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:
The era of autonomous agents: Auto-GPT and BabyAGI win developers’ hearts.
Open-source AI — StabilityAI is paving the way for a different way to work with AI.
Microsoft is developing its own chip.
Today’s edition is supported by our knowledge partner, Singularity Group.
80 Leaders. 20 Experts. 5 Days. 80 leaders like you who share a passion for innovation and a curiosity about the future. 20 experts to teach you about the next wave of exponential technologies. 5 days of intense collaboration that will change how you see the world and your role in it. This is the Singularity Executive Program. Limited seats remain for our April and June cohorts.
Interested in sponsorship? Get in front of 261,000+ executives, investors and technology leaders. Book a call with our team and learn about EV sponsorship opportunities.
Sunday chart: Autonomous AI
OpenAI’s GPT-4 API was released last month. It enables developers to connect the AI to other software and essentially play with new ideas. One of the experiments that is gaining most traction is “agentic AI”. Autonomous agents (which we wrote about in EV#418) are designed to generate and carry out multistep plans by interacting with their environment and using it to learn and iterate.
You can give these apps a high-level objective and they will figure out the sub-tasks needed to get things done… and then they will try to figure out the sub-sub-tasks below that. It is sort of like a to-do list that does itself. A video explainer is here.
Although they currently struggle to complete jobs without human intervention, often veering off-task, these projects are significant as proofs-of-concept. (For an example of the brittleness of these systems, have a read of this.)
The two most prominent open-source projects are Auto-GPT and BabyAGI. Auto-GPT, already has 100k stars on GitHub and is the 43rd most-starred project at the time of writing. This is a promising indication of the interest and experimentation within the ecosystem.
The explosion in experimentation around GPT is significant. Not only will it birth a slew of collaboration, interesting applications and new ways of using digital tools, but their open-source model could also serve to democratise AI technology. Just don’t connect the AI to anything important, or stop overseeing your experiments.
Key (AI) reads
Training data drama. The use of training data in large language models has come under scrutiny once again. The Washington Post analyzed Google’s C4 data set, which includes the content of 15 million websites used to train models such as Meta’s LLaMA. OpenAI does not disclose which datasets were used to train ChatGPT. The article uncovered some troubling findings, including the presence of websites promoting white supremacy, Russian state-backed propaganda, and other hate speech proponents. Additionally, the copyright symbol appeared over 200 million times in the data set, raising concerns about potential legal challenges. MIT Technology Review’s Melissa Heikkilä has argued that OpenAI can’t prove that it’s not infringing Europe’s GDPR laws. Further reading: Dean Baker’s take on adapting copyright laws to compensate creative workers and abolish copyright monopolies.
Bigger is no longer better. Sam Altman argues that the race for ever-bigger language models is already over and that new approaches are necessary for the continued development of AI. One major challenge is the shortage of high-quality language data, which a paper found may be exhausted before 2026. This, along with the high cost of computing power, electricity, and skilled labour, makes the development of large language models unsustainable. To move forward, Altman suggests grounding these models in a cognitive foundation through curriculum learning, which can strengthen their mental models. See also: Nvidia’s text-to-video plans.
The horse is out of the stable. StabilityAI has released a collection of open-source language models called StableLM. This is significant because StableLM is not just another language model, but an open-source model that can be used for commercial purposes. According to a statement by Stability AI, developers can freely inspect, use, and adapt StableLM models for research or commercial purposes, subject to the terms of the CC BY-SA-4.0 license. While StableLM models are still being tested and are not performing very well at present, the release of the models as open-source could enable developers to modify and fine-tune their AI models at a fast pace. See also: Alphabet merges DeepMind and Google Brain AI research units, putting Demis Hassabis in charge of the conglomerate’s AI efforts.
Weekly Commentary: The State of the Silicon Valley
I’ve just come back from a week in Silicon Valley where I have been a regular visitor since 1996. It was my first trip since the covid pandemic and I was able to catch up with old friends and meet new ones.
I enjoyed more than 24 hours of conversations with locals. My interlocutors included people who had founded, backed or built some of the most successful Internet and tech products in history (like Sun Microsystems, Juniper Networks, Microsoft Office, and LinkedIn) as well as taken bets on the next set of potential winners (like OpenAI, Commonwealth Fusion Systems, Pachama and Replay Genomics).
I’ve distilled my impressions on the mood in the Valley, including key thoughts on the development of AI as a platform, in a short note. This will go out in my weekly commentary for Members-only on Monday. You can sign-up below if you aren’t a member.
Market data
The patent cliff. The world’s 10 biggest drugmakers stand to lose around 46% of their revenues due to patents expiring at the end of the decade. Expect to see them buy emerging pharma firms to compensate.
What gas crisis? The EU managed to decrease its gas usage by 17.7% between August and March — greater than its target of 15%.
A troubling increase. A record 363 journalists were imprisoned in 2022, up 88% since 2010.
Cash was king. 90% of the UK population uses contactless payment every day.
No one wants to BeReal. Daily app usage in March has dropped 61% since its peak in October 2022.
Russian disinformation network claims social media platforms detect fake accounts only 1% of the time.
Short morsels to appear smart at dinner parties
🔋 Chinese company CATL, the world’s biggest maker of EV batteries, unveils its strongest battery to date.
🍟 Microsoft is working on its own chips to rival Nvidia’s.
📱 Caring about lifespan: why Apple dominates the US Smartphone market.
🪲 Magnolias are much older than bees, so beetles pollinate them.
🧠 The myth of the homunculus. New research shows that we were wrong about how the brain controls different body parts.
🪐 Beautiful images of the Landsat 8 orbiting the Earth, taken by the WorldView-3 satellite. Also, try out this virtual space lift.
End note
We also published the first part of our second chartpack. A chartpack is a deeper dive into key areas of interest for the EV team. This week
looks at the principle of modularity and its economic impact. Recommended.A
What you’re up to — community updates
A brilliant essay from Ludwig Siegele in The Economist argues that AI could change computing, culture and the course of history.
Claudia Chwalisz discusses citizen assemblies and representation by lottery on a recent episode of Hilary Sutcliffe's podcast.
Gianni Giacomelli invites members to test a prototype of the application of generative AI for problem-solving through MIT’s collective intelligence lens.
Rafael Kaufmann presents Digital Gaia’s team position on the only feasible alternative to an inevitable slip towards AI-powered chaos or control.
Martin Varsavsky had an unsolicited software update on his Tesla, which left him stranded and unable to drive home.
Share your updates with EV readers by telling us what you’re up to here.
This didn't appear as an email, is that an oversight, or a problem my end?