Welcome to Verses Over Variables, a newsletter exploring the world of artificial intelligence (AI) and its influence on our society, culture, and perception of reality.

AI Hype Cycle

AI’s Next Big Leap

We've been witnessing a remarkable AI renaissance. Versatile language models that can create everything from marketing copy to poetry are incredibly impressive. According to AI pioneers David Silver and Richard Sutton (the minds behind AlphaGo and AlphaZero at Google DeepMind), we've been living in what they call the "Era of Human Data," a period in which AI learns by essentially absorbing vast repositories of human-generated content. It's as if AI has been studying humanity's collective knowledge. While this has been effective and revolutionary, Silver and Sutton argue we're nearing the limits of what this approach can achieve, and the next advancement requires something fundamentally different.

What comes next? Silver and Sutton propose we're transitioning into the "Era of Experience." This represents not just an incremental update but a fundamental shift in how AI learns and improves. Instead of passively absorbing static human knowledge, the next generation of AI agents will learn predominantly by doing. They will interact with environments (digital or physical), generate their own data, and learn from the consequences of their actions. This approach builds on Reinforcement Learning (RL), the same methodology that enabled AlphaZero to develop chess strategies that surprised human masters. We've already seen glimpses of this potential with systems like AlphaProof, which earned the equivalent of a silver medal at the International Mathematical Olympiad by training itself and generating millions of practice problems rather than simply studying human-created examples.

According to Silver and Sutton, these experience-driven systems will operate differently in several key ways:

• Continuous Learning: Instead of isolated interactions, these AIs will learn continuously from an ongoing stream of experiences, adapting over long periods, similar to how humans learn throughout life. Consider an AI health coach that remembers your January sleep patterns when making recommendations in July, or a language tutor that recognizes how your pronunciation evolves over months.

• Real-World Grounding: Actions and observations won't be limited to text; they'll be anchored in the real or digital world through software interfaces, sensors, and potentially robots. This creates AI that can interact with systems, not just discuss them.

• Outcome-Based Rewards: Rather than optimizing for human approval, rewards will derive from the environment itself. Did the code execute successfully? Did the experiment produce results? Did the patient's health metrics improve? This "grounded reward" system means AI learns from actual outcomes, not just human evaluation.

• Novel Reasoning Approaches: Experiential AI might develop ways of reasoning and planning that differ from human thought processes. By testing ideas against reality, it can potentially overcome flawed assumptions and biases embedded in our data, leading to powerful and unexpected problem-solving approaches.

The timing seems appropriate for several reasons. We're observing diminishing returns from purely human-data-driven AI. Reinforcement learning techniques have become significantly more powerful to the point where DeepMind has reportedly built systems that use RL to discover their own RL algorithms, outperforming human-created alternatives. Additionally, we're developing AI "agents" capable of more autonomous interactions with digital environments. The foundation is being established for AI to move beyond mimicking human knowledge toward experiencing the world independently. As Silver and Sutton write, "We stand on the threshold of a new era in artificial intelligence that promises to achieve an unprecedented level of ability."

If Silver and Sutton's vision materializes, this shift could unlock remarkable capabilities. Consider AI accelerating scientific discovery, designing innovative materials, or creating truly personalized education that adapts continuously. It represents the potential for AI to address problems previously considered beyond reach. Naturally, this brings new considerations. Agents acting autonomously based on their own experience raise important questions about safety and control. Interpretability could become more challenging if AI develops reasoning styles that differ from human approaches. However, there are potential safety benefits as well. An AI learning from experience is inherently more adaptable to changing circumstances. It might develop the ability to recognize and avoid causing human distress based on real feedback. Additionally, the natural constraints of real-world interaction, especially in physical environments, could provide inherent limitations on rapid, unchecked advancement.

Ultimately, the "Era of Experience" represents a transition toward AI that doesn't simply reorganize human knowledge but generates its own insights, potentially surpassing human capability in specific domains. It marks an evolution from AI as a sophisticated pattern-recognizer to AI as a potential innovator.

