复制或以任何其他方式使用本文全部或部分,侵权必究。

美债收益率冲击下,欧洲必须自寻出路

埃里安:欧洲不能只盼着大西洋彼岸的借贷成本回落,它必须着手解决自身问题。

AI与通胀:央行行长该担忧到什么程度?

凯恩斯:这取决于一个价值万亿美元问题的答案:AI是否真能带来预期的生产力提升?

印度企业界的乐观情绪站得住脚吗?

印度企业界公开表达的乐观情绪,与家庭和投资者的焦虑情绪之间的落差日益明显。

Lex专栏:中国药企在药品开发领域获得一席之地

外国制药集团如今开始购买中国本土开发创新药物的授权。

特斯拉与太空探索技术公司:以马斯克为中心的庞大企业星系

埃隆•马斯克旗下上市公司的合并有一定道理,但其他股东最终可能不如现在有影响力。

火箭、机器人和风险:马斯克的3.6万亿美元豪赌

全球首富试图促成迄今最大胆的一笔合并交易,但他还需要说服股东和全球监管机构。
设置字号×
最小
较小
默认
较大
最大
分享×