Lamini, a Palo Alto-based startup building a platform to help companies deploy generative AI technology, has raised $25 million from investors including Stanford computer science professor Andrew Ng.
The mineco-founded several years ago by Sharon Zhou and Greg Diamos, presents an interesting sales pitch.
Many generative AI platforms are far too general-purpose, Zhou and Diamos say, and lack the right solutions and infrastructure to meet business needs. In contrast, Lamini was designed from the ground up with businesses in mind and is focused on delivering high accuracy and scalability of generative AI.
“The top priority of almost every CEO, CIO and CTO is to leverage generative AI within their organization with maximum ROI,” Zhou, CEO of Lamini, told TechCrunch. “But while it’s easy to get a working demo on a laptop for an individual developer, the path to production is littered with failures left and right.
According to Zhou, many companies have expressed frustration with the obstacles faced in meaningfully adopting generative AI in their business functions.
According to a month of March survey According to MIT Insights, only 9% of organizations have widely adopted generative AI, while 75% of them have experimented with it. The main obstacles range from lack of IT infrastructure and capacity to poor governance structures, insufficient skills and high implementation costs. Security is also a major factor – in a recent investigation According to Insight Enterprises, 38% of businesses said security impacts their ability to leverage generative AI technology.
So what is Lamini’s answer?
Zhou says “every piece” of Lamini’s technology stack has been optimized for enterprise-wide generative AI workloads, from hardware to software, including the engines used to support the orchestration, fine-tuning, execution and training of models. “Optimized” is a vague word, sure, but Lamini is pioneering a step that Zhou calls “memory tuning,” which is a technique for training a model on data such that it recalls certain parts exactly. of this data.
Memory tuning can potentially reduce hallucinationssays Zhou, or cases where a model invents facts in response to a request.
“Memory tuning is a training paradigm – as effective as fine tuning, but goes beyond – for training a model on proprietary data that includes key facts, figures and figures so that the model has a high precision”, Nina Wei, AI designer. at Lamini, told me via email, “and can memorize and recall the exact correspondence of any key information instead of generalizing or hallucinating.”
I’m not sure I’d buy this. “Memory tuning” seems to be more of a marketing term than an academic one; There is no research paper on this – at least none that I have managed to find. I’ll let Lamini demonstrate that his “memory tuning” is better than other hallucination reduction techniques that are/have been attempted.
Fortunately for Lamini, memory tuning isn’t its only differentiator.
Zhou says the platform can operate in highly secure environments, including isolated environments. Lamini allows businesses to run, refine and train models on a range of configurations, from on-premises data centers to public and private clouds. And it scales workloads “elastically,” scaling up to more than 1,000 GPUs if the application or use case demands it, Zhou says.
“Incentives are currently misaligned in the market with closed-source models,” Zhou said. “We aim to Put control back in the hands of more people, not just a few, starting with the companies that care most about control and have the most to lose from their proprietary data owned by anyone another one.
Lamini’s co-founders are, for what it’s worth, quite accomplished in the field of AI. They also worked separately with Ng, which undoubtedly explains his investment.
Zhou was previously a professor at Stanford, where she led a research group on generative AI. Before earning her PhD in computer science under Ng, she was a machine learning product manager at Google Cloud.
Diamos, for his part, co-founded MLCommons, the engineering consortium dedicated to creating standard benchmarks for AI models and hardware, as well as the MLCommons benchmarking suite, MLPerf. He also led AI research at Baidu, where he worked with Ng when the latter was chief scientist there. Diamos was also a software architect on the Nvidia project. CUDA team.
The co-founders’ industry connections appear to have given Lamini a fundraising advantage. Besides Ng, Dylan Field, CEO of Figma, Drew Houston, CEO of Dropbox, Andrej Karpathy, co-founder of OpenAI and, interestingly, Bernard Arnault, CEO of luxury giant LVMH, have all invested in Lamini.
AMD Ventures is also an investor (a bit ironic given Diamos’ Nvidia roots), as are First Round Capital and Amplify Partners. AMD got involved early, supplying Lamini with data center hardware, and today Lamini manages several of his models on AMD Instinct GPUs, going against the grain industry trend.
Lamini loudly claims that its model training and running performance is comparable to equivalent Nvidia GPUs, depending on workload. Since we are not equipped to test this claim, we will leave it to third parties.
To date, Lamini has raised $25 million in seed and Series A rounds (Amplify led the Series A). Zhou says the money is intended to triple the company’s 10-person team, expand its IT infrastructure and kickstart development toward “deeper technical optimizations.”
There are a number of enterprise-oriented generative AI vendors that could compete with aspects of Lamini’s platform, including tech giants like Google, AWS, and Microsoft (via its OpenAI partnership). Google, AWS, and OpenAI, in particular, have been aggressively courting the company in recent months, introducing features like streamlined fine-tuning, private fine-tuning of private data, and more.
I asked Zhou about Lamini’s customers, revenue and overall go-to-market dynamics. She wasn’t willing to reveal much at this somewhat early stage, but did say that AMD (via the AMD Ventures tie-up), AngelList and NordicTrack were among Lamini’s early (paid) users, as well as several undisclosed government agencies.
“We are growing quickly,” she added. “The number one challenge is serving customers. We only processed incoming request because we were inundated. Given the focus on generative AI, we are not representative of the overall technology slowdown – unlike our peers in the hot AI world, our gross margins and business are more corporate-like classic technology.
Mike Dauber, General Partner at Amplify, said: “We believe there is a huge opportunity for generative AI in enterprise. While there are a number of AI infrastructure companies out there, Lamini is the first I’ve seen to take the business problems seriously and create a solution that helps businesses unlock the enormous value of their private data while meeting even the most stringent compliance requirements. and security requirements.
