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

DeepSeek and You Shall Find

Now, if you haven't heard of DeepSeek yet, you're about to see why it's making waves. These folks have managed to create an AI model that's 27 times more cost-effective than OpenAI's o1 model. That's not a typo - we're talking about the kind of efficiency gain that makes hardware engineers weak in the knees. It's like someone just figured out how to make an electric car that can drive across the country on a single charge. This kind of breakthrough efficiency is what's driving the excitement around DeepSeek, but it's not the whole story. They're also pushing the boundaries of AI training and championing an open-source approach that's shaking up the industry. (For a deeper dive, our favorite listens were Ben Thompson’s Stratechery and Lex Fridman’s podcasts. Although we warn you, Fridman’s was over 5 hours.)

DeepSeek is a Chinese AI startup making waves by developing and releasing open-source large language models (LLMs). Think of LLMs as the brains behind chatbots like ChatGPT. They're trained on massive amounts of text data and can generate text, translate languages, write different kinds of creative content, and answer your questions informatively. DeepSeek has developed several LLMs, with DeepSeek-R1 currently garnering the most attention.

DeepSeek distinguishes itself in several key ways. Its approach to efficiency, training, and open-source strategy sets it apart from the crowd. They've implemented a Mixture of Experts (MoE) system that works like a highly efficient hospital. Instead of having every specialist examine every patient, they only call in the brain surgeon when there's actual brain surgery to be done. No more wasting computational resources on tasks that don't need them. They also use Multi-head Latent Attention (MLA). Think of it as the Marie Kondo of AI memory management - it keeps only what sparks joy (or in this case, what's immediately relevant for processing). Specifically, DeepSeek-R1 excels at tasks requiring reasoning and "chain-of-thought" processes, demonstrating strong performance in coding, mathematical problem-solving, and general logical reasoning. While detailed benchmarks are still emerging, early indications suggest it's highly competitive with other leading LLMs in these areas.

But DeepSeek's innovations go beyond just architecture. They've also pioneered some fascinating training techniques. One particularly noteworthy approach, highlighted by Thompson, involves using one AI to train another. DeepSeek used R1 to generate synthetic data, which was then used to improve their V3 model. This "AI training AI" method is a significant advancement in how these models are developed. While the exact composition of DeepSeek's training data isn't fully disclosed, it's understood they leverage publicly available datasets. Understanding the specifics of their training data is crucial for assessing potential biases and limitations, an area that requires further investigation.

DeepSeek also made significant progress with pure reinforcement learning (RL). While most LLMs need human feedback (RLHF) to learn, DeepSeek's R1-Zero learned to reason and think through problems on its own. Its "aha moment”—when the model autonomously recognizes a flawed approach, adjusts its strategy, and dedicates more processing power to finding a better solution—demonstrates a capacity for learning how to solve problems more effectively, a key characteristic of intelligence.

Another key differentiator is DeepSeek's commitment to open source. Their models are open-weight, meaning anyone can download and use them, even for commercial purposes. DeepSeek released their models under an MIT license, essentially handing out their secret recipe to anyone who wants to use it, even commercially. In an industry where proprietary technology is guarded like dragon's gold, this is a pretty radical move. It's putting pressure on giants like OpenAI and Llama to rethink their closed-source strategies. This open approach has significant implications for the AI ecosystem, fostering collaboration and accelerating innovation.

DeepSeek's rise also has a geopolitical dimension. Their success highlights the growing competition in the AI field, particularly between the U.S. and China. Thompson even suggests that U.S. export controls on advanced chips might have inadvertently spurred DeepSeek's innovation in efficient AI training.

However, even with DeepSeek's impressive efficiency gains, translating that into scalable and affordable service presents significant challenges. Running sophisticated AI at scale isn't like copying a recipe—it's more like trying to run a five-star restaurant that serves thousands of people simultaneously while keeping prices at food truck levels. While DeepSeek-R1 is more cost-effective than some competitors, the inference costs—the cost of running the model to generate responses—remain substantial, especially for complex reasoning tasks. As evidenced by the experiences of other companies attempting to serve DeepSeek models, scaling inference is a major hurdle. The computational demands of "chain-of-thought" reasoning and the memory requirements of long context lengths can strain even the most optimized systems. Those high margins at other AI companies aren't just about padding the bottom line. Running a massive AI infrastructure is like maintaining a small city's power grid. Sure, DeepSeek's innovations are like installing solar panels and smart meters, but you still need the basic infrastructure, and that doesn't come cheap.

