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

To an AI, Science Is a Straight Line and Fiction Is Origami

When a Large Language Model reads a physics textbook, it’s effectively traversing a flat plain. When it reads a short story, it’s navigating a mountain range. This isn’t a metaphor about "depth" or "soul." It is a literal geometric measurement of how neural networks represent different types of information. A new study from researchers has quantified a structural difference between creative and informational text. Their paper, "Unveiling Intrinsic Dimension of Texts," uses a metric called Intrinsic Dimension (ID) to map the shape of text within an AI’s internal mathematics. The findings confirm a suspicion held by many who work with these models: creative writing is computationally distinct from academic writing.

To understand the study, you have to know how LLMs think. Models convert text into vectors which are lists of numbers that exist in a high-dimensional space. But text doesn't use all that space. It lives on a specific "manifold," a lower-dimensional surface curved within the larger void. The "Intrinsic Dimension" measures the complexity of that surface. It calculates the minimum number of coordinates needed to describe the data without losing information. The researchers found a stark stratification across genres. Scientific and technical texts are geometrically simple, clustering around an intrinsic dimension of ~8. Encyclopedic entries (like Wikipedia) sit slightly higher, around 9. But fiction, opinion pieces, and personal narratives? They spike to 10.5 and above. While 2.5 points sounds small, it is actually massive in topology. It indicates that creative text requires significantly more "degrees of freedom" to represent. Scientific writing is constrained; it follows rigid templates and predictable logical flows. Creative writing wanders. It uses a much wider array of semantic movements to get from point A to point B.

This provides a mathematical explanation for why AI-generated fiction often feels flat. Contemporary LLMs are heavily trained on low-dimensional text (academic papers, news, documentation). They are optimized to minimize loss by finding the most probable path, which naturally gravitates toward the geometric median: the "straight line." Humor, satire, and unreliable narration are high-dimensional constructs. They rely on subverting patterns rather than following them. When a model attempts a joke, it often tries to map a high-dimensional trajectory onto a low-dimensional plane. The result is usually coherent but geometrically "crushed," structurally sound, but semantically lifeless.

We often talk about creativity in terms of magic or intuition. We can also talk about it in terms of topology. Human creativity has a shape, and that shape is measurably more complex.

Back to Basics

Gemini 3 and Nano Banana Pro: When Typography Actually Works

Last week, Google released Gemini 3: their most capable model yet. Two days later, Nano Banana Pro dropped, which was built on Gemini 3, but designed for image generation. AI model releases in late 2025 arrive like meteor showers. OpenAI dropped GPT 5.1 the week before. Anthropic released Opus 4.5 yesterday. Every announcement promises revolutionary capabilities. But the gap between what's promised and what actually changes in your daily workflow can be vast enough to contain a small galaxy.

Gemini 3 is, by most measures, the current state-of-the-art foundation model. The benchmark scores represent the first time any model has reliably completed 10 to 15 coherent logical steps without losing the thread. Previous models, including earlier Gemini versions, would start to drift around step five or six, according to DeepMind CEO Demis Hassabis. The model ships with a 1-million-token context window and handles text, images, video, audio, and PDFs natively. Google is positioning this as the foundation for agentic AI, the kind that can execute complex tasks autonomously rather than requiring constant human steering.

The improvements in reasoning are noticeable but incremental. What I care about, what every designer I know cares about, is Nano Banana Pro. This is the model that finally solves typography. Every previous image model has treated text as decoration. You'd get beautiful compositions with text-shaped blurs, letters that looked almost right until you actually tried to read them. Nano Banana Pro renders actual, legible, typographically coherent text. Not just single words. Paragraphs. Multiple languages. Different fonts, textures, styles. (Try for yourself: ask it to generate an infographic of the solar system in a cyberpunk style, for instance.)

The integration with Gemini 3's reasoning engine means the model understands semantic relationships between objects, not just visual correlation. It's the difference between a star catalog and an orrery. One shows you positions, the other shows you how things move together, how they're gravitationally bound. Nano Banana Pro is starting to understand the orbital mechanics of visual composition.

The original Nano Banana, which dropped in August, went viral for turning selfies into hyperrealistic 3D figurines, and in actuality it was a great image editing tool. Nano Banana Pro is designed for professionals. It supports up to 4K resolution versus 1024px for the original. You can upload up to 14 reference images to establish visual context, Load your entire brand guide at once: logos, color palettes, character turnarounds, product shots. The model maintains consistency it all. The genuinely novel part: it can ground images in real-world data using Google Search. Ask it to create an infographic about how to make cardamom tea, it pulls an actual recipe. Request a diagram of a specific plant species, it verifies botanical accuracy. The shift from aesthetic generation to informational visualization matters because it changes what these tools are actually for. For years, AI image generation has been about making cool stuff. Nano Banana Pro is the first model that feels like it's built for communicating ideas.

