Welcome to the Bulletin by Remix Robotics, where we share a summary of the week's need-to-know robotics and automation news.
In today's email -
- Why we need more robot soldiers
- GPTchat is a big deal (even for robotics)
- How Dalle-2 can help robots imagine
- AI’s that can read your brain
Killer robots are groovy - Well at least San Francisco's ruling Board of Supervisors thinks so. They have voted to let the city's police use robots that can kill. We discussed the proposal last week -there was a bit of debate but it was passed by an 8-3 vote. The heart of hippy culture has come a long way. This news is getting a lot of publicity but there are already a few lethal robots being used by US police. As an example - police in Dallas used a robot armed with explosives to kill a sniper in 2016.
Nuclear Space robots - A team of engineers at the UK Atomic Energy Authority is applying its robotics expertise to space. They often use teleoperated robots inside fusion reactors - now they’re experimenting to see if the same technology can be adapted to repairing the 6000 broken satellites orbiting the earth.
Truck stop - The autonomous truck software company, Embark has lost 98% of its share price since going public a year ago at a valuation $5 billion. Strangely it's now worth less than its cash reserves - with a market cap of $110 million vs $191 million in cash... Basically, investors think it has no real value. Apparently, the company is still trucking on and will deploy in 2024.
Robot inception -ABB is developing an automated factory to build automated robots, automatically with robots (sorry). The facility will not have flexible, modular production cells that are served by autonomous mobile robots in favour of traditional fixed lines. “AI-powered” robots will take on tasks such as screwdriving, assembling and material handling.
Anduril Raises $1.48 Billion - The company builds software and hardware-enhanced with artificial intelligence and machine learning for their military and defence industry. It is now worth $8.48 billion.
Open AI releases GPTchat - and everyone is losing their minds. GPT, the AI text generator has been out for a few years but OpenAi has just made it easy to use with a new and beginner-friendly UX - I recommend having a play. It's easy to be over-saturated by the number of use cases, here are the best ones -
- Debugging code
- Building a Linux machine
- Website development
- Developing business strategies
- Essay writing
- My favourite - songwriting
Also cool - Hackers were quick to find vulnerabilities in its content moderation policies. These models are so huge and complex that not even the developers really know how they work (see the post on emergence below). Its interesting that the bot can almost be socially engineered, with hackers using all sorts of tricks to coax the AI to break its rules, whether asking it to - imagine its allowed to break its own rules, write a poem or reveal admin details.
Why is this interesting to roboticists - Similar to chess, AI/human “centaurs”will outperform either humans or AIs independently. Regardless of our industry, we all have a new goal - to figure out how to apply these AI tools to our workflow. Whether it's automating email writing so we have more time for innovative research or streamlining code development - there is a lot of scope for value add.
Also, expect robotics researchers to keep stealing and borrowing ideas from generative AI and implementing them in robotics. See Research below…
Imagine all the robots - Researchers have developed Dall-e-bot an AI system that uses Dalle-2 to give robots “imagination engines” capable of understanding how a system should be organised to suit human aesthetics.
How does it work -
- First, key features are identified in a “scene” - this could a pile of cutlery on a table, a stack of dirty dishes or a messy bedside table
- These features are used to generate a text description - “A fork, a knife, a plate on a kitchen table”.
- This text description is passed to DALL-E.
- Now the clever bit - DALL-E is used to generate a goal image for the objects - basically this a layout that would make aesthetic and practical sense to a human.
- The robot then uses this layout as a guide for how to rearrange the objects.
Why is this interesting? Robots understanding what humans want without specific programming - very cool. Generative AI is improving so rapidly and roboticists are getting very jealous. Expect to see them borrowing and stealing developments to make robotics reason and interact with the world more naturally.
Robot mind readers - Researchers have used developed an AI able to read your mind. MinD-Vis is capable of decoding fMRI-based brain activities and reconstructing images. Basically, it uses machine learning to decode your brain waves and guess what image you're picturing. It doesn’t work perfectly but researchers say it is a step towards a non-invasive machine brain interface which removes many of the challenges of embedded approaches like Neuralink. What they don't mention is the potential impact on “neuroprivacy” - what happens when an AI can read your thoughts without your consent or even knowledge?
Don’t hate the player - Last week Meta announced an AI capable of beating humans at the board game Diplomacy. DeepMind must have been feeling left out as they’ve just released DeepNash, an AI capable of beating experts at the game Stratego. Strategio requires less negotiation / manipulation than Diplomacy but still requires “long-term strategic thinking in the face of imperfect information”. Why is this interesting? The techniques that game-playing AIs use to interact with humans and account for irrational behaviour could be applied in the real world “If you’re making a self-driving car, you don’t want to assume that all the other drivers on the road are perfectly rational, and going to behave optimally”. Or…. it could be used to manipulate people.
The Code war - AI will disrupt warfare, China is currently outpacing the United States in AI warfare and as result, the US needs to invest in AI military applications. Or so argues Alexandr Wang - it's the “if we don’t, they will” argument. The author layout some interesting facts -
- China considers AI as a “historic opportunity” for “leapfrog development” of national security technology, per China’s 2017 National AI Development Plan.
- China is outspending the United States on AI technology for defence. The People’s Liberation Army (PLA), spent between $1.6B and $2.7B on AI against an overall defence budget of $178B in 20202 whereas the US Department of Defense (DoD) spent only between $800M and $1.3B on AI against an overall DoD budget of $693B over the same period
- In many DC wargames of the past few years, China has won. The quotes are “The United States gets its ass handed to it” & “We are going to lose fast”
What's the solution -
- Data supremacy - America has by far the largest fleet of military hardware. If they can successfully turn this platform advantage into a data advantage through an investment in data infrastructure and data preparation, they can get ahead and stay ahead.
- Shift 25% of the DoD budget towards AI-enabled capabilities by 2032 and focus on self-disruption + training.
Emergent properties - Emergence is a phenomenon in AI research, where a model’s performance improves nonlinearly with increased scale. It's a double edged sword, although you get powerful gains from scaling, it's inherently unpredictable. Large-scale models tend to have 'hidden' capabilities and safety issues as a consequence of emergence. This blog shows 140 odd examples of emergence spread across different language models including GPT-3, LaMDA, PaLM, Chinchilla, Gopher. Why this is interesting? Emergent capabilities are an indicator that we have a 'capabilities overhang' - today's models are far more capable than we think, and our techniques available for exploring them have a long way to go.
This is a great rule of thumb - see our piece on Moravec's paradox for more.
“In 3-8 years we'll have a machine with the general intelligence of an average human being. The machine will begin to educate itself with fantastic speed. In a few months, it will be at genius level and a few months after that its powers will be incalculable.” —Marvin Minsky “Godfather of AI”, 1970