Eric Landau is the CEO & Co-Founding father of Encord, an lively studying platform for laptop imaginative and prescient. Eric was the lead quantitative researcher on a world fairness delta-one desk, placing 1000’s of fashions into manufacturing. Earlier than Encord, he spent practically a decade in high-frequency buying and selling at DRW. He holds an S.M. in Utilized Physics from Harvard College, M.S. in Electrical Engineering, and B.S. in Physics from Stanford College.
In his spare time, Eric enjoys taking part in with ChatGPT and huge language fashions and craft cocktail making.
What impressed you to co-found Encord, and the way did your expertise in particle physics and quantitative finance form your strategy to fixing the “data problem” in AI?
I first began fascinated with machine studying whereas working in particle physics and coping with very massive datasets throughout my time on the Stanford Linear Accelerator Heart (SLAC). I used to be utilizing software program designed for physicists by physicists, which is to say there was loads to be desired when it comes to a pleasing consumer expertise. With simpler instruments, I’d have been in a position to run analyses a lot quicker.
Later, working in quantitative finance at DRW, I used to be liable for creating 1000’s of fashions that had been deployed into manufacturing. Much like my expertise in physics, I discovered that high-quality information was essential in making correct fashions and that managing advanced, large-scale information is tough. Ulrik had the same expertise visualizing massive picture datasets for laptop imaginative and prescient.
After I heard about his preliminary thought for Encord, I used to be instantly on board and understood the significance. Collectively, Ulrik and I noticed an enormous alternative to construct a platform to automate and streamline the AI information growth course of, making it simpler for groups to get the perfect information into fashions and construct reliable AI programs.
Are you able to elaborate on the imaginative and prescient behind Encord and the way it compares to the early days of computing or the web when it comes to potential and challenges?
Encord’s imaginative and prescient is to be the foundational platform that enterprises depend on to rework their information into purposeful AI fashions. We’re the layer between an organization’s information and their AI.
In some ways, AI mirrors earlier paradigm shifts like private computing and the Web in that it’ll turn out to be integral to workflows for each particular person, enterprise, nation, and business. In contrast to earlier technological revolutions, which have been largely bottlenecked by Moore’s regulation of compounded computational development of 30x each 10 years, AI growth has benefited from simultaneous improvements. It’s thus transferring at a a lot quicker tempo. Within the phrases of NVIDIA’s Jensen Huang: “For the very first time, we are seeing compounded exponentials…We are compounding at a million times every ten years. Not a hundred times, not a thousand times, a million times.” With out hyperbole, we’re witnessing the fastest-moving expertise in human historical past.
The potential right here is huge: by automating and scaling the administration of high-quality information for AI, we’re addressing a bottleneck stopping broader AI adoption. The challenges are paying homage to early-day hurdles in earlier technological eras: silos, lack of greatest practices, limitations for non-technical customers, and a scarcity of well-defined abstractions.
Encord Index is positioned as a key instrument for managing and curating AI information. How does it differentiate itself from different information administration platforms presently accessible?
There are just a few ways in which Encord Index stands out:
Index is scalable: Permits customers to handle billions, not hundreds of thousands, of information factors. Different instruments face scalability points for unstructured information and are restricted in consolidating all related information in a corporation.
Index is versatile: Integrates instantly with non-public information storage and cloud storage suppliers equivalent to AWS, GCP, and Azure. In contrast to different instruments which can be restricted to a single cloud supplier or inside storage system, Index is agnostic to the place the information is situated. It permits you to handle information from many sources with acceptable governance and entry controls that enable them to develop safe and compliant AI purposes.
Index is multimodal: Helps multimodal AI, managing information within the type of pictures, movies, audio, textual content, paperwork and extra. Index will not be restricted to a single type of information like many LLM instruments at the moment. Human cognition is multimodal, and we consider multimodal AI can be on the coronary heart of the following wave of AI developments, which can supplant chatbots and LLMs.
In what methods does Encord Index improve the method of choosing the suitable information for AI fashions, and what affect does this have on mannequin efficiency?
Encord Index enhances information choice by automating the curation of huge datasets, serving to groups determine and retain solely probably the most related information whereas eradicating uninformative or biased information. This course of not solely reduces the scale of datasets but in addition considerably improves the standard of the information used for coaching AI fashions. Our prospects have seen as much as a 20% enchancment of their fashions whereas attaining a 35% discount in dataset dimension and saving lots of of 1000’s of {dollars} in compute and human annotation prices.
