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.

We’ll publish intermittently until the end of the year and then return to our regular cadence in 2025.

AI Hype Cycle

Process over Polish: The Rise of Beautiful Imperfection

You know that feeling when you bite into a homemade cookie, and it's slightly imperfect – maybe a touch too crispy on the edges, or the chocolate chips aren't perfectly distributed? That's exactly what we're going to be craving more of in our AI-saturated future. Not the cookie (though I wouldn't say no), but that beautiful imperfection that screams, "A human made this."

I've been turning this over in my mind for months, but it really crystallized during a recent webinar with Sari Azout, founder of Sublime, which explores the intersection of technology and human curation. As I watched AI seep into every corner of our digital lives, something about her perspective clicked – that perfect missing piece that helped me understand what I'd been seeing. We've rocketed from "wow, AI can do that?" to "yeah, yeah, another AI-generated thing" in record time. But as the novelty wears off, something far more interesting is emerging in its place: a renewed appreciation for the distinctly human elements of creation.

Remember math class – that perpetual frustration when teachers insisted we "show our work" instead of just writing down the answer. Those teachers were architects of a future they couldn't have imagined. In a world where AI can spit out perfect answers in milliseconds, the journey to the solution – the messy process, the false starts, the moments of inspiration – that's becoming the real treasure.

We're entering an era where provenance – the origin story of ideas and creations – matters more than ever. It's not just about what was made, but how it came to be. Those pencil sketches in an artist's notebook aren't just preliminary work anymore; they're proof of the human journey. The early novel drafts, complete with crossed-out passages and margin notes? Pure gold. The process of your favorite creator, whether they're a musician, writer, or scientist, holds compelling power beyond their end results. Understanding their influences, their struggles, their moments of doubt and breakthrough – that's what draws us in. AI might be able to mimic their style perfectly, but it can't replicate their unique path to discovery.

I'm seeing this shift already. Developers are starting to document their thought processes more thoroughly, not just their code. Designers are sharing their mood boards and iterations, not just the final mockup. Scientists are making their lab notebooks public, complete with dead ends and failed experiments. Social media is full of “get ready with me” videos and threads of creators trying out new tools. (Ironically, even AI tools now flaunt their 'process' with Canvases, Artifacts and Reasoning – a synthetic mimicry of human transparency.) We're all beginning to realize that our humanity isn't in the perfection of our output – it's in the beautiful mess of our process.

The irony isn't lost on me: as AI gets better at producing polished results, we're learning to value the unpolished, the raw, the real. We're rediscovering that creativity isn't just about the destination; it's about the journey, complete with wrong turns and unexpected discoveries. This doesn't mean we should reject AI or its capabilities. Instead, we need to recognize that AI is pushing us to articulate and celebrate what makes us uniquely human. Our value isn't in our ability to compute faster or generate more content – it's in our capacity to think differently, to make unexpected connections, to bring our lived experiences and emotions into our work.

So here's what I'm betting on: in the coming years, we'll see a renaissance of process-sharing. Creators will invite us into their minds, showing us not just what they made, but how they got there. We'll develop new ways to authenticate and value the human journey of creation. Because in a world of artificial perfection, our beautiful mess of humanity becomes the most precious thing we have to share.

The World (Model) Is Not Enough: AI's Quest to Understand Everything

Let's talk about the hottest trend in AI that's getting everyone from Meta's bigwigs to ambitious startups hot under the collar: World Models. We're witnessing what feels like a digital gold rush, with tech companies large and small staking their claims in this fascinating frontier. (And no, we're not talking about America's Next Top Neural Network here.)

Think of World Models as AI's attempt to understand reality the way we humans do - messy, complex, and full of surprises. Unlike their more specialized AI cousins (looking at you, ChatGPT), World Models aim to develop a general understanding of, well, everything. It's like trying to teach a computer to have common sense, which anyone who's ever watched their smart TV struggle with basic tasks knows is harder than it sounds. Meta's AI Chief has been particularly bullish on this front, suggesting World Models could be our ticket to human-level AI within the next decade. Though given the tech industry's track record with timelines (remember when self-driving cars were "just two years away"?), we're keeping our expectations cautiously optimistic. (Some of the World Models announced this month include DeepMind’s Genie 2 and Worldlab.

What makes World Models special is their ability to learn through observation and prediction, kind of like how we learned not to touch hot stoves as kids (though hopefully with less painful trial and error). They combine visual, auditory, and textual information into one cohesive understanding - imagine if your smartphone's camera, microphone, and autocorrect actually worked together instead of seemingly plotting against you. This multimodal approach helps address some of AI's most embarrassing current limitations. You know how language models sometimes confidently spout complete nonsense? World Models aim to ground these systems in reality, reducing those facepalm-worthy "hallucinations" we've all come to know and mock. They're also designed to handle physical environments better than current AI, which often has all the spatial awareness of a caffeinated puppy in a china shop.