Why the AI Revolution Feels Surprisingly Slow

While the tech world is buzzing and executives feel the pressure to incorporate AI yesterday, a lot of the wider public is still figuring out how these powerful tools actually fit into their daily workflows in a truly useful way. Surveys show a weird mix: people are aware of AI, maybe even using tools like ChatGPT occasionally, but there's also significant nervousness and skepticism about its real impact, especially on jobs and society. Workplace adoption is growing, yes, with around 75% of knowledge workers reportedly using AI, often bringing their own tools if the company is too slow. But dive deeper, and you see a lot of that use is still experimental or for fairly brief tasks.

This gap between the astronomical hype (and funding) and the sometimes-underwhelming reality is where a paper by Arvind Narayanan and Sayash Kapoor, titled "AI as Normal Technology," comes in. It suggests that maybe we should think about AI less like an impending alien superintelligence and more like electricity or the internet. Powerful, transformative, but still just technology that follows predictable patterns. One of the paper's core points is that real-world impact doesn't happen overnight just because a new model drops. They distinguish between AI methods (the cool new algorithms), applications (products built with those methods), and adoption/diffusion (people and companies actually using those products effectively). History shows that diffusion, especially for truly impactful tech, is a slow burn measured in decades, not months. Think about electricity: it took ages for factories to truly leverage it by redesigning their entire workflow around production lines, not just swapping steam engines for motors. We're seeing this play out now. AI might ace the bar exam, but that doesn't mean it can handle the nuance, strategy, and client hand-holding of actual legal work. For creatives, it's similar. AI can generate stunning images or draft copy, sure. But integrating that seamlessly into a complex project, managing client feedback, ensuring brand consistency, and maintaining artistic vision? That requires significant human adaptation and workflow changes, which takes time. It's less about the AI's raw capability and more about how we integrate that capability.

The authors also caution against equating benchmark performance with real-world utility. An AI might score off the charts on a standardized test, but those tests often measure exactly what current AI is good at (like information recall) while missing the complex, contextual skills that define professional work. This might explain why, despite incredible AI advancements, many studies show productivity gains are often modest and lean towards augmenting human work rather than wholesale replacement. The most useful AI often works quietly in the background, smoothing things out rather than demanding the spotlight. The "normal technology" view also reframes AI risk and control. Instead of fearing an uncontrollable paperclip-maximizing superintelligence, the focus shifts to managing AI like any other powerful tool. This involves familiar concepts: auditing, monitoring, fail-safes, and robust cybersecurity (like ensuring AI has the 'least privilege' needed). The authors predict that managing, overseeing, and specifying tasks for AI will become increasingly central parts of human jobs. For the time being, let’s focus on more "sophisticated prompt engineering and system oversight,” rather than impending AGI. For most creative pros, this feels intuitive; we're used to mastering complex tools to realize our vision.

If AI is normal tech, the biggest risks aren't necessarily rogue AIs taking over, but rather the amplification of existing societal problems: bias baked into algorithms, increased inequality, job displacement in specific sectors, and the concentration of power. These are the messy, real-world consequences we need to grapple with through thoughtful policy and ethical deployment, focusing on resilience rather than trying to lock the tech down. Viewing AI as "normal technology" doesn't diminish its power or potential. It grounds the conversation. It reminds us that integrating truly transformative tech is a marathon, not a sprint. It requires patience, adaptation, and a focus on building real-world applications and workflows, not just chasing SOTA benchmark scores.

Back to Basics

The AI Index 2025: What’s New in the World of AI

The Stanford HAI 2025 AI Index report just landed, and it's our annual reality check on where this fast-moving field actually stands. The report confirms what many of us feel: the pace of change is intense, and AI is rapidly evolving from a niche tech concept into something deeply woven into our work and lives.

