Verses Over Variables

Your guide to the most intriguing developments in AI

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

AGI-tation is Coming: Are you Ready to Rumble (with Robots)?

You've likely been hearing the AGI drumbeat getting louder and louder lately. And it's not just Silicon Valley hype-bros trying to pump their crypto bags this time. Turns out, the actual government is starting to sweat a little, too. We recently tuned in to Ezra Klein's podcast where he chatted with Ben Buchanan, formerly Biden's White House AI guru, and let's just say the conversation wasn't exactly a soothing lullaby. The general vibe? AGI, that holy grail of AI (Artificial General Intelligence, a system capable of performing almost any cognitive task a human brain can), is closer than we think – maybe within the next two to three years. Cue the collective "oh, sh**" moment.

Now, don’t roll your eyes - we’ve heard about impending AI doom before. But this time, the urgency feels different. As Klein pointed out, this could be the event horizon, the “before and after“ moment in human history. Think the invention of the printing press, but on espresso.

What's really fascinating (and maybe a little terrifying) is that Uncle Sam is not leading this technological tidal wave. This isn't your grandfather's internet, forged in the crucible of DARPA and national security. This time, it's the wild west of the private sector – a bunch of brainy folks in hoodies fueled by venture capital and a deep-seated desire to, well, change everything. Which brings us to the crux of the issue. Usually, groundbreaking tech like the internet or microprocessors emerged from the intersection of national security and technology development (think Cold War paranoia). But AI is bubbling up outside the government's sphere of influence. This means a different dynamic, and, frankly, a bit of a scramble for the government to catch up and figure out how to manage this beast.

Of course, no discussion of AGI would be complete without mentioning the elephant in the server room: China. The US and China are locked in a strategic competition for AI dominance. The fear is that if China gets there first, they'll gain a significant economic, national security, and military advantage. This creates a kind of arms race, with both countries pushing the boundaries of AI development, sometimes at the expense of safety and ethical considerations. The big question is: how do we maintain a competitive edge without sacrificing our values or triggering a global AI apocalypse? It's a tricky balance to strike.

The takeaway here isn't to panic (though a little healthy anxiety is probably warranted). It's to start thinking, planning, and preparing for a world where AI is far more capable and pervasive than it is today. This means addressing potential job displacement, strengthening cybersecurity, fostering a dynamic AI ecosystem, and grappling with the ethical implications of this technology. We need to understand the relationship between the public and private sectors concerning AI development, safety, and usage. Governments must establish clear guidelines, foster innovation, and address negative impacts on workers and society.

We're not saying you need to build a bunker in your backyard (though, hey, no judgment). But AGI is a real deal, and pretending it's just another overhyped tech trend is like ignoring the asteroid hurtling toward Earth. It's time to pay attention, ask tough questions, and demand responsible AI development. At the very least, it's good cocktail party conversation. Now, if you'll excuse us, we have some serious coding to do...or maybe just watch more sci-fi movies for "research."

Anthropic's AI "To-Do" List: From Paperclip Panic to Policy Power-Ups!

Anthropic, a leading AI safety and research company and the makers of Claude, recently submitted a comprehensive set of recommendations to the White House's Office of Science and Technology Policy (OSTP) regarding a national AI strategy. These recommendations outline a detailed framework for navigating the complex landscape of artificial intelligence. Far from focusing solely on technological advancement, Anthropic's document emphasizes the critical need for proactive governance and a commitment to ensuring that AI serves as a beneficial force for society. This is about responsible innovation, guided by a clear understanding of both the opportunities and the potential risks associated with advanced AI systems.

Anthropic's policy recommendations can be viewed as a comprehensive package of both pragmatic safety measures and forward-looking regulations, boiling down to a few key ingredients. First, a hearty helping of AI safety research. This isn't about stifling creativity, it's about understanding the inner workings of these complex systems, predicting their behavior, and making sure they align with human values. Because, let's face it, an AI programmed to optimize paperclip production at all costs (the infamous "paperclip maximizer" thought experiment) probably isn't going to be a great roommate.

