This Week in Robotics 19/03

Welcome to this Week in Robotics - providing insight into robotics, autonomy and AI.

In today's email -

• GPT-4 & its robotics application
• Investors unworried by robot-related job loss
• Rapid robot growth
• Are AIs a hot mess?

News

CEOs automated out - Last August, NetDragon Websoft, a billion-dollar gaming firm, appointed an AI CEO to lead their main subsidiary. The new boss was responsible for all of the typical duties of a CEO - analysis, decision-making, assessing risks, etc. So far, having an AI CEO has worked pretty well - the company has outperformed Hong Kong’s stock market... If only all CEOs worked 24/7, didn’t sleep and accepted $0 annually.

Has the hype been met? Open AI finally released the latest iteration of their AI, GPT-4. Some highlights -

• It's multimodal?and now accepts both image and text inputs
• It has human-level performance in pretty much every academic benchmark, from SATs to the Advanced Sommelier qualification
• Its already being used by - Stripe to combat fraud,?Duolingo as a language partner and Morgan Stanley as a knowledge database
• It seems to have more confidence in its capabilities, even where they might be unfounded

Should roboticists care? See Research for Microsoft's answer.

Robots on the rise - The Association for Advancing Automation reports that robot sales in North America have increased for the second year in a row. Last year 44,196 robots were purchased, worth ($2.38 billion). The automotive industry was a driving force behind this, accounting for more than 50% of sales. This trend matches the skills gap the manufacturing industry has faced since 2020 but hasn't been reflected in productivity, which decreased by 2.7% last year.

So many acronyms - Matt Clifford MBE, chair of the UK's Advanced Research and Invention Agency (ARIA) and co-founder of Entrepreneur First (EF), has been appointed to advise the UK government as it establishes a taskforce on foundation models (FMs?). Its goal is to harness AI's benefits to grow the economy, create jobs, ensure security, and benefit society. The taskforce's first step will be advancing the UK’s AI capability and addressing the challenges of controlling and regulating such systems.

Opinions

Knowing without understanding - Cognitive scientist and linguist Noam Chomsky believes that GPT will never truly understand -

The human mind is not, like ChatGPT and its ilk, a lumbering statistical engine for pattern matching, gorging on hundreds of terabytes of data and extrapolating the most likely conversational response or most probable answer to a scientific question. On the contrary, the human mind is a surprisingly efficient and even elegant system that operates with small amounts of information; it seeks not to infer brute correlations among data points but to create explanations.

Machine learning is great at using big data and statistics to describe and predict but it can't propose causal mechanisms or physical laws. As such, it will always be stuck in a "non-human phase of cognitive evolution". Without true understanding, it is unlimited in what it can memorise but will either over-generate, producing both truths and falsehoods equally easily or under-generate, exhibiting non-commitment to any decisions. The real question is, at a certain scale, does it matter?

Robots can't steal your job - Famed investor Andreessen Horowitz isn't worried about robot-related job loss. His arguments -

1) Same as it ever was - The fear of new technology causing overall unemployment has existed for centuries, but historically, jobs and wages have continued to rise. Two anti-technology job scares occurred in the last 20 years with outsourcing and robots, resulting in the best global economy with the most jobs and highest wages.?Counter - Just because things have been one way doesn't mean they'll stay that way.

2) It's illegal - AI can't cause employment because it's effectively "illegal" due to regulations, monopolies, and cartels in education, healthcare, housing, etc. (see red). These sectors do not allow technological innovation to push down prices. In contrast, sectors with less regulation (blue) allow technology to improve quality and lower prices. Over time, the regulated sectors continuously grow as a percentage of GDP, while the less regulated sectors shrink.?Counter - just because they're not letting technology bring down prices doesn't mean they're not using technology to become more efficient.

The hot mess theory of misaligned AI - Many worry that super-intelligent AIs will narrowly pursue the wrong goal to disastrous consequences. Should we fear the opposite? Humans are (seemingly) the smartest creatures in existence and are a total mess. We pursue inconsistent and non-static goals and constantly engage in self-sabotaging behaviour. Most work on AGI misalignment risk assumes that, unlike us, smart AI will not be a hot mess.

Could this be wrong? When large language models behave unexpectedly, it is rarely because there is a clearly defined goal they are pursuing instead of their instructions. They are normally doing something which is both poorly conceived and sensitive to seemingly minor details of prompt phrasing, sampling technique, and random seed.

Why might incoherence be the more likely outcome? Complex systems are harder to control than simple systems. Requiring a system to act only in pursuit of a well-defined goal or only to maximize a utility function is an extremely strong constraint on its behaviour. This constraint could become harder to satisfy as the system becomes more intelligent and more complex.

Research

My drone is too Agilicious for ya - The Robotics and Perception Group at the University of Zurich have developed Agilicious, an open-source framework for designing high-speed, agile drones. Their platform includes hardware, software and simulation plans and has been used in over 30?papers for acrobatic flight, drone racing, and obstacle avoidance.

Chating with your bot - Is ChatGPT relevant to robotics? Microsoft thinks so. Their idea is that instead of a technical engineer programming in the loop, a non-technical user can "prompt engineer" the robot on the fly. Microsoft is building a few different tools to help roboticists with this -

• PromptCraft - an open-source platform where anyone can share examples of prompting strategies for different robotics categories.
• ChatGPT-AirSim - a robotics simulator for testing GPT prompts

The demo is pretty basic, but we're still day 0 and expect to see LLMs across robotics.

Rabbit robots - Disney has demoed their latest robot at SXSW. The demo highlights their long history of using robotics to create characters that tell stories and elicit emotions. The techniques Disney is perfecting might only seem relevant for an entertainment park (West World?) but as robots become more prevalent, expect to see more robots designed with little features that exploit human psychology, in good ways & bad.

Vine by me - Researchers have developed a vine-like robot that can navigate complex paths without touching its environment. Unlike traditional vine robots that use air pouches for steering, this robot features a multi-segment design that allows for selective actuation of individual pouches at the tip. A small magnetic valve connects each pouch to a pressure supply line, allowing it to grow into different shapes and hold them while navigating complex paths.

Tweet

All views and opinions expressed in the briefing reflect those of the author/authors and no other entity.??