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

Are We Optimizing Away the Magic?

My earliest encounters with AI weren’t with chatbots. It was with the image generation tool MidJourney. I was still figuring out how to use Discord, let alone how to prompt images there. However, as I got more comfortable, I also got more adventurous. I would purposely throw in abstract prompts, and sometimes I’d actually forget to type in the right parameters. The results were strange and fun. Some images were warped, others shined with accidental brilliance. What struck me wasn’t polish but unpredictability. The system wasted time and effort, and inside that waste appeared surprise.

The field is now moving in the opposite direction. A new survey, Speed Always Wins, catalogs how researchers are cutting down the Transformer to reduce complexity: mixtures of experts that activate only a handful of specialized sub-networks, hybrids that mix strategies, even diffusion-based approaches for parallel generation. The benefits are clear: faster models cost less, run on smaller hardware, and respond instantly. Still, I think back to those early MidJourney images. The original model was extravagant. Every token interacted with every other token. It was computationally indulgent, but that excess created room for odd connections. Efficient models prune the tree before it grows. They decide in advance which interactions matter. That makes the output smoother, but it also narrows the chance for surprise.

Mixture of Experts is one of the biggest trends in efficiency in the newest AI models. Imagine a party full of specialists: a poet, a lawyer, a coder, a comedian, a chef. In a traditional model, everyone talks at once, which takes a lot of energy but also creates surprising crossovers. MoE changes that dynamic. The system has a router that decides which two or three guests get to speak at each moment. It’s clever, because the model can be enormous on paper, but only a tiny slice is active at a time, which saves huge amounts of compute. The trade-off is that you rarely hear from the whole crowd. The conversation gets narrower. The unlikely combination (say the chef and the poet bouncing ideas that inspire something strange and beautiful) happens less often. Efficiency is gained, but some of the serendipity is lost. Artists have long understood the value of excess. A painter piles on layers of paint only to scrape half of them away, leaving unexpected textures behind. A photographer embraces a light leak that turns a mistake into something iconic. Creativity often comes from what first looks like waste.

Efficiency also changes safety. Waiting for MidJourney to render gave me time to linger on results, laugh at oddities, and decide what to try next. That pause created reflection. Faster systems erase it. They flood us with output and smooth over hesitation. People begin to trust the speed itself, mistaking quickness for certainty. And acceleration carries sharper risks: false information spreads instantly, automated scams refine themselves in real time, markets twitch under the pressure of algorithms that never slow down. In many parts of life we add friction deliberately. We set speed limits, build confirmation boxes, enforce waiting periods. We do it because friction buys us time to think. In AI, we are erasing it.

The alternative is not to glorify inefficiency but to use it where it matters. Models could move quickly when the stakes are low and slow themselves when meaning is at stake. They could keep a few redundant pathways alive. They could let a less likely expert have a voice now and then. Efficiency carries scale, but inefficiency preserves surprise.

Those early MidJourney sessions taught me something. The images were imperfect, but they shimmered with possibility. Speed spreads what already exists. Slack creates the space for something new. If we optimize too tightly, we may end up with outputs that are flawless but hollow. The next time we celebrate a breakthrough that saves milliseconds, it’s worth asking whether those milliseconds once held the spark that made the machine feel alive.

Back to Basics

The Hidden Cost of Curiosity

The first time I heard someone claim every ChatGPT query "boils a glass of water," I laughed and then immediately pictured a river drying up, because I had just asked what to cook for dinner. It's one of those lines that's too sticky to ignore: a little absurd, and exactly the kind of thing that spreads faster on Twitter than actual data. That's been the entire pattern with AI's energy story. The metaphors always outrun the measurements. Depending on which paper you stumbled across, one AI answer might use less energy than a light bulb blinking once, or enough to power your coffee maker for a solid minute. Both could be "true" because the studies were mostly educated guesswork stitched together with model sizes, hardware specs, and wishful thinking about efficiency.

