乔治斯·多里奥特（Georges Doriot）是风险投资（VC）的奠基人之一，他曾将自己的工作描述为“评判‘一个人和一个想法’”。许多希望从人工智能热潮中赚取巨额财富的硅谷投资者也接受了类似的极简主义理念。一批仅凭模糊野心的人工智能明星从风险投资者那里筹集了数十亿美元，离开了成熟的模型制造商，成立了他们所谓的“新实验室”（neolabs）。到目前为止，这些新实验室大多还没有拿出什么成果。
GEORGES DORIOT, a founding father of venture capital (VC), described his job as judging “a man and an idea”. Many Silicon Valley investors looking to make mega-bucks from the artificial-intelligence boom have embraced a similarly minimalist philosophy. A string of AI superstars with little more than a vague ambition have raised billions from VC investors to decamp from established model-makers and set up their own so-called neolabs. So far most have little to show for it.
新实验室时代始于2024年6月，当时OpenAI的联合创始人伊利亚·苏茨克维（Ilya Sutskever）推出了安全超智能（Safe Superintelligence，SSI）。根据风险投资公司激进创投（Radical Ventures）的统计，截至两年后，共有29家新实验室成立。根据数据提供商PitchBook的数据（见图表），这一批新实验室累计融资已达230亿美元。其中包括Reflection AI，该公司于10月5日宣布的开放权重Beam模型被誉为美国对廉价中国模型的回应，这些中国模型在西方市场迅速占据了份额。
The neolab era began in June 2024 with the launch of Safe Superintelligence (SSI) by Ilya Sutskever, a co-founder of OpenAI. Two years later a total of 29 neolabs had been founded, according to a count by Radical Ventures, a VC firm. This cohort has now raised a combined $23bn, according to figures from PitchBook, a data provider (see chart). Among the list is Reflection AI, whose open-weight Beam model, announced on October 5th, was hailed as America’s answer to the cheap Chinese models that have rapidly gained ground in the West.
然而，其他新实验室在融资方面尚未取得显著进展。根据《经济学人》对公开声明的分析，激进创投（Radical Ventures）队列中不到35%的初创公司推出了商业产品。略高于一半的公司发布了开发者可自行集成到自身产品中的代码，但这些代码通常被宣传为仅用于展示新实验室方法的研究演示。
Other neolabs, however, have not yet achieved much with their funding. Under 35% of the startups in Radical Ventures’ cohort have released commercial offerings, according to The Economist’s analysis of public announcements. A greater share—slightly over half—have released code that developers can incorporate into their own products, but these are often billed as research demos meant only to showcase the approach a neolab is taking.
SSI上一次估值达到320亿美元，可能是这批新实验室中最不急于求成的一家。作为人工智能领域最受殊荣的研究者之一，苏茨克维表示，他的新实验室不会在达到“超智能”之前发布任何产品。（目前尚未达到。）他在7月宣布，SSI终于拥有了“值得扩大规模”的研究，但未透露任何细节。芯片巨头英伟达（Nvidia）曾获得一次内部预览，并向该初创公司追加投资50亿美元，使其总融资额达到80亿美元。
SSI, last valued at $32bn, may be the least hurried of the bunch. Mr Sutskever, one of AI’s most decorated researchers, has said that his neolab does not plan to release a product until it reaches “superintelligence”. (It hasn’t yet.) In July he declared that SSI at last had research “worthy of scaling up”, without revealing any details. Nvidia, a chip colossus, got a peek and handed the startup a further $5bn in funding, bringing its total to $8bn.
苏茨克维在OpenAI的前同事米拉·穆拉蒂，如今在自己创办的初创公司“思考机器实验室”（Thinking Machines Lab，简称TML）中，产品推出速度明显更快。去年12月，在以120亿美元估值成功募得20亿美元资金几个月后，TML发布了名为“Tinker”的服务，该服务能让工程师们对其他机构开发的开源模型进行定制化调整。
Mira Murati, a former colleague of Mr Sutskever’s at OpenAI, has been quicker to ship products at her startup, Thinking Machines Lab (TML). In December, a few months after raising its first $2bn at a $12bn valuation, it released Tinker, a service that allows engineers to customise open-weight models developed by others.