This is a critical area to watch as DeepSeek and others strive to make these powerful models more accessible. DeepSeek has shown us what's possible when you approach AI efficiency with fresh eyes and clever engineering. But the gap between "look what we can do" and "look what we can reliably deliver to millions of users" is still pretty wide. They've given us a glimpse of a more efficient future, but turning that potential into reality? That's the real challenge keeping AI engineers up at night.

Good Enough AI: It's Not Rocket Science (But It Is Pretty Smart)

If you're like us, you've probably seen the headlines. AI is taking over! Robots are writing symphonies, painting like Picasso, and probably plotting world domination as we speak. (Or at least, that's what the movies tell us.) But hold on a second. Before we all start building underground bunkers, let's talk about something a little less apocalyptic and a little more practical. Let's talk about "good enough" AI.

We're not saying the quest for Artificial Superintelligence (ASI) is dead. It's still out there, lurking in the imaginations of futurists and the code repositories of Google DeepMind. But while everyone's chasing the AI unicorn, something quietly revolutionary is happening: AI is getting good enough to be incredibly useful right now. Think of it like this: do you really need a rocket to go to the grocery store? A reliable Corolla might just do the trick.

Benedict Evans hit the nail on the head. He points out that "better" AI doesn't always equal "more accurate." Sometimes, you don't need a super-powered brain; you just need something that gets the answer right. Imagine asking an AI to calculate the number of elevator operators in the US in 1980. A super-intelligent AI might spin you a tale of elevator operators battling rogue toasters in a dystopian future. Fun, but not helpful. What you actually need is an AI that can reliably pull the correct data from a census report. That's "good enough."

This brings us to a crucial point: reframing the problem. We're so used to computers being right all the time, that we try to force LLMs into that mold. But they're not built for precise information retrieval. They're probabilistic, not deterministic. Think of it like this: you wouldn't use a hammer to screw in a lightbulb, right? (Unless you're really committed to DIY.) Instead of asking LLMs to be perfect data miners, we should focus on what they are good at: probabilistic reasoning. They can identify patterns, make educated guesses, and even tell you when they're likely to be wrong. And that, my friends, is a superpower in itself. It’s like having a colleague who says, "I'm not sure about this, but…" instead of confidently giving you the wrong answer.

And just like you wouldn't use a hammer for a screw, you also wouldn't use the same LLM for every task. Choosing the right tool for the job is crucial. There's no one-size-fits-all AI. Some LLMs are better at generating creative text, while others are better at analyzing data. It's like choosing the right chef for the right dish. You wouldn't ask a pastry chef to grill a steak (unless you're into some seriously experimental cuisine).

This isn't to diminish the incredible progress in AI. We've seen mind-blowing advances in image recognition, natural language processing, and even code generation. But the real game-changer isn't necessarily the flashy AI that can write poetry; it's the practical AI that can automate tedious tasks, analyze mountains of data, and free us up to do more interesting things. Think of it as the rise of the "infinite intern," as Evans puts it. Sure, the intern might occasionally get things wrong (who hasn't?), but they can handle the grunt work, freeing you to focus on the strategic stuff.

And that brings us to another key idea: disruption. Early personal computers weren't great at running mainframe applications. But they were amazing at something else: personal productivity. The same is true for LLMs. They might not be perfect replacements for traditional software, but they excel at something new, something we haven't quite fully grasped yet. The "something else" for generative AI remains a topic of much discussion. It's more than just a fancy API call. Using LLMs to simply replicate old software patterns feels like using a Ferrari to deliver groceries. Sure, it works, but you're not exactly maximizing its potential.

Now, some folks are worried. They're saying, "But what about the jobs? What happens when the robots take over?" (Cue dramatic music.) And yeah, there will be some disruption. But history tells us that technological progress usually creates new jobs, even as it displaces old ones. Remember when everyone was worried about computers replacing typists? Now we have web developers, UX designers, and social media managers. (And yes, AI might eventually replace them too, but that's a discussion for another day.)

