内容来源：https://techcrunch.com/2025/10/20/top-openai-google-brain-researchers-set-off-a-300m-vc-frenzy-for-their-startup-periodic-labs/

内容总结：

由前OpenAI顶尖研究员利亚姆·费杜斯与前谷歌大脑团队成员埃金·多乌斯·丘布克联合创立的科技公司Periodic Labs，上月正式结束隐匿模式并宣布获得3亿美元种子轮融资，该轮由风险投资机构Felicis领投，多家顶级风投及天使投资人参与。

七个月前，费杜斯与丘布克在硅谷一次对话中决定将"生成式AI颠覆科研"的构想付诸实践。丘布克指出三大技术条件已然成熟：实现粉末合成的机械臂趋于可靠，机器学习模拟能精准复现复杂物理系统，而大语言模型更展现出强大的逻辑推理能力——后者正是费杜斯在OpenAI带领ChatGPT后期训练团队取得的关键突破。

这一技术闭环已初现雏形：通过模拟计算发现新化合物，机械臂执行材料合成，再由大语言模型分析实验结果并动态调整研究方向。事实上，丘布克参与的谷歌2023年突破性研究已验证该路径——其全自动实验室曾根据语言模型建议成功研制出41种新型化合物。

融资过程颇具戏剧性。费杜斯宣布离职后引发投资机构激烈竞逐，甚至有投资人撰写"情书"示好。最终其前OpenAI同事、现任Felicis投资人的彼得·邓在旧金山努埃瓦山谷的徒步交流中当场承诺投资。邓表示："当费杜斯指出'真正做科学就要配备实体实验室'时，这个构想瞬间征服了我。"

目前该公司已组建超过20人的顶尖团队，成员涵盖AI研究与材料科学等多个领域，每周举行跨学科研讨会以促进深度协作。实验室现已投入运行，正专注于新型超导材料的研发探索。丘布克透露，虽然机械臂系统尚需训练周期，但即使实验失败也将成为珍贵的训练数据，这种"重视探索过程而非仅追求成果"的模式可能重塑科研激励机制。

值得注意的是，OpenAI并未参与本轮投资。与此同时，OpenAI副总裁凯文·威尔上月宣布成立科学部门，预示着AI驱动科研赛道的竞争正在升温。

中文翻译：

由OpenAI顶尖研究员利亚姆·费德斯与前谷歌大脑同事埃金·多乌斯·丘布克共同创立的新兴企业Periodic Labs，上月结束隐匿模式时宣布获得3亿美元巨额种子轮融资。本轮由Felicis领投，众多知名天使投资人及顶级风投机构参与其中。

七个月前，费德斯与丘布克（朋友昵称“Doge”）的一次对话催生了这家初创公司。作为谷歌大脑在机器学习与材料科学领域的核心研究员，丘布克亲历了硅谷对生成式AI将彻底改变科学研究的反复探讨。他们判断，实现这一愿景的要素已然齐备——至少足以创立一家勇于尝试的企业。

“大语言模型领域、实验科学和模拟技术近年取得的关键突破，让当下成为最佳时机。”丘布克向TechCrunch坦言。

他列举了三重契机：其一，能够处理粉末合成（即混合与创造新材料的过程）的机械臂近期被证实具备稳定可靠性；其二，机器学习模拟现已能高效精准地构建复杂物理系统模型，满足新材料研发需求；其三，大语言模型展现出强大的推理能力——这部分得益于费德斯及其OpenAI团队的贡献。费德斯不仅是ChatGPT创始团队核心成员，更执掌着OpenAI至关重要的训练后优化部门。

将这些要素串联起来，蓝图逐渐清晰：模拟系统可从理论层面探索新化合物，机器人执行材料混合工序，大语言模型则分析实验结果并实时调整研发方向。人工智能驱动的自动化材料科学已蓄势待发。

事实上，丘布克正是2023年发表突破性论文的研究团队成员之一，该论文记录了谷歌此前开展的先驱性科研项目。该团队通过语言模型生成的配方，在全自动机器人实验室中成功研制出41种新型化合物。