The economic implications have certainly caught investors' attention. World Labs (clearly winning at the creative naming game) has already secured a cool $230 million in funding. Meanwhile, tech giants like Meta, Google, and OpenAI are throwing resources at World Models like they're going out of style - which, given tech trends, they probably will be right after they become successful.

Now, we'd be remiss not to mention the challenges ahead. Building World Models requires computational resources that would make a supercomputer blush, and we're still trying to figure out how human cognition works well enough to replicate it. Not to mention the ethical considerations - we've all seen enough sci-fi movies to know we should probably think this through carefully.

World Models represent a fascinating shift in how we approach artificial intelligence. Instead of building specialized tools for specific tasks, we're attempting to create systems that understand and interact with the world more holistically - like humans, but hopefully with better memory and fewer coffee dependencies.

Back to Basics

Artificial Intelligence, Natural Resistance

This week, the worlds of graphic design and AI finally collided in public (and yes, we are feeling a bit conflicted about it, too). When Paula Scher's team at Pentagram – the design equivalent of getting Meryl Streep to direct your home movies – embraced AI for a website project, it felt like watching your favorite vinyl-only record store start selling MP3s. The whole situation mirrors that seismic moment when Photoshop first crashed the graphic design party in the '90s, complete with declarations that "real designers use X-Acto knives." Yet here we are, with digital tools forming the backbone of modern design (though we still love the smell of rubber cement in the morning).

Scher's team's approach to AI is not unique – deploying it not as a replacement for human creativity, but as a digital intern. They started with handmade elements (traditional Scher territory) and then fed these into Midjourney, essentially teaching the AI to speak their design language. The result: a custom-built tool that could riff on their ideas at scale. The practical reasoning is crystal clear. When staring down the barrel of creating 1,500 icons for a government website, using AI starts looking less like selling out and more like smart resource management. It's the difference between hand-writing 1,500 party invitations and using a printer – though admittedly, a printer that sometimes gets creative with the font choices. She defended her team’s use of Midjourney by stating that “The whole notion of the site was to correct (government bureaucracy) by creating a site that could run all by itself.”

The delicious irony comes from Paula Scher herself, whose career has been built on the kind of hands-on, intuitive design work that seems antithetical to algorithmic generation. (Her fingerprints are all over modern visual culture, from Citibank and The Public Theater to Shake Shack, MoMA, and Tiffany's.) It's like finding out your favorite chef now uses an InstantPot – not necessarily wrong, but definitely surprising.

The response from the design community has been about as mixed as a playlist combining death metal and smooth jazz. Some see it as a betrayal of craft, while others view it as an inevitable evolution. This mirrors broader conversations across creative fields – writing, music, visual arts – creating a sort of anxiety disco where everyone's dancing to different beats of the same drum. The heart of this debate lies in how we define "making" in an age where our tools are increasingly capable of independent generation. Scher's argument that design is fundamentally about creating plans rather than executing every element by hand strikes at the core of our industry's evolution. The essence of design shifts and morphs as our tools advance, yet the fundamental need for human vision remains constant.

We'll admit it – seeing Pentagram's Midjourney process gave us a bit of validation (it's our favorite image generation tool too). Their approach mirrors what we have been doing in the trenches, and there's something satisfying about seeing it elevated to the mainstream. But the copyright implications linger like a bad hangover, and we can't help but think Pentagram could have framed this milestone with more transparency and foresight. In the end, maybe that's the real lesson here: it's not just about how we use AI, but how we talk about using it.

Tools for Thought

Last month, the AI doomers were forecasting that the end was near for AI, and that AI had hit a wall. This week, the holidays have come early, as OpenAI and Google (and many others) have been dropping updates and new tools like a tech Santa on a caffeine high. We can't cover all the new drops, so let's dive into the highlights.

OpenAI’s Holiday Hustle: OpenAI has transformed the usually quiet December tech landscape into a festive AI showcase they're calling "12 Days of Shipmas" - running from December 5th to December 18th. Imagine if Santa's workshop went full silicon valley: every weekday, they're unveiling a new AI breakthrough like clockwork. Unfortunately, each demo comes chock full of dad jokes.

• Day 1: The curtain rose with a double feature that had us spilling our morning coffee - the full release of the o1 reasoning model and a new ChatGPT Pro subscription at $200/month. The Pro subscription is an all-access pass to GPT-4o, Advanced Voice Mode, and a souped-up version of o1. At that price point, it better make me coffee too.

