🔮 ByteDance’s big leap; grande langue; sodium batteries; predictably random ++ #428
Your insider guide to AI and exponential technologies
Hi, I’m Azeem Azhar. As a global expert on exponential technologies, I advise governments, some of the world’s largest firms, and investors on how to make sense of our exponential future. Every Sunday, I share my view on developments that I think you should know about in this newsletter.
In today’s edition:
ByteDance’s massive GPU order;
AI turns words into matter;
Energy storage sorted.
💯 Thanks to our sponsor Fin, a breakthrough AI bot by Intercom, ready to join your support team.
Sunday chart: Bytedance’s $1bn chip order
ByteDance, the parent company of TikTok, has ordered 100,000 GPU chips, worth $1bn, in 2023 from Nvidia. This is a massive commitment to building out an AI infrastructure, representing the entire market for GPUs in China last year. It is sufficient to build one of the largest private AI clouds in the world.
To put this into perspective, Meta uses 16,000 A100 Nvidia GPUs, although with every hyperscaler and many governments rushing to build more and more powerful infrastructure, this is likely out of date. Just this week the US Department of Energy hit a milestone with its 63,000 GPU Aurora supercomputer capable of a theoretical 2 exaflops (2 x 10^18 operations per second) performance. Tesla is building its own AI supercomputer, targeting 100 exaflops, a task that would require 300,000 Nvidia A100 GPUs.
ByteDance’s order is split between A100 chips (ordered before the US government told Nvidia to stop selling its top-performing HPC cards to China, in August 2022) and H800 cards. The H800 is a Hopper-based custom accelerator Nvidia built to comply with export restrictions — an attenuated version of its H100 accelerator. So while powerful, this would represent a lower performance system than Tesla’s ambitions.
Back in 2016, I remember talking to a semiconductor exec who said that his firm’s modelling implied a nine-order-of-magnitude increase in the demand for compute cycles over the coming 5-10 years. We’re well into that territory now.
Key reads
Aux LLMs, citoyens! Check out the pitch that helped the French entrepreneurs Mistral secure a record $113m of seed funding from top-tier US and European venture capitalists. The funding, based solely on this pitch for an open-source software licence, is a sign that the market believes there could be room for sovereign foundation AI models, not just exclusively American ones. Is it a bubble? At this point in the innovation cycle, investor eyes may get bigger than market stomachs. But as the French say: Marchons!
Further reading: EV reader, Chris Stokel-Walker: which countries could and should plan for sovereign AI?
Material world. Researchers have successfully designed a model that translates natural language instructions into robotic actions to synthesise molecules. The model, referred to as ChemCrow, can take one- or two-sentence instructions and use them to synthesise a variety of compounds. It is an interesting architecture: ChemCrow uses an LLM as an orchestrator which figures out which of a series of sub-tools to use to execute, step-by-step, the user’s instructions, creating what the researchers call “a reasoning engine that observes, reflects, and acts.” I’d not overstate the outcomes of this research: it’s more a small-step to prove that even today’s limited LLMs can be powerful cognitive assistants to a wide-range of thinking work. (And, of course, raises the critical safety concerns that such broad tooling is more widely accessible than ever.)
Elsewhere: AI-driven 3D printers can learn to print materials they have never worked with.
A storage story. To decarbonise effectively, the world is going to need a lot more energy storage capacity. Encouragingly, it’s increasingly clear that it will get it. Research firm Rystad Energy now predicts that annual installations of battery energy storage systems will grow tenfold between 2022 and 2030 (equivalent to 33% CAGR).
This growth (and the growth in electric vehicles) is understandably putting pressure on the lithium industry to scale up more rapidly than any extractive industry in history. But as I discussed in this essay, economics can help. Shortages can increase the appeal of alternatives. And so it seems for batteries: sodium-ion chemistries are getting more popular, forecast to quintuple in the next four years. And, in the medium term, alleviating some of the pressure on lithium.
🚀 Today’s edition is supported by our partner, Fin.
Fin can’t burn its mouth on hot pizza. Or wave at someone who wasn’t waving at them. Fin can resolve half of your customer support tickets instantly before they reach your team. Meet Fin. A breakthrough AI bot by Intercom – ready to join your support team today.
Weekly commentary
This week, EV reader Marshall Kirkpatrick shared his best practices for using ChatGPT in innovation and analytical thinking.
🔮 Using AI for analytical thinking
More high-leverage than prompts for content production or translation, is using an AI system of your choice to ponder strategic opportunities and facilitate thinking for innovation. I want to share some strategies and prompts you can use to support strategic thinking and innovation in your own mind and work.
Market data
Ford secured one of the largest loans to a US carmaker in over a decade, $9.2 billion to build three battery factories to boost US electric vehicles competitiveness against China.
In May, the EU’s combined wind and solar power generation surpassed fossil fuels for the first time.
Only 36% of workers anticipate significant skill changes to their roles in the next five years. This contrasts with the prediction by employers that 44% of workers’ skills will be disrupted in the next five years.
Capital Economics’ forecast suggests that American office properties might not achieve their pre-pandemic peak values until 2040 due to reduced demand for desk space.
London has the most AI tech talent in Europe with 12.29% of the talent share. The next highest is Paris at 3.81%.
Short morsels to appear smart at dinner parties
🦖 Scoring the likelihood of humanity surviving LLMs. (Pay attention to the gender-divided in the answers.)
🐙 Helped and inspired by ChatGPT, SoftBank CEO Masayoshi Son is reigniting his love for tech investment, boosting company shares and advocating for emotionally-intelligent AI. (His investment presentation is off the wall, worth looking at.)
🏷️ A fascinating tale of the hidden labour behind AI. This was the topic of my conversation with Kate Crawford from 2021.
🤖 Household robots are learning tasks from YouTube.
👾 Searching for the killer app: Apple launched its visionOS software development kit for developers.
🕸️ Mathematicians found new ways of predicting structure in graphs.
🇮🇳 The new era of India-US tech cooperation.
End note
I was travelling this week and asked Marshall Kirkpatrick to put together our weekly commentary. His essay covering practical approaches to using systems like ChatGPT to improve your analytical thinking is really great. Recommend reading it in full.
Cheers,
Azeem
📢 We’re looking for cool products and companies to share with our audience. If you’ve got a product or service you think our readers would love, complete this form and we’ll get in touch to discuss.
What you’re up to — community updates
Rumman Chowdhury testified on AI in front of the US Science, Space, and Technology Committee.
Lawrence Lundy and Lunar Ventures built a research tool to understand the deep tech landscape, and they’re making it freely available.
Harry Briggs wrote an article on how to prepare for the AI future.
George Darrah argues that the key to nature’s intelligence is artificial intelligence.
Share your updates with EV readers by telling us what you’re up to here.