We're living in a world where AI is getting good enough to make a real difference, even if it's not quite ready to achieve sentience and join us for a game of chess. And honestly, that's probably a good thing. We can focus on building useful tools, solving real-world problems, and maybe even getting a little more sleep at night. The AI revolution isn't about Skynet; it's about spreadsheets that fill themselves, emails that write themselves (sort of), and a whole lot of other "boring" stuff that suddenly becomes a lot less boring.

Back to Basics

Talk to the Hand (and It Will Build You an App)

We've been talking to our computers a lot lately. Like, a lot. And we're feeling a little strange about it. It started innocently enough. "Hey computer, whip me up a landing page for my artisanal dog toy emporium." Next thing you know, we're deep down the rabbit hole of "vibe-based" coding, a term coined by Justine Moore in their recent exploration of the AI-powered web revolution. (Yes, "vibe-based." We're still processing that one, too.) But the more we explore this world of talking to machines and building apps with our feelings (sort of), the more we realize something significant is happening. This isn't just a quirky new trend. It's a fundamental shift in how we create and interact with technology, and it's raising some pretty big questions about the future of development.

The core idea is simple: instead of writing code, you describe what you want to build using natural language. Think of it as having a conversation with your computer. You say, "I want a website that sells artisanal dog toys and has a built-in scheduling system," and the AI figures out how to make it happen. Moore explains that this new approach is creating a new "app stack" – instead of just lists of libraries and frameworks, people are working with pixels and natural language. It's a huge shift, making web development accessible to a much broader audience. As Nick Floats put it, "YOU CAN NOW BUILD APPS AND WEBSITES JUST BY OPENING UP A BROWSER AND TALKING TO YOUR COMPUTER AND THIS IS VERSION ZERO OF THIS TECH, DO YOU NOT REALIZE WHAT'S ABOUT TO HAPPEN?????" (Yes, he used all caps. It felt appropriate.)

But it's not all sunshine and rainbows: there are some real challenges. Debugging can be tricky, integrations can be finicky, and sometimes the AI just gets stuck in a loop of errors. It's like having a junior developer who's enthusiastic but still needs guidance. And It's not just hardcore coders using the tools. (We have tried it several times, and we are definitely not coders.) It's everyday people building personalized apps, developers using AI to speed up their workflow, and even consultants creating websites for small businesses. The examples are fascinating – everything from bedtime story generators to custom finance trackers. And the market is validating this trend. Just look at Cursor. They've become the fastest SaaS company ever to hit $100M ARR, according to Sacra, achieving this milestone in a mere 12 months (compared to OpenAI which took 6 years). While they're focused on AI coding tools rather than direct web app generation, their success speaks volumes about the appetite for AI-driven development solutions.

Looking ahead, we see even more potential. We predict more specialized tools, better integrations, and even more sophisticated design controls. We also think a lot about whether these AI tools will stay separate or become integrated into existing platforms like Figma or Replit. Will design tools start generating functional code? Will coding tools become more conversational? The lines are blurring, and the possibilities are endless.

Ultimately, we believe we're in the early stages of a major shift in how we build and interact with the digital world. And if you're even remotely interested in technology, it's definitely worth paying attention. This isn't just about 'talking to your computer.' It's about unlocking creativity and empowering a whole new generation of builders. The future of development is conversational, it's accessible, and it's happening now.

Tools for Thought

Google’s Big Bang: Gemini’s Updates

What it is: Lately, we have been shocked to find ourselves solely in the Google AI ecosystem, as we’ve been pleasantly surprised with the functionality of the new Gemini models, NotebookLM, and our favorite AI Studio. This month Google dropped a whole constellation of Gemini 2.0 models, and it's like they're trying to cover every possible AI use case. From the Flash model, which is now the default (think of it as the everyday superhero of AI), to the Pro Experimental with a context window so big it could probably hold the entire Library of Congress, they've got a Gemini for every occasion. And if you're pinching pennies (aren't we all?), there's even the Flash-Lite. But the real head-scratcher (in a good way) is the Flash Thinking Experimental. It's like they strapped a GoPro to its brain and let us watch the AI think. And, of course, they all speak multimodal, which means they're not just about text anymore.