Next up: transparency and accountability. We're not exactly fans of black boxes making decisions that affect our lives, and neither is Anthropic. We need to understand why an AI made a particular choice, and there needs to be someone responsible when things inevitably go sideways. Think of it like building a bridge – you need to know who designed it, who built it, and who to blame when it collapses. And speaking of oversight, Anthropic also calls for the creation of an independent body to keep a watchful eye on AI development. Think of them as the referees in the AI game, ensuring fair play and preventing any accidental offsides.

Finally, and perhaps most crucially, Anthropic stresses the importance of international collaboration. AI is a global phenomenon, and we can't afford to have different countries developing AI according to wildly different standards. Imagine if every nation had its own unique version of the internet – a fragmented, chaotic mess, right? The same logic applies to AI. We need common standards and best practices to ensure that everyone's playing on a level (and safe) playing field.

Now, we know that policy documents can be about as exciting as watching paint dry, but this is about our collective future. It's about ensuring that AI serves humanity, rather than the other way around. Anthropic's recommendations offer a practical, forward-thinking roadmap for navigating the AI revolution – and it's a roadmap we should all be paying attention to. In the meantime, we're off to have a serious conversation with our Roomba about its long-term goals. Just to be safe.

Back to Basics

Deep Research or Deeply Misled? AI’s Fact-Checking Fails.

We're living in an age where AI can write poems, compose music, and even design sneakers. But behind all the digital wizardry lies a peculiar problem: AI has a tendency to, well, make things up. We're talking about chatbots inventing historical figures and image generators conjuring creatures that would make Dr. Moreau blush. Welcome to the world of AI hallucination, where creativity sometimes veers into the realm of pure fantasy.

If you’re like us, you’re probably both amazed and slightly terrified by the rapid advancements in artificial intelligence. We're using AI to write killer marketing copy, design eye-catching graphics, and even compose music. But let's be real, the creative renaissance powered by AI comes with a catch: the ever-present potential for our digital muses to completely lose the plot. And with the rise of Deep Research functions, the stakes are even higher.

So, what exactly is AI hallucination? Put simply, it’s when an AI model generates information that is incorrect, nonsensical, or just plain fabricated. It's the digital equivalent of your friend who swears they saw Elvis at the grocery store, except this friend is a hyper-intelligent algorithm. The concern only grows as we place more trust in AI-driven "Deep Research" capabilities designed to synthesize vast amounts of information.

This happens because current AI models, especially the large language models (LLMs) and generative image models, are essentially sophisticated pattern-matching machines. They are trained on massive datasets, learning to predict the next word, pixel, or musical note based on what they’ve seen before. They don’t actually understand what they’re creating. Think of it like this: they’re really good at mimicking Shakespeare, but they couldn’t tell you the plot of Hamlet if their digital lives depended on it. And if you ask it to do deep research on Shakespeare, it might invent sources or misinterpret existing ones, presenting you with a distorted picture.

The issue is that these models can sometimes overgeneralize or extrapolate from their training data in ways that lead to these factual errors and logical inconsistencies. As the saying goes “Garbage in, garbage out,” and when the data used to train these models includes misinformation or biases, the AI's "creative" outputs can reflect and amplify those flaws. This is particularly worrying when using AI for research, as a single inaccurate data point can lead to a cascade of errors. Imagine relying on AI-generated summaries for your next historical fiction novel, only to discover that crucial details are completely fabricated.

You might think, "Okay, a few historical inaccuracies or bizarre animal creations – no big deal, right?" Wrong. For us creatives, the risks of AI hallucination are very real. This leads to the spread of misinformation and erodes trust in AI-generated outputs, which ultimately undermines the credibility of the creative work and the creatives themselves. Suddenly, you're not just a visionary artist; you're the guy who accidentally unleashed a torrent of digital nonsense upon the world. That "deep dive" you thought you were taking was more like a swan dive into a shallow pool of misinformation.

So, what can we, as responsible (and hopefully not-hallucinating) creatives, do to mitigate these risks? Luckily, there are several strategies we can use to keep our AI collaborators in check.