今年7月，TML又推出了自家研发的模型“Inkling”，公司称该模型“在多个领域表现优异”，即便算不上综合实力最强的模型。据一位了解相关数据的人士透露，TML目前的年化营收已突破2亿美元。这一数值虽远超其他同类初创企业，但与据悉该公司新一轮融资所设定的约400亿美元估值相比，仍显得微不足道。
It followed up in July with its own model, Inkling, which it defended as being “strong across many domains”, even if not the best overall. According to a person briefed on the figures, TML’s annualised revenue has crossed $200m. That is much more than other neolabs, but still a trifling figure compared with the roughly $40bn valuation that TML is reportedly targeting in a new fundraising round.
“Inkling”未能入选人工智能交易平台OpenRouter上备受关注的“20款最受欢迎模型”榜单。该榜单上唯一的同类初创企业产品是“Jev”——这是一款成本极低的模型，虽无法生成文本，却擅长充当“任务调度员”，将各类任务分配给其他模型。研发“Jev”的初创企业Typesafe在9月份发布该模型前仅募得4000万美元，但本月又从风险投资机构处获得了8.7亿美元的新资金。
Inkling misses the cut-off for a closely watched list of the 20 most popular models on OpenRouter, an AI marketplace. The only neolab offering on that list is Jev, an ultra-cheap model that does not output text but is skilled as a traffic cop, directing tasks to other models. Typesafe, the neolab behind Jev, raised only a modest $40m before releasing the model in September, but this month brought in an additional $870m from VC investors.
这些初创企业及其投资方坚信，规模更小且目标更明确的团队更有可能取得突破性创新成果。不过他们也面临着诸多挑战，其中之一便是如何赶在模仿者之前推出新产品。例如自“Jev”发布以来，市面上已出现了不少与其类似的竞品。
Neolabs and their backers are betting that smaller and focused teams are more likely to produce novel breakthroughs. But they face various challenges. One is staying ahead of copycat offerings. A number of competitors to Jev, for example, have already surfaced since its release.
另一个挑战是与成熟实验室争夺算力。这促使许多新实验室转向英伟达，后者可以提供其芯片的使用权。据激进风投（Radical Ventures）估计，英伟达已对初创模型制造商进行了32笔投资，超过任何其他公司或风险投资公司。Reflection 已从英伟达筹得8亿美元，据报道正就另一笔大额投资甚至收购进行谈判。英伟达的竞争对手AMD也十分活跃。9月，AMD斥资80亿美元收购了由另一位AI超级明星费飞飞创立的新实验室World Labs。
Another challenge is competing with established labs for computing power. That has pushed many neolabs into the arms of Nvidia, which can grant access to its chips. It has made 32 investments in startup model-makers, reckons Radical Ventures, more than any other company or VC firm. Reflection, which has already raised $800m from Nvidia, is reportedly in talks about another big investment or even an acquisition. AMD, a rival to Nvidia, has been active, too. In September it paid $8bn to purchase World Labs, a neolab founded by Fei-Fei Li, another AI superstar.
OpenAI和Anthropic花了数年时间才将研究转化为商业产品。因此，投资新实验室的人——尤其是那些从事新颖研究方法的人——可能需要有耐心。许多初创模型制造商将会失败。但那些幸存下来的，随着时间的推移，有可能为其早期投资者带来巨额回报。
It took OpenAI and Anthropic years to turn their research into commercial offerings. Investors in neolabs, especially those working on novel research approaches, might therefore need to be patient. Many startup model-makers will fail. But those that don’t could, in time, generate enormous returns for their early backers.
在其他领域，如生物技术，风险投资机构对承担此类风险感到自如。但正如一位投资者指出的，生物技术公司有预定义的里程碑，进展可以据此衡量。如果人体试验的第一阶段进展顺利，初创公司可以期望获得更多融资。新实验室尚未存在足够长的时间以发展出这样的流程，且风险投资机构尚未坚持要求建立这样的流程。如果新实验室时代要继续下去，这些初创公司就需要开始产出更频繁的成果。■ 要追踪塑造商业、工业和技术的趋势，请订阅《底线》——我们的全球商业周刊新闻通讯。
In other fields, such as biotech, VCs are comfortable taking such risks. But as one investor points out, biotechs have pre-defined milestones against which progress can be measured. If the first phase of human trials goes well, a startup can expect to raise more funding. Neolabs have not been around long enough to have developed such a process, and VCs have yet to insist on one. If the neolab era is to continue, the startups will need to start producing more frequent results. ■ To track the trends shaping commerce, industry and technology, sign up to “The Bottom Line”, our weeklyr-only newsletter on global business.