🔮 Open-source AI surge; UBI surprises; OpenAI losses; Murdoch’s empire drama, waste-eating flies & learning ++ #484

An insider’s guide to AI and exponential technologies

Ideas of the week

Meta’s Gambit. Llama 3.1’s release has narrowed the gap between open-source and closed-source AI, matching proprietary models on key benchmarks.¹ Zuckerberg champions open-source AI, displaying altruism while executing a shrewd corporate strategy. Meta employs a “commoditise your complement” approach, a strategy to make LLMs more generic and accessible. Core AI models would be freely available which reduces their market value — and that of their competitors — while positioning Meta to profit from the surrounding ecosystem of tools, platforms and services that rely on these models. But can open-source models sustain this progress? Open-source must innovate. Open-source must scale. Open-source must find sustainable revenue. As AI progresses through two more generations, costs will soar (for more on this, see my essay on the scaling ceiling). Closed-source models can justify these expenses through the promise of direct monetisation, whereas Meta’s open-source strategy relies on monetising adjacent offerings. Will this be enough?

Source: Maxime Labonne

See also:

• A good overview on the state of Chinese AI. They are innovating significantly in computational efficiency, maximising AI performance under US-imposed hardware constraints.

• You can now sign-up to test OpenAI’s SearchGPT – its Perplexity rival.

New studies on UBI. A comprehensive universal basic income (UBI) experiment in the US, giving $1,000 a month to 1,000 low-income individuals for three years, yielded underwhelming results: slightly reduced work hours with no significant improvements in health or human capital investments. It did, however, increase generosity among recipients — spending on others saw the largest categorical increase. 

Source: OpenResearch

This isn’t good news for UBI, at least in an American context – no large economic or social benefits were seen, suggesting that a more targeted approach for dealing with wealth inequality may lead to better outcomes. A contrasting 11-year study in Bangladesh demonstrated the potential of targeted interventions. By providing assets (cows) to 6,000 extremely poor households, researchers identified a threshold above which families could transition to better occupations and escape poverty. Targeted approaches that address specific issues relating to poverty and social mobility are likely to be more effective than cure-all efforts. Yet, to accelerate technology transitions, we will need a way to stabilise social and political concerns. I discuss UBI’s potential for this in my recent essay.

Invisible hands, invisible policies. In today’s interconnected global economy, America’s open market unwittingly imports foreign industrial policies, reshaping its industries by default rather than design. As the “global consumer of last resort,” the United States absorbs excess global savings, necessitating persistent trade deficits and suppressing domestic savings rates. This economic position effectively imports foreign industrial strategies, silently restructuring American industries—particularly in manufacturing—without explicit consideration of domestic needs. The US now faces a critical juncture: it must choose whether to forge its own comprehensive industrial strategy, acquiesce to foreign economic blueprints, or radically redefine its role in the global economy. Each path carries profound implications for the country’s economic future and its position on the world stage.

MathBot gets silver. Fields Medalist Terence Tao’s vision of AI as a highly-effective mathematical co-pilot (as highlighted in EV#478) has been validated by DeepMind's latest breakthrough. AlphaProof and AlphaGeometry 2 solved four out of six problems at the International Mathematical Olympiad, achieving silver-medal performance. These AI systems tackled complex algebra, number theory and geometry problems. Sir Timothy Gowers, an IMO gold medalist, Fields Medal winner, and mathematics professor, lauded the system’s non-obvious problem-solving approach. This achievement, using reinforcement learning and formal mathematical language, aligns with Tao’s prediction of AI becoming a powerful assistant for mathematicians. 

AI Ouroboros. While LLMs continue to devour web-scraped data, they’ll increasingly consume their own digital progeny as AI-generated content continues to flood the internet. This recursive loop, experimentally confirmed, erodes the true data landscape. Rare events vanish first. Models churn out likely sequences from the original pool while injecting their own unlikely ones, warping performance and breeding bias. To break this echo chamber, we must safeguard human-created data, trace content origins, and curate diverse, top-notch training sets in the AI-saturated world.

---

Data

OpenAI is projected to lose $5 billion this year – expected, given its high Azure compute costs.

Students show a 17% performance drop after LLMs are taken away.

Tesla’s Q2 automotive gross margin fell to 15% (excluding regulatory credits) – its lowest level in 10 years.

A 20-year-old solar panel was found to be capable of retaining 90% of its original capacity.

Amazon lost over $25 billion selling devices (Echo, Kindle, Fire TV, etc.) from 2017 to 2021.

Ireland’s data centres used more power than all its urban homes in 2023.

---

Short morsels to appear smart at dinner parties

🤾‍♀️ What is sport?

📰 Turns out Succession was a documentary. The battle at the heart of the Murdoch family empire is real.

🚢 Global shipping routes are still very vulnerable. Even Russian vessels, which were meant to have safe passage, are targeted by Houthi militants. 

🌌 I revisited David Deutsch’s 2005 TED Talk recently:

🧨 One firm solved its own Crowdstrike cybersecurity crisis with a simple barcode. Snap smarts!

🫠 Google gained exclusive access to Reddit. I’m hearing echoes of the Balkanisation of the internet…

🪰 Genetically engineered flies can help manage organic waste.

---

Preview

While away, I asked my long-time EV friend Mike Maples to mind the shop for a day. Mike is a renowned Silicon Valley investor whose portfolio includes Twitter, Twitch and Lyft. Together, we exclusively bring you Chapter 6 of Mike’s new book. This chapter looks at how “living in the future” is a key skill of breakthrough entrepreneurs.

1. There is debate about whether it is truely open-source or not - e.g. there are restrictions on usage and model naming (you must have Llama-{xyz} in the title).