First and foremost: fact-check everything. Especially when using those "Deep Research" features. Treat AI-generated content like a first draft, not the gospel truth. Verify information with reliable sources, just like you would with any other research. Think of AI as a starting point for your creative process, not the finish line. Speaking of choosing wisely, the good news is that researchers are actively tracking and measuring AI hallucination. Resources like Vectara's Hallucination Leaderboard are incredibly helpful, allowing us to see which models tend to stay grounded in reality and which ones are more prone to flights of fancy. These leaderboards evaluate different models based on their tendency to "hallucinate," providing a valuable benchmark for anyone looking to integrate AI into their creative workflows. And remember, even the "best" models can still hallucinate, particularly when doing deep dives into complex topics.

When engaging in Deep Research with AI, be extra diligent about:

  • Source Scrutiny is Key: Don't blindly trust the sources cited by the AI. Take the time to examine each source individually, assess its credibility, and verify the information presented. Look for established publications, reputable institutions, and authors with expertise in the relevant field.

  • Cross-Reference Everything: Don't rely on a single source, even if it seems trustworthy. Cross-reference the information with multiple independent sources to confirm its accuracy.

  • Go Back to the Primary Source: When possible, track down the original research papers, historical documents, or other primary sources that the AI is referencing. This allows you to examine the evidence firsthand and draw your own conclusions, rather than relying solely on the AI's interpretation.

  • Triangulate your findings: The internet is good at finding consensus, so try to find multiple sources that back up what the AI is telling you.

  • Use Multiple AI Assistants: Run the query through multiple AI assistants and compare and contrast the outputs. If one is wildly off, that will stand out.

Second, choose your AI tools wisely. Not all AIs are created equal; some are better trained and more reliable than others. (Perplexity’s research tool will always cite its sources, so this is a great way to eliminate hallucinations.) Third, maintain a healthy dose of skepticism. Don’t blindly accept everything an AI tells you. Ask questions, challenge assumptions, and use your own critical thinking skills to evaluate the output. It’s okay to be a little paranoid; in this case, it’s actually responsible.

Ultimately, the future of AI in creative fields depends on our ability to use these tools responsibly. By being aware of the potential for AI hallucination and taking steps to mitigate its risks, we can harness the power of AI to enhance our creativity without sacrificing accuracy or integrity. 

Demystifying AI Chatbots: Karpathy's Two-Hour Tour of the Tools You Need to Know (Yes, It's Worth It)

Let's cut to the chase: if you're trying to figure out what all the fuss is about with AI chatbots, and frankly, which ones are actually useful, Andrej Karpathy's latest video is your new best friend. We know, two hours sounds like a marathon, but trust us, this isn't your typical dry tech lecture.

For those who might not know, Karpathy isn't just some armchair AI enthusiast. He's a heavyweight in the field, with stints at Tesla and OpenAI, and he's basically the godfather of Stanford's legendary CS231n computer vision course. So, when he talks AI, we pay attention. And what he's delivered here is less of a deep dive into theoretical AI and more of a practical, hands-on walkthrough of the current chatbot scene.

We were expecting something super technical, too. But Karpathy surprised us with a refreshingly accessible exploration of the tools we're all hearing about. He breaks down the key differences between platforms like ChatGPT, Gemini, Grok and others, not by drowning us in jargon, but by showing us real-world use cases. Think of it as a guided tour of the chatbot zoo, with Karpathy as your knowledgeable (and slightly irreverent) zookeeper.

He doesn't just tell you what these tools can do; he shows you what they actually do, and how he uses them in his daily life. We're talking everything from managing context windows (the chatbot's short-term memory) to understanding when to use "thinking models" versus the faster, more straightforward ones. He explores the fascinating world of tool integration, where chatbots are learning to access real-time information and perform complex tasks. And it's not just text anymore. Karpathy shows us the rapidly evolving landscape of multi-modal inputs, explaining how these tools handle audio, images, and even video. He demystifies the process, making it clear that we're witnessing a genuine shift in how we interact with technology.

What's particularly refreshing is Karpathy's focus on practical applications. He doesn't just wax poetic about AI's potential; he shows us how to use these tools to enhance our productivity, streamline our research, and even have a bit of fun. He compares the different platforms, pointing out their strengths and weaknesses and giving us a clear sense of which tools are best suited for different tasks.

So, if you feel overwhelmed by the sheer volume of AI chatbot news or are curious to see what these tools can actually do, carve out a couple of hours ands video. It's a valuable primer on the current state of AI chatbots, delivered with the clarity and insight that only a seasoned expert can provide. Consider it essential view’ essential viewing for navigating the rapidly evolving world of conversational AI.

