Welcome to the Robot Remix, where we provide weekly insight on robotics, autonomy and AI
This week -
- An artificial brain made from... real mouse brains
- Modular moon robots
- In hand dexterity
- Graphs & a Meme
Research
Almost AGI - Researchers have benchmarked GPT-4 and found that it exhibits "sparks of artificial general intelligence". It's a long paper filled with amazing examples but the most interesting sections are on the model's limitations, its -
- Great at incremental tasks - which can be solved by gradually adding one word at a time. e.g. summarising text, answering factual questions, composing a poem, and solving a math problem with a standard procedure
- Bad at discontinuous tasks - that require a ”Eureka” moment or a leap to progress. E.g. writing a joke or a riddle, coming up with a scientific hypothesis, solving a math problem with a novel formula
Why? There's a fundamental flaw holding it back - the model relies on a local, "greedy" process of generating the next word without any global or deep understanding of the task or output. To add to this, some other limitations -
- Continual learning: The model lacks the ability to update itself or adapt to a changing environment. One can fine-tune the model on new data, but this can cause degradation of performance or overfitting
- Transparency, interpretability and consistency: Not only does the model hallucinate, make up facts and produce inconsistent content, but it has no way of verifying whether it is consistent with the training data or whether it’s itself. It is also very good at post-rationalising incorrect answers
- Cognitive fallacies and irrationality: The model exhibits the same cognitive biases and irrationality as people, probably inherited through its training data
Will the rate of progress continue or will Large Language Models (LLMs) reach the S-Curves shoulder? See the Graph below for more.
Google getting bodied - Google has launched PaLM-E, a multipurpose LLM that hopes to create intelligent, embodied robots. The key innovation here is that the model can ingest robot sensor data directly, as well as text and images. The result is a general-purpose model that enables robot learning while maintaining language-only capabilities.
Shape rotators - Touch Dexterity is a new system that can perform in-hand rotation by touch alone. Their approach is interesting - they use cheap binary sensors rather than strain/compression sensors to surpass the state-of-the-art. Their approach provides great sensing coverage and minimises the gap between Sim2Real while also reducing costs/complexity.
Cyborg biocomputing - Researchers at the University of Illinois have created a "living computer" that uses 80,000 mouse brain cells to solve complex problems and recognize patterns of light and electricity. Why? Biocomputers are much more efficient than digital neural networks and far quicker to train. It's unsurprising that biological matter surpasses our best attempts to copy it digitally but its still a bit creepy.
We built a zoo(bot) - MIT researchers won the Best Paper Award at the IEEE Aerospace Conference for their modular lunar robot called WORMS. The team behind it wanted to avoid “a zoo of machines” with specialised robots for every task imaginable.
Each module is inspired by a different animal and can be swape to perform different tasks. They have a spider for crawling in tight tunnels, an elephant for hauling heavy equipment and an Ox for transporting solar panels.
News
OpenAi invests - 1X Technologies, a humanoid robotics company formerly Halodi Robotics, has raised $23.5 million in funding led by the OpenAI Startup Fund. The funds will be used to build the upcoming bipedal android model NEO which aims to have human-level dexterity.
The secret history - Semafor has released some fairly salty gossip about how and why OpenAI became public. The timeline-
- In early 2018, Musk offered to take control of OpenAI and run it himself, as he thought the venture had fallen fatally behind Google.
- Altman and OpenAI’s other founders rejected Musk’s proposal.
- Musk, in turn, walked away from the company and reneged on a massive planned donation. That conflict created a public rift between, Musk and Altman.
- Later, OpenAI made a big decision to pivot toward transformer models, requiring enormous amounts of training data, so OpenAI created a for-profit entity to raise enough money to pursue the most ambitious AI models.
- Less than six months later, OpenAI took $1 billion from Microsoft, which could provide funding and infrastructure know-how. Together they built a supercomputer to train massive models that eventually created ChatGPT and the image generator DALL-E.