But there's a ghost in the machine, and it reveals something important about how native multimodality actually works under the hood. The system gets sticky when switching modalities. I'll ask it to generate an image of, say, a NYC metro map. Great result. Then I ask a follow-up question about the concept in text. The system hesitates, hangs, or tries to route my text query back through the image generator. The model has mass now, and that mass creates inertia. Once you're in visual mode, you're caught in a gravitational well. Getting back to pure text conversation requires genuine effort.

This breaks the promise of seamless multimodality. True multimodality should feel like a conversation where I hand you a photo, you hand me back a paragraph, I sketch something in response. Currently, Gemini 3 feels like it has to shift gears mechanically. You can hear the clutch grinding. What this tells us is that there's still a router making decisions about which model to call, and that router is biased toward whatever modality was last active. It's not actually native. It's stitched together, and the seams show when you switch contexts quickly.

Despite this friction, despite the obvious architectural seams, Nano Banana Pro represents something genuinely new. This is the first time I can use AI image generation as a serious communication tool rather than an aesthetic flourish. The ability to render accurate text combined with high-fidelity object permanence means I can actually design with it. If Google can solve the modality handoff problem, if they can make the switching truly fluid, they'll have the first real AI design partner. Not a tool you use for specific tasks, but something you can genuinely collaborate with across the full spectrum of creative work. That's the promise. Right now we're about 70% of the way there. The image generation has arrived. The conversation infrastructure is still catching up.

Tools for Thought

AntiGravity: Ghost in the Browser

What it is: Released last week alongside Google’s Gemini 3 update, Antigravity is a new development platform that splits the screen into two distinct modes: the "Editor View" for pros and the brilliant "Manager View" for the rest of us. This latter mode replaces lines of code with a simple task board where you direct a squad of autonomous AI agents. You simply type a request like "fix the mobile navigation" or "make the header stick on scroll," and the system dispatches an agent to do the work. It validates its own changes and presents you with "Artifacts" (screenshots and interaction logs), so you can approve the work without ever knowing what a <div> tag actually does.

How we use it:  We aren't developers, so we have been using Antigravity to completely revamp our website without touching a single file. We just open the integrated "Live Browser" window and watch the agents work in real time. It is honestly mesmerizing to see the site transform before our eyes as we type instructions. The best part is seeing the updates in the browser live, effectively turning web development into a creative direction session where we provide the vision and the agents handle the construction.

NotebookLM: The Automated Analyst

What it is: While Google’s NotebookLM has been out for a while now, the team keeps adding capabilities, which continue to make it indispensable for us as a research tool. As of last week, the platform integrates "Deep Research" as a dedicated tool that actively plans and executes web searches to supplement your uploaded notes. It pairs this new brain with the Nano Banana Pro image model to handle the visuals. Users can now upload Google Sheets, Word docs, or PDFs and ask the system to convert that raw data into cited reports, polished slide decks, and infographics. The tool handles the heavy lifting of structuring information and formatting it for presentation.

How we use it: We have pivoted to using NotebookLM as our primary visualizer for dense academic material. We absolutely love the new infographic function. We use it to spin up instant visual summaries for complicated research papers that would otherwise take hours to deconstruct. We are also testing the auto-generated slide decks. The content organization is impressive. However, the design has not quite nailed our specific style yet. We currently use the slides as structural blueprints rather than client-ready deliverables.

ChatGPT Group Chats: The Polite Observer

What it is: OpenAI rolled out Group Chats which allow you to invite up to 20 people into a single thread. An AI Agent also sits quietly in the background, listening to the conversation. Unlike the standard "reply-guy" bot behavior, this version of GPT-5.1 is trained to be "socially aware," as it only chimes in when explicitly tagged or when it detects a clear consensus gap it can fill. The goal is to move AI from a 1:1 consultation tool to a team mediator that can settle debates, summarize conflicting viewpoints, or generate itineraries without anyone leaving the thread. AI as a multi-player experience.

How we use it: To be honest, we are still trying to find the "killer app" for this one. We spun up a few group threads, and while the technology works (the AI is surprisingly good at not interrupting), we aren't fully sold on the workflow yet. It currently occupies a weird middle ground: it lacks the structured threading of Slack and the permanence of a Google Doc. We found it useful for quick, low-stakes brainstorming sessions where we needed an impartial tie-breaker, but for now, it feels more like a fun novelty than a replacement for our existing collaboration stack. We are keeping it in the rotation, but it hasn't replaced our "real" work channels just yet.