With the fast integration of cutting-edge applied sciences like Meta’s Phase Something Mannequin, how does Encord keep forward within the fast-evolving AI panorama?
We deliberately constructed the platform to have the ability to adapt to new applied sciences rapidly. We concentrate on offering a scalable, software-first strategy that simply incorporates developments like SAM, making certain that our customers are at all times outfitted with the most recent instruments to remain aggressive.
We plan to remain forward by specializing in multimodal AI. The Encord platform can already handle advanced information varieties equivalent to pictures, movies, and textual content, in order extra developments in multimodal AI come our manner, we’re prepared.
What are the commonest challenges corporations face when managing AI information, and the way does Encord assist handle these?
There are 3 principal challenges corporations face:
- Poor information group and controls: As enterprises put together to implement AI options, they’re usually met with the fact of siloed and unorganized information that isn’t AI-ready. This information usually lacks robust governance round it, limiting a lot of it from being utilized in AI programs.
- Lack of human specialists: As AI fashions sort out more and more advanced issues, there’ll quickly be a scarcity of human area specialists to organize and validate information. As an organization’s AI calls for improve, scaling that human workforce is difficult and expensive.
- Unscalable tooling: Performant AI fashions are very data-hungry when it comes to information wanted for fine-tuning, validation, RAG, and different workflows. The earlier era of instruments will not be outfitted to handle the quantity of information and forms of information required for at the moment’s production-grade fashions.
Encord fixes these issues by automating the method of curating information at scale, making it straightforward to determine impactful information from problematic information and making certain the creation of efficient coaching and validation datasets. It makes use of a software-first strategy that’s straightforward to scale up or down as information administration wants change. Our AI-assisted annotation instruments empower human-in-the-loop area specialists to maximise workflow effectivity. This course of is especially essential in industries equivalent to monetary providers and healthcare, the place AI trainers are expensive. We make it straightforward to handle and perceive all of a corporation’s unstructured information, lowering the necessity for guide labor.
How does Encord sort out the problem of information bias and under-represented areas inside datasets to make sure truthful and balanced AI fashions?
Tackling information bias is a essential focus for us at Encord. Our platform robotically identifies and surfaces areas the place information could be biased, permitting AI groups to deal with these points earlier than they affect mannequin efficiency. We additionally be certain that under-represented areas inside datasets are correctly included, which helps in creating fairer and extra balanced AI fashions. By utilizing our curation instruments, groups might be assured that their fashions are educated on numerous and consultant information.
Encord lately secured $30 million in Collection B funding. How will this funding speed up your product roadmap and enlargement plans?
The $30 million in Collection B funding can be used to drastically improve the scale of our product, engineering, and AI analysis groups over the following six months and speed up the event of Encord Index and different new options. We’re additionally increasing our presence in San Francisco with a brand new workplace, and this funding will assist us scale our operations to assist our rising buyer base.
Because the youngest AI firm from Y Combinator to boost a Collection B, what do you attribute to Encord’s fast development and success?
One of many causes we have now been in a position to develop rapidly is that we have now adopted an especially customer-centric focus in all areas of the corporate. We’re continually speaking with prospects, listening intently to their issues, and “bear hugging” them to get to options. By hyper-focusing on buyer wants slightly than hype, we’ve created a platform that resonates with high AI groups throughout varied industries. Our prospects have been instrumental in getting us to the place we’re at the moment. Our potential to scale rapidly and successfully handle the complexity of AI information has made us a beautiful answer for enterprises.
We additionally owe a lot of our success to our teammates, companions, and buyers, who’ve all labored tirelessly to champion Encord. Working with world-class product, engineering, and go-to-market groups has been enormously impactful in our development.
Given the growing significance of information in AI, how do you see the function of AI information platforms like Encord evolving within the subsequent 5 years?
As AI purposes develop in complexity, the necessity for environment friendly and scalable information administration options will solely improve. I consider that each enterprise will finally have an AI division, very like how IT departments exist at the moment. Encord would be the solely platform they should handle the huge quantities of information required for AI and get fashions to manufacturing rapidly.
Thanks for the good interview, readers who want to study extra ought to go to Encord.