• Day 2: OpenAI dropped something that had developers buzzing - reinforcement learning fine-tuning capabilities. Think of it as giving AI models a personal trainer for specific tasks. The techies got particularly excited about this one since it means they can finally customize models for niche applications without starting from scratch.

• Day 3: Enter, Sora, the text-to-video generator that's been living rent-free in our minds since its announcement. The demand was so intense they had to temporarily close the velvet rope to new users.

• Day 4: Canvas for ChatGPT got a major glow up. They've thrown open the gates to all users (no more VIP list!), and you can now run code right inside. You can also use Canvas in custom GPTs. It's like they took our workflow wishlist and checked off everything at once.

• Day 5: The Apple Intelligence integration announcement dropped, and suddenly the tech ecosystem feels a bit more connected. Unfortunately, as 100 million iPhones came online, ChatGPT went down. Turns out even AI needs a breather sometimes.

Google’s Big AI Energy: Not to be outdone, Google stepped into the ring with their own impressive lineup of AI releases. And these updates deserve some serious attention.

• Gemini 2.0: Google kicked things off with a bang, unleashing their most advanced AI model yet. Designed for the "agentic era," it's like they've given AI a graduate degree in getting things done.

• Gemini 2.0 Flash: Their experimental model that's somehow running twice as fast as its predecessor while performing better on nearly every benchmark. We're talking about a 1-million token context window (that's entire-book territory), the ability to handle everything from images to video in real-time, and benchmark scores that make other models look like they're running on a calculator from 1995.

• Project Astra: This is Google's take on a universal AI assistant that actually knows its way around the Google ecosystem. Imagine having a digital concierge that seamlessly works across Search, Lens, and Maps.

• Project Mariner: This Chrome extension gives your browser an AI copilot that can navigate the web better than we can after 20 years of experience.

• Deep Research: This feature in Gemini Advanced is basically the research assistant we wished we had in grad school. It dives into complex topics with the tenacity of a caffeinated PhD student, but without the existential crises.

Amazon’s AI Arsenal: While OpenAI and Google may have been stealing the headlines, Amazon just quietly dropped some serious AI firepower at re:Invent 2024. Between all the announcements this month, two of Amazon's releases stand out as genuine game-changers. First up is the Nova family of foundation models. The range starts with Nova Micro for basic text processing and scales all the way up to Nova Premier (coming early 2025) which handles everything from text to images to video. These models are running at least 75% cheaper than their competitors, and they support over 200 languages.

The real showstopper is Project Rainier - Amazon's upcoming "Ultracluster" supercomputer. Built in collaboration with Anthropic (after an $8B investment), this beast will pack hundreds of thousands of Trainium2 chips and deliver more than five times the computing power used to train Anthropic's current models. When it comes online in 2025, it'll be less like a computer and more like a digital sun - generating enough AI processing power to make today's models look like pocket calculators.

Meta’s Efficiency Play: Meta just pulled off something genuinely impressive with Llama 3.3. They've managed to shrink their 405-billion-parameter beast into a 70-billion-parameter package without losing any punch. That's like getting the power of a sports car in a compact - same performance, way less fuel. The new model handles eight languages with ease, from English to Spanish to Hindi, making it a serious contender for global applications. By slimming down the computational requirements, Meta's essentially democratizing high-end AI. Small businesses that couldn't afford to run the bigger models can now play in the same sandbox as the tech giants.

Anthropic’s Style Game: While everyone else was busy unleashing new models, Anthropic took a different approach, focusing on making Claude more personable - or rather, as personable as you want it to be. They've rolled out custom styles for Claude, letting users shape how their AI communicates. The update brings three preset flavors - Formal (for the suit-and-tie crowd), Concise (for the get-to-the-point folks), and Explanatory (for when you need things broken down step by step). And you can now create your own custom style by uploading writing samples and specific instructions.

Midjourney tries World-Building: Midjourney just dropped Patchwork - an infinite canvas where up to 100 people can simultaneously build entire fictional universes, complete with characters, locations, and storylines. The killer feature is their "scraps" system - smart sticky notes that generate images in ten different styles. Users simply drop a story beat, describe a character, or sketch out a location, and watch as AI art brings it to life. The creative chaos of 100 simultaneous world-builders adds to the magic, turning individual imagination into collective storytelling. With future plans for video generation and personalized models, Midjourney's betting big on collaborative creativity.

Photo Meet Physics from Worldlabs: World Labs unveiled its latest innovation: a system that transforms ordinary photographs into interactive 3D environments. Led by AI pioneer Fei-Fei Li, the team has developed technology that goes beyond simple 3D rendering to create explorable virtual spaces – imagine stepping into a photograph and walking around inside it. The system lets users navigate these transformed environments in real-time, adjusting camera angles, tweaking lighting, and even modifying object colors. While the current output has a stylized aesthetic with some exploration limitations, World Labs emphasizes this is just an initial preview of the technology's capabilities.