ChatGPT Shopping: Just in time for the holidays

What it is: OpenAI has officially turned ChatGPT into a dedicated shopping agent. Released alongside a broader GPT-5.1 update, the new "Shopping Research" feature changes the e-commerce loop. It bypasses the traditional "search, click, tab, repeat" cycle by using a specialized agent to interview you about your needs before scours the web. The system proactively generates a "Buyer’s Guide" tailored to your specific constraints (budget, aesthetic, or technical specs) and presents a comparative analysis rather than a list of blue links. It is rolling out now to all logged-in users with nearly unlimited usage through the holidays.

How we use it: We aren’t, as we haven’t started shopping yet. However, our colleagues have found it quite efficient. The UI is adaptive, and it asked totally different clarifying questions for the gift search. They also loved the interactive rating system that pops up while it searches, allowing you to refine the results in real-time by thumbing up or down initial finds. The final output is a detailed justification for each product, which feels like a significant step up from the "answer engine" listicle format we see in competitors like Perplexity Shopping.

Intriguing Stories

The Genesis Mission: Yesterday, The White House announced the Genesis Mission, it’s plan to turn AI policy into a national industrial strategy. Genesis will wire together federal scientific data, DOE’s national labs, and the country’s biggest supercomputers into a single AI platform for discovery. This will allow foundation models and AI agents to be trained on decades of government science data, pointed at problems like fusion, new materials, semiconductors, and biotech. At the center is what they are calling the American Science and Security Platform, run by the Department of Energy. It will pool national lab supercomputers, secure cloud AI environments, domain-specific models, and even robotic or automated labs, then aim all of that at at least 20 “national science and technology challenges” from advanced manufacturing to quantum information science. Framed explicitly as an Apollo or Manhattan-level effort, Genesis is also a signal to industry. Chipmakers, cloud providers, and frontier-model labs are being invited into a common stack with standardized rules for IP, data access, and security tiers that range from open science data to tightly held national security datasets. We’ll have to watch and see whether this becomes a genuine shared instrument for scientists, startups, and universities, or a highly centralized asset that mostly benefits a small circle of contractors and labs. Either way, it marks a clear shift that the next wave of AI progress is being framed around who controls the compute, data, and workflows that sit closest to scientific discovery itself.

OpenAI’s 5% problem: OpenAI now has two public faces: Sam Altman on the research and compute side, and Fidji Simo running what is essentially the “apps company” built on top of it. Simo’s vision seems like a description for turning a weird, expensive research lab into something that looks and behaves like a product business. Simo’s mandate is clear: close the gap between what the models can do and how much people actually use them. Since joining, she has pushed out Pulse (your calendar-aware AI briefing layer), a jobs and certification platform, and a bigger focus on mental health and safety inside ChatGPT, while also owning the awkward questions around future ads, data use, and content quality. The financial backdrop is brutal. OpenAI is spending staggering sums on compute and reportedly losing billions each year, even as ChatGPT hits around 800 million weekly active users. Note that reporting suggests only about 5 percent of those people are paying for anything at all. Simo’s bet is that if ChatGPT really becomes a team of helpers in your pocket, the willingness to pay will eventually catch up to the hype. Unfortunately, before then, get ready for targeted advertising in your sessions.

GPT Moves into the Lab: OpenAI has released a new paper which gathers 13 case studies across mathematics, physics, biology, materials science, and computer science where the model accelerated scientific reasearch. In several examples, it helped researchers spot patterns, propose mechanisms, or even suggest key steps in proofs that experts later verified and published. In one case, GPT-5 assisted UCLA mathematician Ernest Ryu in solving a 40-year-old optimization problem. Ryu describes the model not as an oracle, but as a restless collaborator that keeps generating new angles while he filters, checks, and stitches together the pieces that hold up. Similar dynamics show up in biology and physics: GPT-5 suggests mechanisms for unexpected immune-cell behavior, proposes follow-up experiments, and helps interpret fusion and black-hole simulations by surfacing hidden structure in complex outputs. The OpenAI for Science team is careful to set expectations. They frame GPT-5 as a tool that can compress weeks of conceptual search into hours, not a system that runs a research program on its own. Every case study emphasizes human verification, literature checks, and traditional authorship, even when the model contributed ideas that changed the direction of a project.

— 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.