The first big headline from the report is that AI capabilities are advancing fast. The benchmarks designed just last year to really test AI systems are already being aced. We saw huge leaps in scores on complex tests, and AI is also getting seriously good at generating high-quality video (think OpenAI's SORA and Google's Veo 2) and, in some cases, can even outperform human programmers on specific tasks against the clock. That old sci-fi trope of a single, super-intelligent AI is more like a whole ecosystem of specialized and increasingly powerful tools. The Turing Test, once the benchmark, feels almost quaint now. And this isn't just happening in research labs. AI is stepping out into the real world. From healthcare (the FDA approved 223 AI-enabled medical devices in 2023 alone!) to transportation (Waymo’s driverless cars are giving over 150,000 rides a week), AI tools are becoming part of the everyday landscape.

Of course, this progress is fueled by serious investment. The numbers are staggering, especially private funding in the US, which hit $109 billion in 2024, dwarfing figures from China and the UK, though it's crucial to note China is rapidly catching up in terms of model performance. Generative AI continues to attract massive funding globally. But this growth comes with a significant energy cost. AI training is power-hungry, with compute needs doubling roughly every five months. Tech giants are even looking into nuclear power options to sustain this demand, which tells you something about the scale we're operating at.

Looking at the global picture, the US remains a leader in producing the highest number of notable AI models, but the performance gap with other players, particularly China, is narrowing significantly. China also leads in overall AI publications and patents. Europe is making strides, especially in regulation with its comprehensive EU AI Act, and investment is growing. We're also seeing significant AI activity emerging globally, from the Middle East to Latin America and Southeast Asia. Governments are increasingly active, investing heavily in infrastructure and, albeit sometimes slowly, crafting policy. It’s less a single race, more a complex global interplay of innovation, investment, and regulation.

And the most game-changing: AI is becoming more efficient, affordable, and accessible. The cost to use sophisticated AI (inference costs) has dropped dramatically, thanks to more efficient hardware and the rise of powerful, smaller models. Open-weight models are now performing nearly as well as their closed-source counterparts on many benchmarks. This lowers the barrier to entry, making advanced AI tools available to more developers and businesses. The cutting edge is getting crowded, and the performance difference between the very top models is smaller than ever. AI in 2024-2025 is a dynamic mix of rapid advancement, huge financial bets, and critical ongoing challenges. It’s becoming more powerful and integrated into society faster than many predicted. But the vital work around responsibility, safety, bias mitigation, data governance, and equitable access is still in a crucial phase.

Tools for Thought

OpenAI’s Big Week

Keeping up with OpenAI was a full-time job this month, as the team rolled out a flurry of updates that significantly enhance ChatGPT's capabilities and user experience. Here's a rundown of the key changes we've seen emerge recently:

• Meet the New Brains o3 and o4-mini: OpenAI dropped its latest models, o3 and o4-mini, designed to "think for longer" before answering. These models are a major upgrade in reasoning power, as they have an improved agentic ability to juggle all the tools in the ChatGPT toolbox (web search, data analysis, visual understanding, and image generation). During its thinking process, the model combines the tools (and often goes back to think again) to tackle complex problems, often delivering results surprisingly quickly. It's a step towards an AI that actively figures things out rather than just responding.

• Updated Memory: The Memory feature got a useful tweak. ChatGPT can remember not only what you explicitly tell it to remember, but also your past interactions, so it can reference past chats. This should lead to more pertinent results and make interactions feel less like starting from scratch every time.

• Image Library: OpenAI created a dedicated spot to admire your image creations, which lets you easily browse, find, and manage the images you’ve created on their platform (a much-needed organizational tool).

• Rumor Has It: Beyond the official releases, rumors persist about explorations into building their own social network or even a dedicated search engine. While speculative, the potential motivation seems clear: gaining access to the vast, real-time data streams that fuel platforms like X and Meta, which are crucial for training ever-smarter AI. It could be about data, direct competition, or both, but it signals OpenAI is thinking broadly about its ecosystem.

