Last Week in AI

News in AI from the week of 28 April

May 05, 2025

Welcome back for week three of last week in AI. This has consistently gotten the most readership, so I'll provide as much as I can every Monday. Please let me know what else you’d like to see.

This week, From Good Tech / Bad Tech

Models & Capabilities

The drive for bigger, faster, and more specialized AI models continues. While capabilities advance, so do concerns about security and real-world utility.

New Models, Familiar Players: Amazon rolled out Nova Premier, Microsoft unveiled the smaller but supposedly potent Phi 4, and Baidu launched models aiming to undercut competitors on price. The focus remains on performance metrics, but the tangible benefits for everyday users are less clear.
Niche Focus: DeepSeek doubled down on math with Prover and Prover-V2, while JetBrains targeted developers with its Mellum coding model. Specialization marches on, but does it solve fundamental problems or just create more complex tools?
Exponential Growth... or Hype? Claims that AI can handle tasks twice as complex every few months fuel the hype cycle. Meanwhile, Nvidia's new tool for turning 3D scenes into AI images offers another avenue for generation, but raises questions about creative ownership and the value of synthetic content.
Security Theater & Lingering Doubts: Meta's LlamaFirewall aims to patch AI vulnerabilities, a tacit admission of the risks inherent in these systems. Researchers also found weaknesses in AlphaFold 3, a reminder that even celebrated AI breakthroughs aren't infallible. Amidst speculation about DeepSeek's R2, whispers suggest OpenAI might be retiring GPT-4, perhaps signaling a shift or simply the rapid obsolescence cycle in AI.
Share

Industry and Business

Corporations are betting big on AI, forecasting astronomical revenues and pushing adoption, but the actual impact on workers and the focus on genuine human needs remain questionable.

Meta's Pivot: Abandoning the metaverse narrative, Meta is now all-in on AI. Boasting 1.2 billion Llama downloads, hosting LlamaCon to woo developers away from OpenAI, launching an API, and projecting $1.4 trillion in AI revenue by 2035 – it's a full-blown AI gold rush. But will this translate into better products for users or just more targeted ads and engagement loops?
State-Sponsored AI & Automation: China pushes for technological self-reliance, funneling investment into AI and robotics, seen in Shoucheng's car park robots and Ant Group's talent hunt. Xi Jinping's visit to Shanghai's AI incubator (SMC) underscores the national priority.
AI Everywhere? From Toyota and Waymo's autonomous vehicle plans (complete with a potential 'AI social network' for cars) to 20-foot construction robots, vehicle prototyping via digital twins, AI for climate-stressed supply chains, and even an AI 'care companion' in Danish hospitals, AI is being pitched as a solution for everything. The question remains whether these applications genuinely improve lives or primarily serve corporate efficiency and automation goals.
The Elusive AGI: Even as practical applications spread, the dream (or nightmare) of Artificial General Intelligence persists, with small European startups joining the chase.

Environmental and Financial Cost

Building the infrastructure for AI demands staggering resources, raising serious environmental concerns and questions about the sustainability of this technological path.

Trillions in Investment, Looming Constraints: Estimates suggest AI infrastructure spending could hit $8 trillion, while mega corporations like Microsoft already foresee capacity shortages. The build-out these companies say is necessary requires immense power, prompting efforts like Google funding electrician training.
Searching for Efficiency: While some startups explore building AI without data centers, others look to technologies like photonics or "AI Native" processing to optimize. Debates continue over whether AI's energy use is a real crisis or manageable, but the sheer scale of investment suggests significant environmental impact.
Chip Wars & Geopolitics: The hardware itself is a battleground, with Nvidia raising alarms about Huawei's capabilities, highlighting the geopolitical stakes intertwined with AI dominance.

Ethics and Societal Impact

As AI integration accelerates, the gap between utopian promises and unsettling realities widens, raising urgent questions about security, privacy, fairness, and the very nature of work.

New Avenues for Fraud & Attack: Realtime deepfake fraud is here, AI is poised to become a potent exploit coder, and AI-generated code risks poisoning the software supply chain through vulnerabilities like 'package confusion'. Nation-states like China are reportedly using AI to sharpen cyberattacks.
Ethical Minefields: An unauthorized AI persuasion experiment on Reddit users sparked outrage, highlighting the ethical vacuum in some AI research. Concerns about inherent model bias, the potential for AI to corrupt scientific peer review, the fairness of popular AI benchmarks, and the reliability of AI in critical human-in-the-loop scenarios (like medicine, where doctors fail to catch AI errors) persist.
Regulation vs. Reality: While companies like Anthropic push for tighter chip controls (clashing with Nvidia's interests), the legal system grapples with fundamental questions, like whether AI-generated output constitutes protected speech. Meanwhile, government adoption plows ahead, sometimes clumsily (SSA's training video), sometimes controversially (deploying AI agents via DOGE recruiters or tasking college students with AI-driven deregulation), and sometimes raising surveillance alarms (NYC's predictive subway cameras).
Jobs, Hype, and Normalization: Is AI revolutionizing work or just shifting tasks? Economists suggest generative AI isn't replacing jobs or hurting wages yet, and studies indicate time saved is often offset by new AI-related work. While some call for treating AI as "normal" technology (MIT Tech Review's take), the hype continues (tracked by the AI Hype Index), particularly around AI in software development, which brings both potential gains and significant risks. (MIT Tech Review also covered stereotypes and coding).

Research & Development

Fundamental research pushes AI into new domains, from pure mathematics to the search for life beyond Earth.

New Frontiers: DARPA wants AI to accelerate mathematics, while roboticists create swarms that mimic materials.
Scientific Discovery: AI is being used to 'explain' the behavior of sticky proteins linked to disease and even deployed as an AI 'scientist team' in the SETI effort. The focus is often on capability, with less discussion on the wisdom or long-term consequences of these applications. (Getting Real About AI Processors paper also discussed hardware).

Overall Sources for the week-

Good Tech / Bad Tech