The Jetsons - It's been six months since NVIDIA unveiled the Jetson Orin Nano GPU, a high-performance, low-power AI processor that provides AI on edge. NVIDIA is moving quickly - their latest release has promised an 80x performance increase from the last release.
Hammering it out - The U.S. Air Force has awarded Machina Labs a contract to automate composite manufacturing. Many companies have gone after this challenge in the last decade with mixed results. Although composite manufacture is critical to important industries like aerospace, energy, defence etc - it's still highly manual due to the dexterity and skill required. Composite tooling can cost over $1 million and take 10 months to build.
Machina Labs tackle this challenge by using hammer-wielding robots that can make bespoke tooling with the dexterity of a blacksmith at the scale of a modern factory.
Standing on its own two legs - The biggest news from this week's ProMat trade show is a bipedal robot named Digit. Agility Robotics has been working on the system since 2019 and it's now one step away from being released. It's got a few updates -
- A good head on its shoulders -The head is reminiscent of Rethink Robotics, the cobot OG and provides a focal point so humans can naturally understand what Digit is doing
- Single-minded - Although Digit is destined to be multipurpose, the company has bounded its first use case to something simple - moving plastic totes around a warehouse
Mimicking the human body is easy (ish), but reaching our level of efficiency is hard - Bipedal robots are more complex/expensive/energy intensive than wheeled robots. Agility argues that the form factor is necessary if robots are to slot seamlessly into our lives without infrastructure change.
Opinions
It's still Day 1 - Greg Smith, the President of Teradyne (who owns Mir and Universal Robot) presented on the growth potential of robotics.
By his estimates, robot arms are at 2% market penetration and mobile robots at 3%...
The Library of Babel - As generative AI reduces the cost of writing and image generation to zero, Lincoln Michel asks how our relationship with content and the internet will change.
The internet is already filled with noise and misinformation and generative AI stands to make these issues worse. Counter-intuitively - this might solve the problem in the long run...
Many chatbot use cases work for individuals but fail with mass adaptation.
Google just demoed an AI that reads your emails, Microsoft an AI that writes your emails, people have automated tinder... the list goes on.
Each of these works if you have a monopoly on AI. All of them are futile if everyone has equal access. So there are two outcomes-
- A never-ending “AI” arms race. One person pays to use an AI to generate; another pays an AI to filter
- The platform becomes unusable & dies
The result is roughly the same - content as we know it changes, the internet in its current form becomes unusable and we adapt to a new paradigm. In that disruption, we might end up with something that solves the misinformation and spam of the current web... Can't tell if this an optimistic take or pessimistic one? Not can I...
The one assumption here is that AI is evenly distributed. If power is concentrated, the dynamics change for the worst.
ClosedAI - Futurism argues that OpenAI has shifted its ethics and is now prioritising profit over safety and equality. The author reminds us of OpenAI's founding ideal -
OpenAI is a non-profit artificial intelligence research company. Our goal is to advance digital intelligence in the way that is most likely to benefit humanity as a whole, unconstrained by a need to generate financial return
Since then, the company has essentially become for-profit, stopped being transparent about its methods and moved away from open source.
Why the big change? Last year OpenAI was nearly overtaken by Stability AI, Midjourney, et al. Today it feels like OpenAI is racing ahead. The author speculates that when the need to win kicks in, ideals can easily be left behind. The Silicon Valley ethos of "Move fast and break things" lives on. See OpenAi's recent data leak more. Somewhere Yudkowsky is screaming into a pillow
Graph
Slowly then, all at once.
Long Read
Large Language Manuscripts - In February, Stephan Wolfram (of WolframAlpha, Mathematica + a lot more), found the time to write a 20,000-word blog post on LLMs. As you might imagine - it's worth reading.
The blog post is available for free, but he's just released it as a paperback if you'd rather consume it in physical form.
Meme of the Week
All views and opinions expressed in the briefing reflect those of the author/authors and no other entity.
