I thought it was interesting - issue 2

Continuing with the theme of interesting developments in AI. Recent developments in AI include major product upgrades, new research insights, and even a historic political appointment. Anthropic rolled out transformative features to Claude, including file creation, contextual memory, and privacy controls, while Albania appointed the world’s first AI-powered government minister. Research highlighted both the promise and risks of AI, from reducing nondeterminism in LLM inference to concerns about cognitive offloading. Meanwhile, evaluations of Microsoft 365 Copilot show that while AI boosts satisfaction and accessibility, its productivity gains remain uneven. Finally, Ethan Mollick reflects on the paradox of trusting AI “wizards” that are powerful but opaque.

Latest Developments in AI

Claude Becomes a Full-Featured Work Assistant

Anthropic launched a major upgrade to Claude AI, enabling it to create and edit Excel spreadsheets, Word documents, PowerPoint decks, and PDFs directly within Claude.ai and its desktop app.

Albania Appoints the World’s First AI-Made Minister

Albania has introduced Diella, an AI-generated figure functioning as the virtual Minister for Public Procurement, with the aim of increasing transparency and eliminating corruption in public tenders.

On Working with Wizards — AI’s Power vs. Opacity

In this essay, Ethan Mollick explores a core paradox: the most capable AI tools (“wizards”) are also the least transparent. As we increasingly rely on them in areas where verification is hardest, we move more into a realm of “provisional trust”—accepting “good enough” despite inherent uncertainty. 

Latest Research Findings

The Hidden Risks of Cognitive Offloading

Generative AI’s convenience might weaken human cognitive capacities over time. As firms increasingly offload memory, reasoning, and creative work to AI, potential long-term impacts on human cognition are drawing scrutiny.

Defeating Nondeterminism in LLM Inference

Thinking Machines Lab has published findings explaining how batch-processing variability at inference time contributes to unpredictable LLM outputs. They propose methods for reducing this nondeterminism to improve consistency and reproducibility.

Microsoft 365 Copilot Pilot Report — Mixed Results

The UK Department for Business and Trade conducted a pilot study of Microsoft 365 Copilot from Oct–Dec 2024. Users reported high satisfaction (72% satisfied or very satisfied) and time savings on tasks like drafting and summarising, especially for neurodiverse and non-native English speakers. However, measured productivity gains were inconsistent across different tasks, with some even seeing additional time costs for lower-quality outputs.

Latest AI Tools & Applications

Math Inc.’s Gauss Breaks New Ground in Mathematical AI

Math Inc. unveiled “Gauss,” an AI system that in just three weeks generated 25,000 lines of verified math code, solving a challenging proof sequence. They’re now aiming to scale this up by 100–1000×. 

Question for the Week

With AI stepping into roles as significant as a virtual cabinet minister, offering capabilities like persistent memory and drastically changing the nature of work, what elements of human cognition, trust, and oversight do we need to preserve—and how should we embed them into these systems?