Even the companies played along with the mystery. Sam Altman tossed off a number earlier this year (0.34 watt-hours and a sip of water per ChatGPT query). Reassuring if you're building a business case for AI adoption, but he conveniently left out what he actually measured. Meanwhile, Mistral showed up with the academic playbook, peer-reviewed study with the French environmental agency, and landed at 1.14 grams of carbon and 45 milliliters of water per response. Same basic activity, completely different universe of impact. Then Google did something radical in our opacity-obsessed industry: they actually measured their own stuff in production. They instrumented their Gemini fleet and counted everything: the accelerators doing the heavy lifting, the CPUs feeding them data, the idle machines sitting around in case traffic spikes, and the cooling systems humming constantly in the background. Their result: the median Gemini prompt in May used 0.24 watt-hours, emitted 0.03 grams of CO₂, and consumed about five drops of water. Less energy than watching nine seconds of TV.

Here's the part that floored me: a year earlier, that same prompt was thirty-three times more expensive in energy. Forty-four times more in emissions. When tech companies talk about efficiency gains, they usually mean shaving off a few percentage points per year. Maybe doubling performance if they're feeling ambitious. But this was a total rewrite of the physics.

The reason is efficiency as obsession. Engineers have been quietly rewriting the rules while we argued about the old ones. They're using mixture-of-experts models that only wake up the neural network pieces they actually need, like having a newsroom where only the reporter and editor show up for a breaking story, not the entire staff. There's speculative decoding, where a smaller model drafts an answer before the big one checks it, like having a junior writer take the first pass. Distillation creates lean serving versions from hulking research models, turning a PhD thesis into a crisp executive summary. Even the data centers joined the efficiency party, squeezing their overhead so low that there's barely daylight between total power and actual compute power.

What all this engineering wizardry shows me is that the footprint of a single prompt has become laughably small. Watching ten seconds of Netflix outweighs it. Brewing your morning coffee demolishes it. But here's where scale gets interesting: billions of queries a day turn those drops into rivers, which is why these efficiency gains aren't just nice-to-have metrics for Google's sustainability reports. They're the main story. I don't buy the doom-and-gloom version where AI is an energy monster devouring the grid, and I definitely don't buy the Silicon Valley shrug that treats energy like a solved problem. The truth sits somewhere more interesting than either extreme: a single query won't sink the grid, but it does leave a trace. The real question is whether we keep shrinking that trace as usage explodes.

Tools for Thought

Runway Game Worlds: Interactive Storytelling Reimagined

What it is: Runway is introducing Game Worlds, a new beta feature that transforms users into authors of their own interactive adventures. Forget static, pre-written narratives; this tool lets you create and play through non-linear stories where the world and its characters are generated by AI in real-time. With both a text-heavy "Chat Mode" and a more visual "Comic Mode," Game Worlds is designed to be an intuitive platform for crafting dynamic experiences. It's a step toward a new kind of gaming where your choices actively shape the universe around you.

How we use it: The creative possibilities are vast. You can jump into one of Runway's preset worlds, from a high-stakes heist in "Last Score" to a historical mystery in "Gallic Storm," or you can build your own from scratch. For writers and game designers, this is a powerful tool for prototyping interactive narratives and exploring complex storylines without writing a single line of code. As you play, the AI generates images that bring the story to life, and with future updates promising even richer visuals and real-time video, the line between playing a game and directing a movie is about to get very blurry.

Google’s AI Mode: Your New Personal Assistant

What it is: Google is supercharging its AI Mode with powerful new agentic capabilities, transforming it from a helpful assistant into a proactive personal agent. Now, your AI can handle complex, multi-step tasks like securing restaurant reservations, booking a wide array of services (from salon appointments to car repairs), and even snagging tickets for events, all on your behalf. This expanded functionality is powered by the sophisticated Project Mariner, seamless partner integrations, Google's vast Knowledge Graph, and the comprehensive accuracy of Google Maps, making it a globally available, one-stop solution for managing your life with unprecedented ease.

