From Your Shower to Your Robot: This Week’s Most Radical AI Developments
This week in radical tech: DIY frontier models, the surprising energy footprint of AI, and how AGI expectations are already reshaping our economy
Dear Friend –
AI truly has become a tale of two cities (“It was the best of times, it was the worst of times…”). On one hand, we have the never-ending barrage of advancements (Claude 3.7 earlier this week, ChatGPT 4.5 yesterday, Meta announcing a standalone AI app, Microsoft launching CoPilot as a macOS app…), and on the other hand, all of these companies are burning absolutely eye-watering amounts of money without a clear path to profitability – and seemingly nobody is using AI in a truly scaled-up production environment. We truly are in a liminal space when it comes to AI – meanwhile, I counted how much I personally use AI on a given day… My primary interface for using AI is Raycast, the macOS app launcher that gives me access to all frontier models in a singular interface. On an average day, I have easily 100 interactions with AI – meanwhile, my use of Google Search (or in my case, more specifically DuckDuckGo) has dropped dramatically. Truly a tale of two cities…
And now - this:
Headlines from the Future
Train Your Own O1 Preview Model Within $450 ↗
The only way is (seemingly) down: A team at UC Berkeley has trained their own O1-preview style model for a mere $450:
Remarkably, Sky-T1-32B-Preview was trained for less than $450, demonstrating that it is possible to replicate high-level reasoning capabilities affordably and efficiently.
We are getting closer and closer to a world where you either just use one of the omnipotent frontier models or simply roll your own.
—//—
AI vs. an Extra Minute in the Shower ↗
You read the headlines stating that GenAI is a true energy hog? Still questioning what your personal use of AI means in terms of energy consumption? Here is your answer:
Let’s proceed, then, with two types of users:
A conservative user: Uses a model that has an energy use of 2 mWh per token and that leans towards 200 tokens on average per response. The user performs 10 queries per day.
A heavy user: Uses a model that has an energy use of 9 mWh per token and that has longer responses of on average 1000 tokens. The user performs 500 queries per day.
With the numbers found above, the conservative user would have an energy footprint of 4 Wh per day, from their use of LLMs. The heavy user, on the other hand, will have a footprint of 4.5 kWh per day. 4 Wh is less than an efficient LED bulb will use in an hour, while 4.5 kWh is about the amount of energy my panel heater uses to keep my bedroom at 22 °C on a typical winter day. (I live in Norway.) The average data center uses 1.7 liters of water per kWh of energy consumed [2], which means the conservative user spends an extra 7 mL of water a day on their LLM use, while the heavy user spends 7.6 L — about the minutely water consumption of an efficient shower.
—//—
Strategic Wealth Accumulation Under Transformative AI Expectations ↗
Here is a fascinating paper examining the impact that future assumptions about Artificial General Intelligence (AGI) (or “Transformative AI (TAI)” as outlined in the paper) becoming real will have on people’s behavior today.
The main takeaway: Just the belief that transformative AI is coming could push interest rates much higher, even before the technology actually exists. This, in turn, could affect how central banks manage the economy and overall financial stability.
The train of thought works like this:
The key idea is that when advanced AI arrives, it will replace human workers, and the money that used to go to workers will instead go to people who own AI systems
The more wealth you have when AI arrives, the more control you’ll have over AI systems and their earnings
The researchers used economic models based on current predictions about when this powerful AI might arrive
They found that even moderate predictions about this future scenario are causing some interesting effects today:
Interest rates could rise much higher (to 10-16%) compared to normal rates (around 3%)
People are willing to accept lower returns on investments now because they’re focused on building wealth to control future AI systems.
—//—
GenAI Is Coming for Your Robot ↗
In case you missed it, GenAI promises to be a boon for robotics. One of the significant challenges in robotics is providing robots with a comprehensive understanding of the real world, which is often quite messy. By using multi-modal GenAI models, robots can gain a better understanding of their environment and respond more effectively.
Microsoft Research released Magma, a foundational model for multimodal AI agents:
Magma is a significant extension of vision-language (VL) models in that the former not only retains the VL understanding ability (verbal intelligence) of the latter, but is also equipped with the ability to plan and act in the visual-spatial world (spatial intelligence) and to complete agentic tasks ranging from UI navigation to robot manipulation. […] Magma creates new state-of-the-art results on UI navigation and robotic manipulation tasks, outperforming previous models that are tailored specifically to these tasks.
