Lamini A Startup

A Palo Alto startup named Lamini has recently secured a notable investment of $25 million from various investors, among them being Andrew Ng, a renowned professor in computer science at Stanford University. Lamini specializes in developing a platform tailored to assist enterprises in effectively implementing generative AI technology.

Founded several years ago by Sharon Zhou and Greg Diamos, Lamini offers a unique proposition in the realm of AI technology. Zhou and Diamos emphasize that many existing generative AI platforms lack specificity, serving as overly generalized solutions that fail to address the particular needs of corporations. In contrast, Lamini stands out by being purposefully designed with enterprises in mind right from its beginning. The company prioritizes delivering exceptional accuracy and scalability in generative AI, aligning closely with the demands of modern businesses.

“The top priority of nearly every CEO, CIO and CTO is to take advantage of generative AI within their organization with maximal ROI, but while it’s easy to get a working demo on a laptop for an individual developer, the path to production is strewn with failures left and right.” said Lamini’s CEO Sharon Zhou.

Zhou’s observation resonates with the sentiments of numerous companies facing challenges in effectively integrating generative AI into their business operations.

A March survey conducted by MIT Insights underscores this issue, revealing that although 75% of organizations have experimented with generative AI, only a mere 9% have truly adopted it. The obstacles encountered vary widely, ranging from insufficient IT infrastructure and capabilities to governance frameworks, scarcity of skilled personnel, and the considerable costs associated with implementation. Security concerns loom large as well, as indicated by a recent study from Insight Enterprises, where 38% of surveyed companies cited security issues as a significant barrier to harnessing the potential of generative AI technology.

So what is Lamini’s answer to all this?

Zhou emphasizes that Lamini has meticulously fine-tuned every aspect of its technology stack to cater specifically to the demands of enterprise-scale generative AI workloads. This optimization encompasses not only hardware and software but also extends to the engines utilized for tasks such as model orchestration, fine-tuning, running, and training.

While the term “optimized” may seem ambiguous, Lamini is at the forefront of pioneering a groundbreaking technique referred to as “memory tuning,” as highlighted by Zhou. This innovative approach involves training a model on data in a manner that enables it to recall specific details from that dataset accurately.

Read More: Snowflake’s Enterprise-Grade Leap in Generative AI

Zhou suggests that memory tuning holds the potential to mitigate occurrences of “hallucinations” within models, where they fabricate information in response to queries. By incorporating this technique, Lamini aims to enhance the reliability and accuracy of its generative AI solutions, setting a new standard for performance and efficacy in the industry.

Nina Wei, an AI designer at Lamini, said in an email that “Memory tuning is a training paradigm — as efficient as fine-tuning, but goes beyond it — to train a model on proprietary data that includes key facts, numbers, and figures so that the model has high precision and can memorize and recall the exact match of any key information instead of generalizing or hallucinating.”

“Memory tuning” may lack academic recognition, as it appears to be more of a marketing term without accompanying research papers to support its efficacy. However, Lamini will need to demonstrate evidence of its effectiveness compared to other hallucination-reducing techniques.

Thankfully for Lamini, its strengths extend beyond memory tuning. According to Zhou, the platform excels in operating within highly secure environments, including air-gapped ones. Lamini offers companies the flexibility to run, fine-tune, and train models across various configurations, from on-premises data centers to public and private clouds. Additionally, the platform demonstrates impressive scalability, capable of dynamically scaling workloads to over 1,000 GPUs to meet specific application or use case demands, as highlighted by Zhou.

Zhou said, “Incentives are currently misaligned in the market with closed source models, we aim to put control back into the hands of more people, not just a few, starting with enterprises who care most about control and have the most to lose from their proprietary data owned by someone else.”

Team and Investors

Lamini’s co-founders boast impressive credentials within the AI sphere, which likely contributed to attracting investments, including from Andrew Ng. Sharon Zhou, formerly a faculty member at Stanford University, led a research group focusing on generative AI. Before earning her doctorate in computer science under Ng’s mentorship, she served as a machine learning product manager at Google Cloud.

Greg Diamos, on the other hand, co-founded MLCommons, an engineering consortium dedicated to establishing standard benchmarks for AI models and hardware. He also played a pivotal role in developing MLPerf, the benchmarking suite. Diamos led AI research at Baidu, collaborating with Ng during Ng’s tenure as chief scientist there. Additionally, Diamos served as a software architect on Nvidia’s CUDA team.

The extensive industry connections of Lamini’s co-founders have undoubtedly facilitated the company’s fundraising success. Alongside Ng, notable investors include Dylan Field, CEO of Figma; Drew Houston, CEO of Dropbox; Andrej Karpathy, co-founder of OpenAI; and surprisingly, Bernard Arnault, CEO of luxury goods conglomerate LVMH.

In addition to prominent investors like First Round Capital and Amplify Partners, Lamini has garnered early support from AMD Ventures, a somewhat unexpected move given Greg Diamos’ background with Nvidia. AMD’s involvement commenced with supplying Lamini’s data center hardware, and today, Lamini mainly operates its models on AMD Instinct GPUs, deviating from the prevailing industry trend.

Lamini boldly asserts that its model training and performance on AMD Instinct GPUs rival those of Nvidia equivalents, depending on the specific workload. However, as we lack the means to verify this claim independently, we defer to third-party assessments.

Thus far, Lamini has amassed $25 million in funding through seed and Series A rounds, with Amplify leading the Series A. Sharon Zhou explains that these funds will primarily fuel the expansion of Lamini’s 10-person team, bolster its compute infrastructure, and initiate efforts toward deeper technical optimizations.

While Lamini asserts its unique value proposition in the enterprise-oriented generative AI market, it faces competition from established players such as Google, AWS, and Microsoft (through its partnership with OpenAI). In recent months, these tech giants have intensified their efforts to court enterprises, introducing features like streamlined fine-tuning and private fine-tuning on proprietary data.

Zhou remained somewhat tight-lipped when asked about Lamini’s customer base, revenue, and overall market traction, citing the company’s relatively early stage. However, she disclosed that AMD, AngelList, and NordicTrack are among Lamini’s initial paying users. Additionally, some undisclosed government agencies are part of Lamini’s early clientele.

She added, “We’re growing quickly. The number one challenge is serving customers. We’ve only handled inbound demand because we’ve been inundated. Given the interest in generative AI, we’re not representative in the overall tech slowdown — unlike our peers in the hyped AI world, we have gross margins and burn that look more like a regular tech company.”

Amplify general partner Mike Dauber said in one statement, “We believe there’s a massive opportunity for generative AI in enterprises. While there are a number of AI infrastructure companies, Lamini is the first one I’ve seen that is taking the problems of the enterprise seriously and creating a solution that helps enterprises unlock the tremendous value of their private data while satisfying even the most stringent compliance and security requirements.”