Understanding the AI Value Chain

Welcome to the Robot Remix, where we summarise the week's need-to-know robotics and automation news.

In today's email -

• Meta's head of AI wants machine learning to suck less
• Helping robots reason the real world
• No more robot soldiers?
• Self-driving stalls

Snippets

“Machine Learning Sucks” - Meta’s AI Chief Yann LeCun has published a  paper on his vision for the future of general AI. He claims that neural networks are not sufficient to create machine intelligence. Their dependence on lots of data makes them brittle and requires over-engineering to deal with edge cases.  LeCun proposes a new model of machine intelligence that uses "common sense" to understand the world around it. It's a very broad paper and is self-admittedly a bit “hand-wavey” but worth checking out. This interview is a bit more accessible.

Let's get physical  - Deepmind has improved its AI’s ability to reason with the physical world by combing Large Language Models and computational physics engines. The language model was trained to convert physics questions on motion, collisions, frictions, etc into code which can be run on Deepmind’s MuJoCo simulator, the results of which are converted back into text. This system known as Mind's eye improved the system's accuracy from 39% to 93%. We can see how this fits into their robotics strategy!

War, what is it good for - Six robot manufacturers have published a letter promising not to weaponize their products “when possible”. Some have questioned the effectiveness of this pledge as it only includes general-purpose technologies - “we are not taking issue with existing technologies that nations… use to defend themselves and uphold their laws” and acknowledged that it is impossible for them to stop misuse. The Genie may be out of the bottle - DIY drone bombs have become commonplace in conflicts including Ukraine.

It’s like watching a car crash - A Crunchbase analysis of 14 public companies in the self-driving ecosystem shows an average post-debut decline of more than 80%. One company even dropped 97% from their debut price 11 months ago. What caused this crash? Crunchbase states that “investors have mostly given up” due to the continued lack of results. If you ask Yann LeCun (see above) self-driving startups have been "a little too optimistic "by thinking one could "throw data at" large neural networks "learn pretty much anything”.

Who needs robots anyways - The International Federation of Robotics annually releases shipments for industrial robots by country, check out this great analysis of 2021’s results.  The takeaway - China accounts for the majority of global demand at 52%, Europe, saw a record year of demand at 84,000 shipments and the UK didn't even make the top 15 countries with around 2000 robots purchased… a 7% drop from 2020… Yikes.

Big Idea

Understanding the AI value chain

In trying to understand who will win the AI race for supremacy Every, a newsletter lays out the industry’s value chain. The AI value chain can be broken into 6 overlapping categories -

• Compute: The chips or server infrastructure required to run AI models
• Data: A data set that a model is trained on
• Foundational Model: The easiest way to think of an AI model is as a digital robot that can do a task. It combines compute and data with fancy math to create the desired output.
• Fine Tune: Often foundational models are not sufficient for a specific use case, but need to be tuned for a specific scenario
• End User Access Point: The model will then be deployed in some sort of application

AI is at its most powerful when companies blend these categories together and we expect this will create new business models for companies including -

• Integrated AI: AI capabilities will be integrated into existing products without dislodging incumbents. Rather than an AI company building a CRM from scratch, it is much more likely that Salesforce incorporates GPT-3 or Adobe builds Dalle-2 into Photoshop, etc.
• Infrastructure as a Service: Evan expects major consolidation at all levels of the value chain. Cloud providers like AWS, Oracle, and Azure could build their own custom AI workload chips, build networking software, and train in-house models that people can reference to solve specific tasks.
• Invisible AI: Invisible AI is when a company is powered by AI but never even mentions it. They simply use AI to make something that wasn’t considered possible before but is entirely delightful - he sights Tick Tok as an example of this.

Video

Chips don't dip

Semiconductors have been the big news story of the last 18 months. COVID highlighted the vulnerability of our supply chains and nowhere was this more pronounced than in control systems.  

It's common knowledge that Taiwan and TSMC are at the forefront of chip development but it may interest you to know that the entire industry depends on a machine made by a single Dutch company, ASML. Their machine combines novel manufacturing technologies with fixed automation and robots, it's a great example of automation being used to create a differentiated product that is driving the news.

Check out this youtube video to learn more about their turbulent path to a $400B valuation.

Podcast

Making hardware less hard

Check out this week's podcast recommendation - an interview with  Eliot Horowitz, ex-founder / CTO of MongoDB and current founder / CEO of Viam Robotics.  Their product is a cloud-based, low-code programming platform that allows developers to quickly and easily program any robotic hardware.  In this podcast, Eliot lays out -

• Why he started a robotics company? Leverage. All the big problems Eliot wanted to solve after leaving MongoDB (climate change, ocean pollution, New York potholes) could be better solved with robots.
• Why robotics is so hard? Hard to use and inflexible software for hardware development. Every new project needs to reinvent the wheel and use a tonne of custom code to stitch together different hardware components.
• His goal? Accelerate the proliferation of robotics companies.

“We want 3 smart 25-year-olds to be just as willing to start a robotics start-up as a social media start-up”.

A truly universal robotics platform is tough to achieve but with a free beta out now - we're interested to see how it progresses!

Tweet

The Genie is definitely out of the bottle