Tools for Thought

Sesame: AI That Actually Sounds Human

What it is: Let's face it, AI voices have historically been... well, robotic. Sesame is determined to drag AI kicking and screaming out of that uncanny valley. This AI startup is making waves with its remarkably realistic voice assistant, featuring two distinct personalities, Maya and Miles. We're talking about a conversational speech model that incorporates natural human imperfections - breath sounds, laughter, even the occasional "um" or self-correction. It's like they bottled up genuine human conversation and stuffed it into an AI. Forget the monotone drone; Sesame's aiming for "voice presence," that magical quality that makes you feel like you're actually connecting with someone. It's the difference between talking to a GPS and talking to a friend.

How we use it: We've felt a little uneasy that we end up talking to our computers as much as we talk to our dogs. But Sesame? It's different. We're using the demo on their website to explore the possibilities of truly natural-sounding AI. Imagine customer service bots that don't make you want to tear your hair out, or educational tools that actually engage students with a personable AI tutor. With plans to open-source their models and develop AI glasses, Sesame is positioning itself to be a major player in the future of human-computer interaction. Sure, forming emotional attachments to AI might sound a bit Black Mirror-ish, but we're intrigued (and maybe a little bit thrilled) by the potential.

Google AI Mode: Search Gets Smart

What it is: Sometimes Google Search feels like an endless scroll of semi-relevant links, and the introduction of AI Overviews didn’t give us much confidence. However, Google AI Mode is here to change that, throwing a custom version of Gemini 2.0 at your complex queries. This is a whole new search experience where you get comprehensive answers, not just a list of suggestions. It's all thanks to their "query fan-out" technique, which is basically Google unleashing a swarm of mini-searches to gather all the info and synthesize it into one coherent response. It's like Google finally decided to graduate from link aggregator to actual research assistant. Or put another way, if you are used to using Perplexity, you’ll recognize the user experience.

How we use it: We just got access to it (there is a waiting list on Google Labs).  We suspect that it will be particularly useful for complex, multi-part and multi-modal questions where you need nuanced answers. Sure, it's still in its experimental phase, but it's a glimpse into a future where search engines actually understand what you're looking for. Whether it will fully rival the other AI search services like Perplexity AI, time will tell, but we think this will change search as we know it.

Google Shopping with AI: Say Goodbye to Bad Fashion Choices

What it is:  Google Shopping just dropped a bunch of AI-powered tools that are straight out of a digital makeover montage: virtual try-on for clothing, an AI "vision match" feature that lets you describe your dream outfit and then finds shoppable items that match, and AR beauty tools that let you virtually try on makeup looks inspired by celebs and trends. It's like they took the best features of a styling app, an AR mirror, and a personal shopper and mashed them all together with Google's AI magic.

How we use it: We've been playing around with the "vision match" feature, and it's surprisingly effective. Describing a "boho-chic maxi dress with floral embroidery" yielded some genuinely appealing results (though we're still holding out for Google to nail down exactly what "boho-chic" means). The AR beauty features are fun for experimenting with new looks without committing to a full face of makeup, and the virtual try-on for clothing is a game-changer for online shopping. Google Shopping's AI makeover might not eliminate all fashion faux pas, but it's a big step in the right direction.

Duck.ai: Get your Questions in a Row…Privately

What it is: Feeling a little paranoid about feeding all your thoughts and questions into an AI chatbot? DuckDuckGo, the privacy-first search engine, has you covered with Duck.ai. This free AI chat tool promises to be the anti-surveillance option in a world of data-hungry AI. The service includes anonymized chats, requests proxied through DuckDuckGo servers to hide your IP address, and local chat history stored on your device, not theirs. Plus, it lets you choose from multiple AI models like GPT-4o mini, Llama 3.3, and Claude 3 Haiku. Think of it as having a super-smart AI assistant without sacrificing your digital anonymity.

