- AI For All
- Posts
- The Robots Are Self-Improving
Hello readers,
Welcome to another edition of This Week in the Future! We have two breakthroughs from NVIDIA and IBM — self-improving robots and a new AI chip architecture.
As always, thanks for being a subscriber! We hope you enjoy this week’s content — for a video breakdown, check out the episode on YouTube.
Let’s get into it!
The Robots Are Self-Improving
Generated by DALL·E 3
LLMs continue to prove useful for robotics. NVIDIA’s research breakthrough, an AI agent dubbed Eureka, automatically generates reward functions to train robots to accomplish complex tasks. The example shown was that of a robot doing a pen-spinning trick with the same dexterity as a human.
The agent uses GPT-4 to generate reward algorithms, outperforming expert human-written ones in over 80% of tasks. It summarizes training results and instructs the LLM to improve its reward functions, making for a system that is effectively self-improving.
Why This Matters
Aside from more capable robots, NVIDIA has demonstrated the utility of LLMs in robotics applications and opened the door to robots that train and improve themselves. Compared to software, robotics has been slow, moving step by step, much like a robot. However, the tables might turn if generative AI hits a ceiling with productivity apps but expedites progress in robotics, unleashing a whole new wave of value.
Our Take
For some, their worst fears might have just been realized — iterative self-improvement. However, if the so-called “Lebowski Theorem” is to be believed, then no super-intelligent system is going to do anything that is harder than hacking its own reward function, which could mean amassing a robot army to overthrow humanity would be too much effort. One risk might be that since LLMs are influenced by language, could using an LLM expose robots to the same hacking and jailbreaking that chatbots have experienced? 🤔
IBM’s North Star
IBM's NorthPole chip, developed over nearly two decades, is a breakthrough in AI chip architecture. Notable features and specs include:
It moves away from traditional von Neumann architecture, integrating memory and processing units to eliminate data shuffling bottlenecks, significantly enhancing efficiency.
NorthPole showcased superior performance in tests with popular AI models, being roughly 4,000 times faster than its predecessor, TrueNorth.
The chip is constructed using a 12-nm node process, housing 22 billion transistors, and is capable of performing 2,048 operations per core per cycle at 8-bit precision.
Our Take
IBM’s new chip can substantially improve AI's efficiency, making it more accessible and cost-effective, especially in enterprise settings. Its design, mirroring the brain's computing style, not only breaks the von Neumann bottleneck but also paves the way for scalable AI hardware systems. However, its focus on inferencing and in-chip memory could limit its versatility for broader AI applications, though multiple chips can be linked to work with large neural networks.
🔥 Rapid Fire
Cisco launches AI-powered Webex
Adept open-sources multimodal Fuyu-8B
Boston Dynamics makes robot tour guide
Datasaur launches LLM lab for enterprises
New details on impending AI executive order
IBM acquires Manta Software for AI governance capabilities
Amazon adds AI image generation
Lenovo CEO says every smart device will have a personal AI
UN announces AI advisory board
OpenAI forms AI risk preparedness team
Frontier Model Forum announces Executive Director
Microsoft’s AI strategy pays off
Apple is preparing to add AI to every product
🎙️ The AI For All Podcast
This week’s episode featured Jannick Malling, the co-founder and co-CEO of Public, who discussed the impact of AI on the world of investing and finance and how AI is democratizing investing and leveling the playing field.
📖 What We’re Reading
This week’s handpicked insights include the right questions to ask when looking for manufacturing solutions claiming to offer AI capabilities, plus Accenture has shared insights into the trends shaping … life, which of course includes generative AI.
The Right Questions To Ask When Evaluating An AI Condition Monitoring Solution (link)
“A discussion has been sparked on the best questions for manufacturers to ask when they are out to invest in an AI-driven manufacturing solution. After all, with few regulations and shifting vocabularies, many vendors are using whatever terms to make a sale – when in fact their AI is weak or non-existent.”
Accenture Life Trends 2024 (link)
“Generative AI is upgrading people’s experience of the internet from transactional to personal, enabling them to feel more understood and relevant than ever.”