Lamini, a startup headquartered in Palo Alto, is developing a platform designed to facilitate the implementation of generative AI technology for businesses. The company has successfully secured $25 million in funding, with contributions from notable investors such as Andrew Ng, a professor of computer science at Stanford.
The founders of Lamini, Sharon Zhou and Greg Diamos, present a compelling proposition.
According to Zhou and Diamos, the majority of generative AI platforms on the market are overly broad and lack the specialized solutions and infrastructure that corporations require. Lamini, however, has been purpose-built with the specific needs of enterprises in mind, aiming to provide highly accurate and scalable generative AI services.
Zhou, serving as the CEO of Lamini, conveyed to TechCrunch, “For most executives, including CEOs, CIOs, and CTOs, leveraging generative AI to its fullest potential within their organizations to achieve the highest return on investment is a key goal. However, despite the ease of creating a functional prototype on a developer’s laptop, navigating the journey to a fully operational state is fraught with numerous obstacles.”
Echoing Zhou’s observations, numerous firms have voiced their challenges in fully integrating generative AI into their operational processes.
A survey conducted by MIT Insights in March revealed that a mere 9% of companies have fully implemented generative AI, although 75% have explored its potential. The obstacles are diverse, ranging from inadequate IT infrastructure and expertise to weak governance frameworks, a shortage of necessary skills, and steep costs of deployment. Security concerns are also significant, with 38% of businesses in a survey by Insight Enterprises citing security issues as a barrier to adopting generative AI technologies.
Lamini’s response to these challenges is comprehensive optimization of its technology stack for large-scale generative AI applications, encompassing everything from hardware to software, and including the mechanisms that facilitate model orchestration, refinement, operation, and training. While “optimized” might seem like an ambiguous term, Lamini is advancing a concept Zhou refers to as “memory tuning.” This approach involves training a model on specific data so that it can accurately recall certain parts of that data.
Zhou suggests that memory tuning could help mitigate the issue of “hallucinations,” or instances where a model fabricates information in response to queries.
Nina Wei, an AI designer at Lamini, explained via email, “Memory tuning represents a training approach that not only matches the efficiency of fine-tuning but also surpasses it. It involves training a model on unique data that encompasses crucial facts, figures, and statistics, ensuring the model retains high accuracy and can precisely remember and reproduce exact pieces of vital information, rather than making generalizations or fabrications.”
I remain skeptical about the concept of “memory tuning.” It seems to be coined more for promotional appeal rather than stemming from scholarly research; at least, my searches haven’t yielded any academic publications on the subject. It’s up to Lamini to provide proof that “memory tuning” surpasses other existing methods aimed at reducing AI errors.
Nonetheless, Lamini has more to offer than just “memory tuning.”
According to Zhou, Lamini’s platform is capable of functioning in extremely secure settings, even those without network connectivity. The company enables organizations to operate, refine, and educate models across various setups, including local data centers and both public and private cloud services. Moreover, it boasts the ability to dynamically adjust computing resources, scaling up to over 1,000 GPUs to meet the demands of specific applications or use cases, as per Zhou’s statement.
Zhou also critiques the current market dynamics, which favor proprietary models, stating, “Our goal is to democratize control, distributing it among a wider audience rather than a select few. This begins with enterprises that have the most at stake regarding the ownership of their confidential data.”
The founders of Lamini are well-regarded figures in the field of AI. Their individual interactions with Ng likely played a role in his decision to invest in their venture.
Before her doctoral studies under Ng’s mentorship, Zhou served on the Stanford faculty, leading a team focused on generative AI research. Her career also includes a stint as a product manager for machine learning at Google Cloud.
Diamos has an impressive track record, having co-founded MLCommons, an engineering group committed to establishing AI model and hardware benchmarks, and creating the MLPerf benchmarking suite. His AI research leadership at Baidu coincided with Ng’s tenure as chief scientist. Additionally, Diamos contributed as a software architect to Nvidia’s CUDA team.
The extensive network of the co-founders seems to have propelled Lamini’s fundraising success. Notable investors include Ng, Dylan Field of Figma, Drew Houston of Dropbox, OpenAI’s Andrej Karpathy, and somewhat unexpectedly, Bernard Arnault of LVMH.
Despite Diamos’ history with Nvidia, AMD Ventures has invested in Lamini, alongside First Round Capital and Amplify Partners. AMD’s early involvement provided Lamini with data center equipment, and presently, Lamini operates many of its models on AMD Instinct GPUs, diverging from common industry practices.
Lamini asserts that its models’ training and operational performance match those of Nvidia’s GPUs, depending on the specific tasks. However, without the means to verify this claim ourselves, we’ll defer to independent evaluations.
As of now, Lamini has successfully secured $25 million through initial funding and a Series A round, with Amplify at the helm of the latter. The capital is earmarked for tripling the startup’s workforce, which currently stands at ten, enhancing its computational resources, and delving into more advanced technical refinements.
Lamini is entering a competitive field with several enterprise-focused generative AI providers, including industry behemoths such as Google, AWS, and Microsoft through its collaboration with OpenAI. These companies have been proactively targeting the enterprise sector, rolling out new functionalities like more efficient model fine-tuning and secure fine-tuning on confidential data.
When inquired about Lamini’s client base, financials, and market traction, Zhou remained reserved, disclosing only limited information at this stage. However, she did mention that AMD (through AMD Ventures), AngelList, and NordicTrack are some of the early adopters who are paying customers, in addition to a few unnamed government entities.
Zhou emphasized the rapid growth of the company, stating, “Our primary focus is on customer service. The overwhelming inbound interest has meant we’ve only been responding to direct inquiries. The surge in generative AI interest means we’re not experiencing the general tech downturn. Our financial health and expenditure are more aligned with traditional tech businesses.”
Mike Dauber, a general partner at Amplify, expressed his conviction in the potential of generative AI within the enterprise sector. “There’s a plethora of AI infrastructure firms out there, but Lamini stands out as the first I’ve encountered that genuinely addresses enterprise challenges head-on. They’re crafting a solution that enables businesses to fully leverage the vast potential of their proprietary data while adhering to the strictest compliance and security standards,” he commented.