Navigating Reddit’s Hive Mind with AI Precision: Reddit’s new feature, Reddit Answers, is here to redefine how we engage with the platform’s sprawling sea of communities and conversations. This AI-powered tool acts as your personal guide, plucking meaningful answers from Reddit’s endless threads and distilling them into concise, easy-to-digest insights. Think of it as a bridge between Reddit’s raw, user-generated wisdom and the streamlined efficiency of AI. Unlike generic web searches, Reddit Answers pulls from moderated, genuine user content. This means you’re not just getting a bot’s take—you’re getting the hive mind of Reddit itself, curated and summarized. Currently in beta for select U.S. users, Reddit Answers is available via the web and iOS app, with plans to grow its linguistic and geographical footprint.

We’ll be talking about our favorite tools, but here is a list of the tools we use most for productivity: ChatGPT 4o (custom GPTs), Midjourney (image creation), Perplexity (for research), Descript (for video, transcripts), Claude (for writing), Adobe (for design), Miro (whiteboarding insights), and Zoom (meeting transcripts, insights, and skip ahead in videos).

Intriguing Stories

NotebookLM Team Writes a New Chapter: The powerhouse team behind Google’s NotebookLM has officially embarked on a new adventure. Raiza Martin, the former team lead, along with designer Jason Spielman and engineer Stephen Hughes, announced their departure to launch a stealth-mode startup aimed at redefining AI’s role in everyday life. Martin revealed that a shared belief in the potential for transformative innovation within the AI space is their main driver. The team’s new project is shrouded in secrecy, with only a minimalist website hinting at what’s to come. While details remain sparse, the group’s history with NotebookLM—famed for its AI-generated podcasts and advanced note-taking tools—hints at their ability to create user-focused, viral solutions. In comments to TechCrunch, Martin emphasized their commitment to building practical, consumer-facing AI products. She highlighted the need for tools that not only showcase the latest in frontier models but also integrate seamlessly into the lives of everyday users. The startup is in its infancy, with no funding announced yet, but early support from investors, academics, and fellow entrepreneurs has been strong, signaling significant interest in the venture’s direction. Given the team’s track record of turning cutting-edge technology into accessible, viral products, this new startup has the potential to influence the AI landscape. And let’s be honest, we loved NotebookLM. With its recent updates, it has become an indispensable tool, so we can’t wait to see what comes next.

Your Selfie’s Secret Bodyguard: Meet Chameleon, from Georgia Tech's research team - a new AI system that generates privacy-protecting masks for photos - essentially teaching one AI to shield your images from other AIs that might try to recognize faces in them. The magic happens through a P3-Mask—a personalized pattern that makes your photos look totally normal to us humans while sending AI recognition systems into a confused tailspin. Unlike other privacy tools that handle your photos with all the grace of a bull in a china shop, Chameleon's approach is more like a master illusionist, making subtle tweaks that preserve your aesthetic while thoroughly bamboozling those nosy recognition algorithms. In an age where facial recognition systems are becoming as common as coffee shops, Chameleon could be a game-changer in preventing everything from identity theft to unauthorized AI training. The researchers are already looking ahead, exploring ways to keep your photos from becoming unwitting participants in training the next generation of AI models.

Perplexity Ponders Hardware: Perplexity AI, the conversational search engine, is toying with the idea of creating an affordable voice-activated device. CEO Aravind Srinivas floated the concept of a "simple, under $50" gadget designed to provide reliable voice-to-voice answers. He suggested that if his post garnered over 5,000 likes, Perplexity would proceed with the project—a threshold that was swiftly surpassed. This move aligns with a broader trend among AI startups venturing into hardware to foster new user interactions. For instance, Midjourney formed a hardware team in August, and OpenAI's Sam Altman is collaborating with ex-Apple design chief Jony Ive on an AI hardware initiative. However, the hardware landscape is fraught with challenges, as evidenced by Rabbit's R1 device, which, despite selling around 130,000 units, faced feature delivery delays and now sees steep discounts on resale platforms. Similarly, Humane's Ai Pin, a wearable AI device, suffered from poor reviews and safety issues, leading to recalls and a search for an acquirer. Perplexity, reportedly nearing a $500 million fundraising round, certainly has the financial muscle to explore hardware. Yet, as history shows, success in this arena requires more than just capital.

— 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.

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.

we hope to be flowing into your inbox once a week. stay tuned for more!

banner images created with Midjourney.