Adept aims to build artificial intelligence that can automate any software process – TechCrunch

In 2016 at TechCrunch Disrupt New York, several of the original developers behind what became Siri unveiled Viv, an artificial intelligence platform that promised to connect many third-party apps to perform almost any task. The pitch was bewildering – but not fully realized. Samsung later acquired Viv, which folded a shortened version of the technology into its Bixby voice assistant.

Six years later, a new team claims to have cracked the code into a universal AI assistant — or at least got a little closer. In the lab of a product called Adept emerged out of disguise today with $65 million in funding, they—in the words of the founders—”[ing] The general intelligence that enables humans and computers to work together creatively to solve problems.”

They are noble things. But Adept’s founders, CEO David Luan, CTO Niki Parmar and chief scientist Ashish Vaswani, cut short their ambition of perfecting an “overlay” inside computers running using the same tools people use. This overlay will be able to respond to commands like “create a monthly compliance report” or “draw the tray between these two points on this chart,” assures Adept, all using existing software like Airtable, Photoshop, Tableau, and Twilio to get the job done.

“[W]We are training a neural network to use every software tool in the world, based on the massive amount of existing capabilities that people have already created,” Luan told TechCrunch in an email interview.[W]With Handyman, you will be able to focus on the work you enjoy most and ask us [system] To take on other tasks… We expect the collaborator to be a good and highly communicative student, becoming more helpful and attuned to every human interaction. “

From Luan’s description, what Adept makes looks a bit like robotic process automation (RPA), or software bots that take advantage of a combination of automation, computer vision, and machine learning to automate repetitive tasks like filling out forms and answering emails. But the team insists that their technology is far more sophisticated than what RPA vendors like Automation Anywhere and UiPath offer today.

“We are building a general system that helps people get things done in front of their computers: a global artificial intelligence collaborator for every knowledge worker… We are training a neural network to use every software tool in the world, based on the massive amount of existing capabilities that people have already created, Luan said. “We believe that the ability of AI to read and write text will continue to be valuable, but the ability to do things on a computer will be significantly more valuable to organizations… [M]Odds who are trained on text can write great prose, but they can’t take action in the digital world. You can’t ask [them] To book a flight for you, cut a check for the seller, or conduct a science experiment. True general intelligence requires models that can not only read and write, but act when people ask them to do something.”

Maher is not the only one exploring this idea. In a February research paper, scientists at Alphabet-backed DeepMind describe what they call a “data-driven” approach to teaching AI to control computers. by take Artificial intelligence monitors keyboard and mouse commands from people who complete computer tasks to “follow instructions”, such as booking a flight, and scientists have been able to show how accurately more than a hundred tasks are performed “at the human level” of the system.

It is no coincidence, Mustafa Suleiman, co-founder of DeepMind recently collaborated He partnered with LinkedIn founder Reid Hoffman to launch Inflection AI, which – like Adept – aims to use artificial intelligence to help humans work more efficiently with computers.

Adept’s phenotypic differentiation is a mental trust of AI researchers drawn from DeepMind, Google, and OpenAI. Vaswani and Parmar helped pioneer Transformer, an artificial intelligence architecture that has received a lot of attention over the past several years. Dating back to 2017, Transformer has become the architecture design of choice for natural language tasks, demonstrating the ability to summarize documents, translate between languages, and even classify images and analyze biological sequences.

Among other products, the language-generating GPT-3 from OpenAI was developed using Transformer technology.

“Over the next few years, everyone piled on the switch, using it to solve many decades-old problems in quick succession. When I led the engineering at OpenAI, we scaled the switch to GPT-2 (the predecessor of GPT-3) and GPT-3, Luan said. Google’s efforts to expand the range of adapter models have resulted in [the AI architecture] Bert, run a google search. Several teams, including our founding team members, have trained compilers that can write code. DeepMind even showed that transformers work for protein folding (AlphaFold) and Starcraft (AlphaStar). Transformers have made general intelligence tangible in our field.”