—//—
Here’s a Thing GenAI Is Actually Good At: Science ↗
Still wondering what the whole GenAI craziness is all about? You are certainly not alone (and some folks, like Ed Zitron, will gladly tell you that it’s all a giant con operation). But there seems to be, aside from the obvious elephant in the room “coding,” at least one area where GenAI shines: science.
Points in case: Google’s AI solved a 10-year superbug mystery in just two days.
[…] While the team knew about this tail-gathering process, nobody else in the world did. Imperial’s revelations were private, there was nothing publicly available, and nothing was written online about it. The scientists then asked the co-scientist AI, using a couple of written sentences, if it had any ideas as to how the bacteria operated. Two days later, the AI made its own suggestions, which included what the Imperial scientists knew to be the right answer.
Meanwhile, Nvidia unveiled an AI system to aid in genetic research.
Scientists have high hopes that such AI technology will dramatically accelerate research by spotting patterns in vast amounts of data that would normally take years to analyse by hand. The system learned from nearly 9 trillion pieces of genetic information taken from over 128,000 different organisms, including bacteria, plants, and humans. In early tests, it accurately identified 90% of potentially harmful mutations in BRCA1, a gene linked to breast cancer.
—//—
Trying to Make Sense of Microsoft’s Quantum ‘Breakthrough’? ↗
Microsoft, this week, announced a breakthrough in their approach to quantum computing in the form of their novel Majorana 1 chip. The announcement made some waves in the eternally “next year is the year of quantum computing” community – and if it leaves you scratching your head, this FAQ from Scott Aaronson might be helpful.
Q6. Is this a big deal?
A. If the claim stands, I’d say it would be a scientific milestone for the field of topological quantum computing and physics beyond. The number of topological qubits manipulated in a single experiment would then have finally increased from 0 to 1, and depending on how you define things, arguably a “new state of matter” would even have been created, one that doesn’t appear in nature (but only in Nature).
What We Are Reading
💥 Is Constructive Dissent the Key to Innovation in Your Team? Innovation sparks when teams master the art of friendly fighting—where different viewpoints clash and collaborate rather than getting swept under the rug. @Jane
🔐 Google Is Replacing Gmail’s SMS Authentication With QR Codes QR codes keep proving they’re not ‘just’ for restaurant menus. Google is aiming to tackle the “rampant global SMS abuse” by replacing authentication text messages with QR codes that you can scan with your phone. @Mafe
🤖 The Handoff to Bots Kevin Kelly offers a thoughtful argument for the AI economy in the long term: a purposeful -- and even necessary! -- economic handoff as the pool of human workers shrinks with declining global fertility rates. @Jeffrey
💭 The Fantasy of a Nonprofit Dating App Apart from learning about Japanese state-funded dating platforms, this article poses an interesting question: how might non-profit alternatives to the now ever-present commercial tech solutions look if they were set up radically differently? While such questions can open entirely new pathways, they are also quite practical product strategy exercises for commercial players. @Julian
🧇 Nike Receives Patent for Waffle-Soled Trainers — Invented in a Waffle Iron Breakthrough shoes made in a kitchen appliance? This is precisely how it started. Nike co-founder Bill Bowerman’s wife was inspired by the waffles she was making and suggested to Bowerman that this could be a good design for a new sole. A ruined kitchen appliance, some polyurethane, and a bit of sewing later led to the Nike Waffle Trainer prototype we all now know. @Pedro
👂 Are Noise-Cancelling Headphones Impairing Our Hearing Skills? Some Audiologists Are Beginning to Worry Discuss “unintended consequences”: The overuse of noise-cancelling headphones, while allowing you to listen to music at lower volumes and reducing strain on your hearing, may cause your brain to struggle with locating the source of sounds. @Pascal
Some Fun Stuff
🍢 The closer to the train station, the worse the kebab – A “Study”
🎹 The last bastion of humanity is quickly eroding: RoboPianist - Dexterous piano playing with deep reinforcement learning.