双重机遇：数据驱动科研变革
同样重要的是，两位创始人意识到失败实验对新公司同样具有价值，因为数据是人工智能的命脉。AI科研将为现实世界的训练数据与训练后数据开辟全新来源。他们认为这或将颠覆现有以论文发表和科研经费为导向、追求成功而非探索的科学激励机制。

“让AI接触现实世界，将实验纳入闭环——我们相信这是下一个前沿阵地。”费德斯向TechCrunch透露。

Felicis赢得投资机会 OpenAI未参与
与丘布克商谈后，费德斯向OpenAI管理层提交辞呈并阐述创业计划。随后他在推特上宣告离职，言辞间暗示获得OpenAI的祝福与投资。

但OpenAI最终并未注资。创始人向TechCrunch确认，OpenAI并非Periodic Labs的投资方。尽管费德斯未说明缘由，实际上他们也无需OpenAI的资金支持。

费德斯的推文引发风投界竞相追逐。“几乎像是被投资人反向路演。有位投资者甚至给公司写了情书。”费德斯笑称，他与丘布克对此都“不知所措”。其他机构则寄来厚达数页的自荐文件。

他们接听的第一个电话来自前OpenAI同事、现顶级种子基金Felicis投资人彼得·邓（邓于2025年初加入Felicis）。“利亚姆在OpenAI地位举足轻重，深受爱戴且极具影响力，”邓向TechCrunch表示，“听闻他离职，我立即联系了他。”

二人在旧金山诺伊谷相约咖啡。在咖啡因与创业激情的催化下，费德斯邀请邓在闻名遐迩的丘陵地带边漫步边深谈。虽然步行路演堪称硅谷经典场景，但这次却真实上演。

微寒的天气逐渐转热。穿着毛衣的邓汗流浃背地试图跟上体格健壮的费德斯，直到创始人说出“让我瞬间止步”的话。“他说‘人人口中都在谈论科研，但真正做科研必须亲身实践’。”邓回忆道。

这意味着需要为AI配备全套湿实验室，在受控真实环境中验证其构想。“这些模型的局限在于其认知始终处于正态分布范围内。我们输入大量数据，它只能复现已知内容。”邓解释道。而探索新知必须包含假设验证环节。

“我当即在诺伊谷的山丘间承诺签下支票。”邓坦言。费德斯也记得邓询问如何参与时，自己提出需要购置笔记本电脑和临时办公室资金，对方立刻回应：“太好了，我现在就打款。这无疑是强烈的信心背书。”

当然邓并未当真在街头开支票。他兴奋地回到公司，却被Felicis律师提醒无法立即签约：这家初创公司尚未注册，连名称都未确定，更遑论收款账户。“这就是我们有多早期。”邓笑道。

很快公司完成所有设立程序，并收到纷至沓来的投资意向书。凭借3亿美元资金储备，丘布克和费德斯组建了逾二十人的顶尖团队，囊括Alexandre Passos（o1与o3模型创造者）、Eric Toberer（已取得关键超导发现成果的材料科学家）及微软两代GenAI材料科学工具开发者Matt Horton等杰出人才。

由于团队成员分属人工智能到物理学等不同领域，每周会由一位专家开展研究生级别的专题讲座。“我们坚信深度协同至关重要。”丘布克表示，他希望全员透彻理解正在构建的每个环节。

Periodic Labs现已建立实验室，正在处理实验数据、运行模拟系统并测试部分预测结果。初期核心使命是探索新型超导材料——这可能是价值连城的发现。改进超导材料有望开启高能效、低能耗的技术新纪元。

不过最终环节的机器人系统尚未投入运行。“它们需要时间训练调试。”丘布克透露。

这无疑是场豪赌。无论是否借助人工智能，科学发现从来不易速成且难以预测。尽管专家团队持有若干积极迹象，但无人能保证他们必定找到目标成果——或在探索途中获得意外发现（甚至仅从失败实验中积累宝贵数据）。