How we use it: If you're like us, you were probably using GPT-4o for, well, everything. It was the Swiss Army knife of AI. But then Gemini came along, and, well, let's just say we've switched teams. The speed of Flash is like going from dial-up to fiber optic. The real game-changer is Gemini's capabilities, especially the Flash Thinking Experimental. It's not just giving us answers; it's showing us how it got there. It's like having a study buddy who actually explains their reasoning instead of just handing you the answer key. This is huge for understanding how AI works and building trust in its responses.

DeepResearch

You might be confused (and we wouldn’t blame you) since the creative minds at OpenAI and Google have both named their research models: DeepResearch. However, last week it was OpenAI’s turn to launch theirs (although for now you have to have the $200/month Pro subscription to use it). We’ll summarize our colleague, Steve Smith’s, take on the model’s usefulness.

What it is: Think of it as a super-powered research assistant, fueled by a souped-up version of their "o3" model (which, by the way, we're very excited about). This isn't your grandma's search engine; this is AI that actually thinks (or at least, does a pretty good impression of it). It can handle complex questions, sift through mountains of data (including your own uploaded files – PDFs, spreadsheets, the whole shebang), and spit out a structured report, complete with citations. It's like having a PhD candidate on speed dial, but, you know, without the student loan debt.

How we use it: If you're like us, drowning in data is a daily occurrence. But Deep Research changes everything. Competitive intelligence becomes a breeze; just give the agent a prompt and let it do its thing. Due diligence, financial analysis, legal research are all suddenly manageable. This isn't just about finding information; it's about understanding it. And the fact that it shows its reasoning process is huge for building trust (and making sure it's not just making stuff up). It's still early days, but we're pretty sure Deep Research is going to revolutionize how we work.

Intriguing Stories

OpenAI: New Look, Same Mission (Mostly)?

OpenAI unveiled its big rebrand. And it's not just a nip and tuck, folks. This is a full-on makeover. But it's not a complete departure either. They've managed to walk that tightrope between "fresh and modern" and "wait, is that still OpenAI?" (Which, let's be honest, is a branding challenge for any company, especially one as high-profile as OpenAI).

First, the new typeface: OpenAI Sans. It's that perfect blend of techy and approachable. (Think less "HAL 9000" and more "friendly neighborhood AI"). The fact that they intentionally made it slightly imperfect is brilliant. It's a subtle nod to the human element in a world that's increasingly dominated by algorithms. (It's like saying, "Yes, we're building AI, but we're still human…ish"). The logo evolution is also smart. They kept the "blossom" (or whatever we're calling it), but refined it. Cleaner lines, more symmetrical. It's like they took the old logo and gave it a good dose of Marie Kondo. (Does it spark joy? Yes, but now it's also more streamlined). Reserving it for Research is an interesting move. OpenAI has effectively created a visual hierarchy that elevates their academic work while allowing their commercial identity to evolve.

The human-centric approach is clear. They're not just selling AI; they're selling the idea of AI as a tool for human progress. The natural color palette and the commissioned photography all contribute to this feeling of organic connection. They're trying to make AI feel less intimidating and more human. (Which, considering some of the things AI can do, is probably a good idea).

OpenAI's Super Bowl debut with "The Intelligence Age" was more than just a commercial; it was a calculated move to solidify their position as the AI brand. The pointillist animation, tracing humanity's technological journey from fire to ChatGPT, cleverly positioned AI as the next logical step in human progress. This wasn't just about showcasing AI's capabilities; it was about normalizing the concept, making it feel less like science fiction and more like an inevitable (and even beneficial) part of our future.

The rebrand is clearly tied to OpenAI's massive growth. They're not just a research lab anymore; they're a global phenomenon. (ChatGPT has officially entered the cultural lexicon.) So, they needed a brand identity that reflects that—something that's both professional and accessible, something that can scale. (Because, let's face it, they're not going to stop here.)

— 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.

banner images created with Midjourney.