At Google, Luan was the overall technical lead for what he called a “big models effort” at Google Brain, one of the tech giant’s prominent research divisions. There, he trained larger and larger adapters with the goal of building a single generic model to run all machine learning use cases, but his team encountered obvious limitations. The best results were limited to models designed to excel in specific areas, such as analyzing medical records or answering questions about certain topics.

“From the beginning of the field, we have wanted to build models with similar flexibility to models of human intelligence that can work on a wide variety of tasks… [M]“Ash learning has seen more progress in the past five years than it has in the previous 60,” Luann said. “Historically, long-term AI work has been the prerogative of big tech companies, and their focus on talent and computation has been unquestionable. Looking ahead, we believe the next era of AI breakthroughs will require problem-solving at the heart of human-computer collaboration.”

Whatever form its product – and business model – ultimately takes, can Adept succeed where others have failed? If possible, the windfall could be significant. according to To Markets and Markets, the market for business process automation technologies – technologies that simplify dealing with customers in the enterprise and back-office workloads – will grow from $9.8 billion in 2020 to $19.6 billion by 2026. One 2020 survey With a process automation vendor, Camunda (given a biased source) found that 84% of organizations expect investment in process automation to increase as a result of industry pressures, including the rise in remote work.

Skilled technology makes sense in theory, [but] Talking about adapters needing to be “able to act” seems to me somewhat of a misdirection, Mike Cook, an AI researcher with the Knives & Paintbrushes research group, which is not affiliated with Adept, told TechCrunch via email. Transformers are designed to predict the next in a series of things, that’s it. For a Transformer, there’s no difference whether that prediction is a character in a text, a pixel in an image, or an API call in a small piece of code. So it doesn’t look like this innovation will lead to AI more than anything else, but it may produce AI that is more suitable to help with simple tasks.”

It is true that the cost of training advanced AI systems is lower than it used to be. With a fraction of the OpenAI funding, recent startups including AI21 Labs and cohere It was able to build models similar to GPT-3 in terms of their capabilities.

Meanwhile, continuous innovations in multimodal AI — artificial intelligence that can understand relationships between images, text, and more — put a system that can translate requests into a wide range of computer commands within the realm of possibility. This is how it works like OpenAI directionsa technology that improves the ability of language models such as GPT-3 to follow instructions.

Cook’s main concern is how to proficiently train his AI systems. He points out that one of the reasons other Transformer models have been so successful with text is that there is an abundance of text examples to learn from. A product like Adept is supposed to need a lot of examples of successfully completed tasks in applications (eg Photoshop) paired with text descriptions, but that’s data that’s not naturally occurring in the world.

In a DeepMind study in February, scientists wrote that in order to collect training data for their system, they had to pay 77 people to complete more than 2.4 million presentations of computer tasks.

“[T]The training data will likely be generated artificially, which raises a lot of questions about who was paid to create it, how scalable this is to other areas in the future, and whether the trained system will have the kind of depth that other Transformer models have, Cook said. “that it [also] Not a “path to general intelligence” by any means… it might make it more capable in some areas, but it is likely to be less capable than a system explicitly trained for a particular task and application.”

Even the best roadmaps can face unexpected technical challenges, particularly in relation to artificial intelligence. But Luan puts his faith in the great founding talent of Adept, which includes the former leader of Google Forms production infrastructure (Kelsey Schroeder) and one of the original architects of Google’s production speech recognition model (Anmol Gulati).

“[W]While general intelligence is often described in the context of human substitution, this is not our North Star. Instead, we believe AI systems should be built with people at the center.” “We want to give everyone access to increasingly sophisticated AI tools that help enable them to achieve their goals collaboratively with the tool; Our models are designed to work alongside people. Our vision is for people to stay in the driving seat: discovering new solutions, enabling more informed decisions, and giving us more time for the work we really want to do.”

Greylock and Addition co-led the Adept funding round. The tour also saw participation from Root Ventures and angels including Behance founder Scott Belsky (Behance founder), Airtable founder Howie Liu, Chris Ray, Tesla Autopilot leader Andrej Karpathy and Sarah Meuhas.