值得注意的是，模型开发者们也在逐步涉足AI科研领域。上月OpenAI副总裁凯文·威尔宣布成立“OpenAI科研部门”，旨在“构建下一代伟大科学工具：加速科学发现的人工智能平台”。

至于那位撰写情书的投资人，最终未能达成交易（尽管费德斯承认书信“令人倍感荣幸”）。其余种子轮投资者包括安德森·霍洛维茨基金、DST、英伟达风投部门NVentures、Accel，以及杰夫·贝索斯、埃拉德·吉尔、埃里克·施密特和杰夫·迪恩等天使投资人。

埃拉德·吉尔将于10月29日在旧金山Disrupt大会上发表演讲，探讨人工智能如何重塑创业生态。

英文来源：

Periodic Labs, a new startup by one of OpenAI’s most respected researchers, Liam Fedus, and his former Google Brain colleague, Ekin Dogus Cubuk, came out of stealth last month with an enormous $300 million seed round. It was led by Felicis and included a who’s who of angels and other top VCs.
The startup began when Fedus had a conversation with Cubuk (whose friends call him “Doge”) about seven months ago. Cubuk was one of Google Brain’s foremost machine learning and material science researchers. After endless Silicon Valley takes on how generative AI would radically change scientific discovery, they decided that the pieces were finally in place to make this a reality. Or at least to found a startup that attempted it.
“There are a few things that happened in the LLM field, in experimental science and in simulations that kind of made this the right time,” Cubuk told TechCrunch.
For one, he said, robotic arms that could handle powder synthesis — the process of mixing and creating new materials — had recently proved themselves reliable. For another, machine learning simulations had become efficient and accurate enough to model complex physical systems such as those needed to develop new materials.
And, third, LLMs now had powerful reasoning capabilities — in part through the work of Fedus and his team at OpenAI. Fedus was one of the small team that created ChatGPT to begin with and was running OpenAI’s uber-important post-training team, which refines models after their initial development.
Stitched together, the picture was clear: A simulation could theoretically discover new compounds, a robot could mix the materials, and an LLM could analyze the results and suggest course corrections. AI-automated material science was ready to be built.
In fact, Cubuk was one of the researchers who published a groundbreaking paper in 2023 documenting a precursor Google research project. The team built a fully automated, robotic-powered lab and created 41 novel compounds from recipes suggested by language models.
2-FOR-1 DISCOUNT: Bring a +1 and save 60%
Google Cloud, Netflix, Microsoft, Box, Phia, a16z, ElevenLabs, Wayve, Hugging Face, Elad Gil, Vinod Khosla — some of the 250+ heavy hitters leading 200+ sessions designed to deliver the insights that fuel startup growth and sharpen your edge. And don’t miss 300+ showcasing startups in all sectors.
Bring a +1 and save 60% on their pass, or get your pass by Oct 27 to save up to $444.
2-FOR-1 DISCOUNT: Bring a +1 and save 60%
Google Cloud, Netflix, Microsoft, Box, Phia, a16z, ElevenLabs, Wayve, Hugging Face, Elad Gil, Vinod Khosla — some of the 250+ heavy hitters leading 200+ sessions designed to deliver the insights that fuel startup growth and sharpen your edge. And don’t miss 300+ showcasing startups in all sectors. Bring a +1 and save 60% on their pass, or get your pass by Oct 27 to save up to $444.
Equally importantly, the founders realized that even failed experiments would be valuable for their new startup because data is the lifeblood of AI. AI science offered an entirely new source for real-world training and post-training data. This could, the founders believe, turn the existing scientific motivation system on its head, which seeks success, not exploration, rewarded via paper publication and grants.
“Making contact with reality, bringing experiments into the [AI] loop — we feel like this is the next frontier,” Fedus told TechCrunch.
Felicis wins the deal; OpenAI does not invest
After that discussion with Cubuk, Fedus went to the powers that be at OpenAI to share his resignation and his plan. He then gleefully tweeted to the world that he was leaving with what appeared to be OpenAI’s blessing and investment.
That investment didn’t actually materialize, however. OpenAI is not a backer of Periodic, the founders confirmed to TechCrunch. And while Fedus declined to say why, they actually didn’t need OpenAI’s money.
Fedus’ tweet set off a frenzy of VCs courting the company. “There was almost like a feeling of being reverse pitched. One investor actually wrote a love letter to Periodic Labs,” Fedus laughed, explaining that neither he nor Cubuk “knew what to make of it.” Others sent multi-page documents pitching themselves.