How we use it: We're playing it safe and haven't handed over the keys to our calendar just yet, but we've been watching early adopters dive in headfirst. They’re using it to completely offload their planning, asking the AI to find and book a table at a top-rated Italian spot in West Hollywood for a Saturday night, and it just gets it done. We've seen it tackle entire weekend itineraries, from snagging tickets to a show at the Hollywood Bowl to lining up a brunch reservation in Santa Monica the next morning. For now, we’re using it as a hyper-intelligent suggestion engine. It’s brilliant for surfacing options and comparing services without the hassle of endless searching, laying out the perfect evening plan and leaving us to make the final click. It’s like having a world-class event planner on standby, even if you’re not quite ready to let it run the show.

Intriguing Stories

Meta is Betting on Midjourney to Fix Its AI Glow Up:  Meta and Midjourney are now officially in the same sentence, which is the kind of crossover event that feels like Marvel and A24 deciding to co-produce a film. On paper, the deal makes sense: Meta gets to infuse its AI products with Midjourney’s aesthetic brilliance, while Midjourney earns validation as the darling of the creative AI world. The result, at least according to the press releases, is that billions of people will soon experience “beauty” at scale. We’re of two minds here. On one hand, Midjourney has been one of our favorite creative tools since its early days. The thought of an API that makes its technology more accessible across platforms is exciting. Imagine a pipeline where Midjourney’s signature style slips effortlessly into other apps, workflows, and production environments. That’s the kind of integration we’ve been waiting for. On the other hand, Meta is still Meta. Which means the same company that optimized your feed into a dopamine slot machine could soon be remixing your aesthetics into another layer of monetization. Privacy concerns linger, as do the inevitable “sponsored experiences” that might sneak their way into anything Meta touches. It’s one thing to create with Midjourney in a focused, ad-free space. It’s another to watch that magic get packaged into feeds that feel more like billboards than galleries. Still, there’s reason for cautious optimism. Midjourney remains independent, self-funded, and deeply rooted in its community. That spirit doesn’t just vanish overnight, even if it’s suddenly licensed to a tech giant. If Meta can resist the urge to turn beauty into banner ads, this collaboration could open up incredible opportunities for creators and audiences alike.

Big Tech Can’t Quit Each Other: Big Tech has a way of circling back on itself, and the latest AI rumors read like a mash-up of unlikely alliances. Apple, the company that spent years touting its privacy-first approach, is reportedly in talks with Google to inject Gemini into a future version of Siri. At the same time, OpenAI, the outfit often painted as Google’s existential rival, has been quietly leaning on Google Search data to fuel ChatGPT’s answers. The Apple piece is especially curious. Siri, which has long lagged behind ChatGPT and Gemini (and Alexa), has struggled to keep up with the pace of AI assistants. Apple’s own models haven’t impressed, and after years of trying to keep things in-house, the company appears to be shopping around. Google’s Gemini is a natural candidate in terms of scale and technical muscle, but it’s an odd fit culturally. Apple’s brand has been built on clean lines and tight control, while Google thrives on sprawling experiments and data-hungry ecosystems. A Siri powered by Gemini would represent Apple quietly admitting that AI is too big to build alone. Meanwhile, across the valley, OpenAI has been borrowing from the very company it’s supposedly challenging. According to The Information, OpenAI has tapped SerpApi to scrape Google Search results, using that material to help ChatGPT answer questions about news, sports, and the markets. Google doesn’t grant OpenAI direct access, but the end result is the same: ChatGPT competes with Search while being propped up by it. In past training runs, OpenAI also relied on YouTube data without Google’s blessing. Apple may end up outsourcing part of Siri’s brain to Google, while OpenAI continues to siphon Google’s knowledge base to fuel its chatbot. Isn’t it ironic.

— 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.