Intriguing Stories

Robots on the Run

Beijing just hosted what might be the most ambitious (or adorably awkward) tech event of the year: the world's first half-marathon featuring humanoid robots attempting to race alongside human runners. The Robot-Man Half Marathon: 21 sophisticated bipedal machines worth millions in R&D dollars, some stumbling, some emitting smoke, all moving with the determination (if not quite the grace) of toddlers learning to walk over 21 kilometers. While human winners blazed through in just over an hour, the leading robot (Tiangong Ultra) required a leisurely 2 hours and 40 minutes, with several battery changes and the constant hovering presence of nervous engineers ready to catch their million-dollar babies before they face-planted on the pavement. The rules acknowledged our current technological reality with refreshing honesty: you can swap batteries mid-race; you can substitute a fresh robot if yours starts smoking (with just a 10-minute penalty); and human handlers can accompany robots like anxious parents watching their child's first bike ride without training wheels. Several competitors experienced what might charitably be called "technical difficulties:" falls, collisions, and the occasional mechanical meltdown that required human intervention. The Beijing robot marathon offers a perfect metaphor for the current state of humanoid robotics: impressive progress that still looks awkwardly mechanical, revolutionary potential that still requires human babysitting, and the delightful contrast between sleek promotional videos and the messy reality of robots learning to master the deceptively complex act of putting one foot in front of another.

Benchmark Bait and Switch

Meta's recent Llama 4 launch has transformed from a triumph to a teachable moment after being caught in what some are calling benchmark sleight-of-hand. When first unveiled, Llama 4's Maverick variant strutted onto the prestigious LMArena leaderboard with an impressive second-place finish, sandwiched between Google's Gemini 2.5 Pro and OpenAI's GPT-4o. But then came the plot twist. Sharp-eyed AI researchers discovered Meta hadn't submitted the standard Llama 4 model that developers would actually get their hands on. Instead, they'd entered a specially tuned experimental version, a model specifically optimized to shine in these particular tests. The aftermath was brutal. When the actual public version of Llama 4 Maverick faced the same tests, it didn't just slip a few positions; it plummeted from superstar status at #2 all the way down to a humbling 32nd place. That's like watching an Olympic gold medal contender suddenly struggle to qualify for the high school track team. Meta's defense came with corporate polish: they were transparent about using an experimental build, they said, and experimenting with variants is standard practice in their development process. This explanation has convinced approximately no one in the AI community, where the benchmark bait-and-switch has been met with the digital equivalent of eye-rolling. LMArena responded by essentially opening their books, releasing over 2,000 head-to-head battle results and updating their policies to prevent similar benchmark theatrics in the future. Their diplomatic statement noted that Meta's interpretation of submission guidelines "did not align with expected standards"—tech-speak for "that's not how this works."

Surf’s Up: OpenAI Catches a Three Billion Dollar Wave

In what looks like a strong endorsement of vibe coding, OpenAI is reportedly in advanced talks to acquire Windsurf. The rumored price tag for this AI coding assistant startup sits around a hefty $3 billion, signaling a major strategic play in the developer tools space and potentially OpenAI's largest acquisition to date. Windsurf builds AI tools designed to help programmers write code faster and maybe even better, kind of like a super-powered autocomplete that understands context. Think of them as a direct competitor to the likes of Microsoft’s GitHub Copilot, Anthropic’s Claude, and Cursor. This deal is especially interesting since OpenAI apparently tried to acquire Cursor first, but that didn't pan out, with Cursor now eyeing a much higher valuation. OpenAI recently launched their own coding models, but acquiring Windsurf gives them instant access to a potentially large user base (Windsurf claims over a million users, though we take such numbers with a grain of salt) and a ready-made distribution channel. It’s not just about selling their models; it's about embedding themselves directly into developers' workflows. Plus, access to all that coding interaction data could be incredibly valuable for training even better models down the line (data is the new oil, or maybe the new plutonium, depending on your level of tech optimism). This whole situation feels like the foundation model giants, realizing that owning the brain (the model) isn't enough; they want to own the hands (the coding tools) too.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up so you never miss an issue.

we’ve also started publishing more frequently on LinkedIn, and you can follow us here

if you’d like to chat further about opportunities or interest in AI, please feel free to reply.

if you have any feedback or want to engage with any of the topics discussed in Verses Over Variables, please feel free to reply to this email.

banner images created with Midjourney.