But the first call that they actually took was from Peter Deng, a former OpenAI colleague turned investor for top-tier seed firm Felicis. (Deng left OpenAI for Felicis at the start of 2025.)
“Liam is a very big deal within OpenAI, very well loved and an extremely impactful researcher,” Deng told TechCrunch. “When I heard he left, I texted him immediately.”
Deng met Fedus for coffee in the Noe Valley neighborhood of San Francisco. Hyped on caffeine and enthusiasm, Fedus invited Deng to finish their conversation on a walk over the area’s famously hilly terrain. Pitch walks may be a Silicon Valley trope, but they also really happen.
The chilly day had turned hot. Deng, wearing a sweater, sweated and scrambled to keep up with the fit and friendly Fedus until the founder said something that “literally stopped me in my tracks,” Deng told TechCrunch. He told Deng that “everyone talks about doing science, but in order to do science, you actually have to do science,” Deng recalls.
In other words, they needed to give AI a fully equipped wet lab to try its ideas in a real-world, controlled setting.
“The truth about these models is that everything that the models know is within normal distribution. We take a bunch of data, and it can just regurgitate what it knows,” Deng said.
Discovering something new has to involve testing hypotheses.
“And I committed on the spot, in the middle of the hills of Noe Valley, to write the check,” Deng says.
Fedus also remembers the moment Deng asked how he could be involved, and Fedus told him the startup needed cash for laptops and a temporary office. And “he’s like, great, I’ll give you money right now. And it was just this huge vote of confidence.”
But Deng didn’t actually whip out his checkbook on the street. He went back to the office elated over the deal only to encounter Felicis’ lawyer, who pointed out that the firm couldn’t promptly sign a contract: The company wasn’t incorporated yet. It didn’t even have a name, much less a bank account to wire funds. “That’s how early we were.” Deng grinned.
Soon they had all of those things and all the term sheets they could handle. With the $300 million war chest, Cubuk and Fedus hired over two dozen of the most prestigious AI and scientific talent like Alexandre Passos (a creator of o1 and o3); Eric Toberer (a materials scientist who has already made key superconductor discoveries); and Matt Horton, a creator of two of Microsoft’s GenAI materials science tools. And the list goes on.
Because the team members are all experts in different areas, from AI to physics, each week one of them gives a grad-level lecture to the others. “We do feel like a tight coupling is extremely important,” Cubuk said. He wants everyone to understand all parts of what they are building.
Periodic Labs has already set up its lab, too, and is working with experimental data, simulations and testing some predictions. The main initial mission is to find new superconductor materials — potentially a gold mine discovery. Improved superconductors could power the next era of potent, but lower energy-consuming tech.
But the last part — the robots — are not yet up and running. “They will take a bit to train,” Cubuk said.
All of this is, of course, a big swing for the fences. AI powered or not, scientific discovery is not typically fast, easy, or predictable. While this team of experts has some indications that they will find what they are looking for — or make other discoveries along the way (or simply generate valuable data on their failures), there’s no guarantees.
And we know that model makers themselves are inching their way toward more AI science. Last month, OpenAI VP Kevin Weil said he was launching an OpenAI for Science unit at the company to “build the next great scientific instrument: an AI-powered platform that accelerates scientific discovery.”
As for the investor who wrote the love letter, he didn’t win the deal (although Fedus did say that the letter was “very flattering.”) The other seed investors include Andreessen Horowitz, DST, Nvidia’s venture capital arm NVentures, Accel, and angel backers like Jeff Bezos, Elad Gil, Eric Schmidt, and Jeff Dean.
Elad Gil will be speaking about how AI has changed the startup landscape at Disrupt in San Francisco on October 29.