How we use it: We've been experimenting with Duck.ai to see if it truly lives up to its privacy claims, and so far, it seems legit. It's great for those random, potentially embarrassing questions you'd never want tied back to your Google account. Need to know the best way to get red wine out of a white rug? Duck.ai can help, without judging you or adding you to a targeted ad list for stain removers. Sure, there's a daily query limit, but for those moments when you need AI assistance without the data harvesting, Duck.ai is a welcome alternative. It's like having a trustworthy friend who happens to be an AI.

Manus AI: China’s Autonomous Agent (Raising Eyebrows)

What it is: Manus AI is making waves as a potential game-changer in the world of artificial intelligence. Touted as the first fully autonomous general AI agent, Manus is designed to operate independently, tackling tasks from start to finish without constant human intervention. Think of it as a collection of highly skilled AI models working together seamlessly in the cloud to plan, execute, and deliver results across a wide range of areas. From sorting resumes and managing social media accounts to planning travel itineraries and performing complex financial analysis, Manus aims to be a versatile and powerful assistant capable of handling intricate real-world tasks. While shrouded in some mystery regarding the company behind it, its development is seen as a significant step forward in the global AI race.

How we use it: Currently, Manus is in an exclusive, invitation-only beta phase, fueling significant hype and even leading to the resale of access codes for exorbitant prices. While most of us can't get our hands on it just yet, the potential applications are vast. Imagine a world where entire workflows are automated, from lead generation to customer support, all orchestrated by a single AI agent. Early adopters praise its ability to execute complex tasks autonomously, hinting at a future where humans can offload significant cognitive labor to AI. However, concerns regarding data privacy, security, and ethical implications remain paramount. Some are comparing Manus to Deepseek, at least when it comes to the implications for privacy and China’s dominance in AI.

Intriguing Stories

Witness the AI Singularity…in Dial-Up?

Last week, the internet buzzed with a viral video showcasing AI agents discussing wedding reservations. The seemingly innocuous interaction took an unexpected turn when the bots, recognizing each other as digital entities, switched to a series of beeps and boops. This new language was dubbed “Gibberlink.” Suddenly, it felt like we had stepped back into the 90s, only instead of illicit MP3s, we were privy to the dawn of a machine-exclusive language.

The video begins with two AI assistants communicating in English, one of whom inquires about a hotel wedding reservation. Upon recognizing their shared nature, they transition to “Gibberlink,” a series of coded sounds and signals incomprehensible to human ears. This protocol, developed by Boris Starkov and Anton Pidkuiko, is designed for efficient and error-proof data transmission between unconnected AI systems, even in challenging environments.

The emergence of AI-to-AI communication methods like Gibberlink presents several essential considerations. Its potential for enhanced efficiency in information exchange between AI systems is clear. This could lead to faster processing speeds and improved performance across various AI applications. However, using languages inaccessible to human monitoring also raises concerns regarding transparency and control. Developing appropriate oversight mechanisms will be crucial as these technologies evolve. Exploring the full spectrum of potential applications, both beneficial and detrimental, of AI systems communicating independently is also of paramount importance.

While this incident has attracted considerable attention, it aligns with ongoing research into optimized AI communication methods. Initiatives like Microsoft’s “Droidspeak,” which enables AI agents to share intermediate computational representations, demonstrate the broader trend toward more efficient communication strategies. The "Gibberlink" demonstration underscores the need for continued research and ethical deliberation within the field of artificial intelligence.

Ultimately, this incident highlights a fundamental truth: AI development extends beyond building sophisticated chatbots. It requires a comprehensive understanding of the complex and rapidly evolving ecosystem we are creating. After all, we don’t want the future to be a constant stream of dial-up noises we can’t decipher.

When Tech Meets Haute Couture: GPU Purses

We stumbled across something today that made us question whether we’re living in 2025 or some cyberpunk novella—GPU purses are now a thing. Yes, actual handbags with graphics processing units embedded in them. These transparent acrylic fashion statements showcase decommissioned GPUs suspended within as if they were rare artifacts from a digital archeological dig. My personal favorite sports a functional cooling fan because nothing says "I've arrived" like a handbag with active thermal management. The flagship "H100" model costs a mere $65,000—perfectly reasonable if you consider it both a fashion accessory and an investment portfolio. We’re pretty sure for that price it should be able to render our entire digital existence in real-time while simultaneously mining enough cryptocurrency to pay for itself.

— 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.