AI’s Future Faces Data Shortage, Security Threats, and Political Pitfalls

Unwind with our weekend tech brief: From AI hiccups to breakthrough gene therapy, sip, click, and explore tomorrow today!

The Thin Wisps of Tomorrow

We Are Running Out of Data: A new study shows that, following the current trajectory, we will exhaust the available supply of human-generated text to train AI models by 2026-2032. What do we do after that? Synthetic data, transfer learning, and data efficiencies might provide relief.

AI and Politics Don’t Mix Well: Beyond deepfakes and the intentional misuse of AI technology to influence voting behavior and politics, AI chatbots are also poor at providing election-related advice. A recent study found errors in 27% of responses related to the 2024 (US) elections. Meanwhile, Anthropic (the company behind the Claude chatbot) published insights into why that is and what to do about it.

Building With AI: After the excellent “Part 1” of the blog post on building with AI, where the authors explored the tactical aspects of the subject, they are back with Part 2—focusing on long-term strategic considerations and Part 3—which is all about strategy. And if building with AI isn’t enough, here’s a resource that lets you dive deep into programming Generative AIs.

Local AI: Curious about running AI on your laptop (locally and for free)? There are several ways to run LLMs locally (on your computer without an Internet connection), but few are as easy to use as Jan.ai.

0-Days Never Felt So Real: We previously discussed the implications of AI on cybersecurity. Now, a team of security researchers has used GPT-4’s code-generating abilities to autonomously hack into websites by exploiting zero-day vulnerabilities. Better batten down those hatches.

AI Safeguards Are All Too Easy To Overcome: There’s a lot of discussion about AI safeguards—the measures we put in place to ensure LLMs don’t respond with dangerous or toxic content. It turns out it’s not that hard to disable those safeguards. Using an approach called “abliteration,” one can easily uncensor any LLM. So much for that…

New Tech, Old Format: Yes, there is a “Miss AI” contest now. And yes, it is exactly what you think it is. <sigh>

You Are What You Ingest: AI, like the human body (or any living thing), becomes what it ingests. For example, recommending glue on pizza was a massive faux pas for Google Search’s AI feature. Widely reported, Google’s algorithms absorbed this information and reinforced the belief that adding glue to keep the cheese from sliding off your pizza is, in fact, a good idea. Snark aside, it’s a good reminder that we have a long way to go when it comes to safeguarding our AI overlords.

Disengagement Engaged: In world which is characterized by a “war for talent”, this is sobering… here is the full report. Meanwhile, FinTech Klarna boasts about AI taking over the work of 700 people.

Can You Hear Me Now: Gene therapy, a technology that seemed perpetually stuck in the “next year is the year” cycle, is having a moment lately. The latest breakthrough enables children with inherited deafness to hear again.

The Deepfakes Are Real: You need to read this—Charlie Guo, an AI expert, breaks down an intricate Elon Musk-Deepfake scam on YouTube. It’s a glimpse into the future, and a good reminder that we need to get better at identifying these schemes.

First the Moon, Now Rockets: India and its startups are on a roll. After successfully landing on the moon, an Indian startup has 3D printed a rocket engine in just 72 hours. Innovation at its best.

---

What We Are Reading

🤖 How AI Can Make Us Better Leaders The “AI-augmented leader” can leverage both the power of AI and their most human qualities to lead most effectively. But there are no shortcuts; you will still need to do the self-work! @Jane

🤦‍♂️ Toyota Apologizes For Cheating On Vehicle Testing And Halts Production Of Three Models Yikes. Toyota, Mazda, and Honda have all admitted to improper certification testing on various vehicle models, leading to production suspensions and numerous apologies. @Mafe

👬 A Digital Twin Might Just Save Your Life This deep dive into the intellectual history, technical state of the art, and promising future of digital twins suggests we might still find tremendous value at this particular convergence of advancing technologies. @Jeffrey

🌐 Statecraft For The 21st Century Approaching the concept of mission-driven governance and collaboration is a broad undertaking. Reports such as this one provide a great perspective on what it can look like in detail. @Julian

🤯 The Strangest Coincidences Of Your Life Probably Aren’t That Strange At All Coincidences can be explained by simple mathematics, and understanding the basic laws of probability helps to demystify surprising events. @Pedro

👨‍💻 The Misfit Who Built The IBM PC The endearing story of Don Estridge, who built the IBM PC against all odds and organizational resistance, and wasn’t thanked for it. @Pascal

---

The Fun Stuff

If you can build it, someone will use it. Here’s a subway-style map of all major Roman roads. It’s impressive, both the map and the roads!

Here’s a delightful (and powerful) reminder that we often get things wrong: The Pessimists Archive, “a project to jog our collective memories about the hysteria, technophobia, and moral panic that often greet new technologies, ideas, and trends.”

Draw an iceberg and observe how it floats. 🧊