Unit ⑧ Letter #20249
We decided to open up to our wider network and share a selection of our resources with our colleagues and followers. 🦾
Dear Unit ⑧ friends,
We have been internally exchanging readings for a while, sharing links relevant to our development as a small organization. We are looking for meaningful directions and keen to keep up-to-date in the realm of technology from a technical, political, ethical, and artistic point of view. Here are some links to the articles we paid attention to these past weeks; we hope you enjoy them!
Google Outlines Plans to help You sort Real Images from Fake
Differentiating which images are authentic and which are fake in this misinformation age we live in, is becoming ever more challenging. The Verge reports on a new technology developed by Google based on C2PA's authentication standard. By integrating C2PA metadata into its search results, this technology can distinguish between images captured by cameras, edited by software, or generated by AI models. This initiative, part of the Coalition for Content Provenance and Authenticity (C2PA), aims to create a technical standard, as well as enhance transparency and trust in digital imaging across platforms like Google Search and YouTube.
What Credit Card Networks can Teach us about Stablecoin Opportunities
In her Substack post, Alana from Back of the Envelope discusses the transformative impact of stablecoins on the financial system. She points out similarities between stablecoin and credit card networks in payment orchestration and market structure. Since stablecoins enable efficient, low-cost, rapid cross-border transactions, they are poised to revolutionize payment systems. In the future, Alana predicts that every financial institution will be able to issue its own stablecoin, and each will operate differently based on issuance processes, reserves, and regulatory framework.
Do Worker Bees need Copilots?
Runtime reports about Microsoft's latest updates to its Copilot feature in Microsoft 365, a new version that includes new features to enhance productivity. Microsoft's request to subscribe to its Copilot service seems aggressive, and, therefore, they have to convince potential business customers of their value, which comes up to a $30 per user per month rate for Copilot on top of existing Microsoft 365 subscriptions.
Council of Europe AI Convention vs the Act
In one of its entries, the EU AI Act Newsletter reports on the current state of the EU AI legislation. The drafting of the General-Purpose AI Code of Practice is attracting nearly 1,000 expressions of interest globally to work on it. Meanwhile, the hiring practices for a lead scientific adviser role within the European Commission are generating some controversy. Lastly, they report on the inaugural AI Board meeting, comparing it with the Council of Europe's AI convention.
OpenAI’s Strawberry
OpenAI has launched its last language model, "Strawberry" (o1-preview). The model employs a technique known as Chain of Thought (CoT), making the assumptions and logic used in calculations explicit. Some of the Substackers we follow reported about it from their own particular perspectives, and here are their views:
Scott Cunningham uses a hypothetical example based on research findings and a choice experiment to highlight Strawberry's Chain of Thought's capacity to break down complex problems into transparent steps. This enhances the model's utility in research contexts where understanding the reasoning process is crucial. Strawberry, unlike ChatGPT-4o, reveals its reasoning process rather than providing just final answers, aiding researchers in performing and understanding complex analyses.
Gary Marcus notes that, despite advancements in AI, like GPT-o1, critiques about limitations and the path to achieving Artificial General Intelligence (AGI) persist among experts. The initial excitement around GPT-o1's release and its capabilities was considerable, yet the limitations of current AI techniques compared to achieving AGI were even more significant. GPT-o1 is still dependent on synthetic data, which has critical implications and raises skepticism regarding the dominance of large language models (LLMs) in AI research.
Michael Spencer points to OpenAI’s significant financial losses–projected to be between $3.5 billion and $5 billion in 2024 due to high operational costs–while acknowledging that the company continues to grow and aims to capitalize on its leading position in AI technology. The Strawberry language model is part of OpenAI's strategy to secure substantial investment and maintain its market position despite financial challenges. The company is negotiating a funding round that could value it at $150 billion, involving significant investors like Apple and Nvidia.
Zvi Mowshowitz discusses the advancements and limitations of GPT-o1 compared to previous models and explains how formal reasoning tasks function and what they entail. At a technical level, the advancement is substantial, particularly in formal reasoning tasks, though it falls short of being the entirely accurate 5-level model that was initially speculated. GPT-o1 leverages an extensive Chain of Thought (CoT) and increased token usage from GPT-4o, allowing it to deal with formal logic and reasoning gracefully. These are the types of skills that would enable one to excel in competitive programming, achieving high scores comparable to human competitors. However, the gains in creative or less structured tasks are not impressive.
In Other News
Google DeepMind trained a Robot to Beat Humans at Table Tennis
We may not be taken over by AI robots in the near future, but we might get beaten by them at table tennis. Using a combination of computer simulations and real-world data, Google DeepMind has successfully trained a robotic arm to play table tennis at an amateur competitive level against human opponents of varying skill levels. The robot won 13 out of 29 games against human opponents, primarily at beginner and amateur levels. MIT Tech Review reports on such a specific AI technology implementation because it represents a significant step toward creating robots capable of performing complex tasks in real-world environments. The table tennis robot continuously improves its capacity to handle spinning balls quickly, and its collision detection is enhanced thanks to real-world data collection and simulation feedback, which remains exceptionally challenging despite the remarkable milestone achieved.
Ape Unit Blog
We spoke with the art collective terra0, composed of Paul Seidler and Paul Kolling. After working on their vision for many years, they have just reached a new stage in which their project will scale up. For the conversation, which we summarized here, they told us about their journey, challenges and plans. If you want to read the full interview, follow the link to our blog post.