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 our perception of reality.

AI Hype Cycle: The Founders Talk Their Books

The AI Optimist: From Anthropic’s Dario Amodei

Not to be outdone by Sam Altman, Dario Admodei, the CEO of Anthropic, decided to pen a 10,000-word essay called “Machines of Loving Grace.” In it, he argues that truly powerful AI could compress a century's worth of human progress into just 5-10 years, potentially transforming our world in ways we've only dreamed of. At the heart of Amodei's thesis is the idea that AI could significantly accelerate scientific discovery, particularly in fields like biology and neuroscience. He envisions AI systems that don't just analyze data but actively design and conduct experiments, potentially leading to breakthroughs in treating diseases like cancer, Alzheimer's, and genetic disorders. The implications are staggering: we could see the elimination of most major diseases and doubling the human lifespan within a decade.

But Amodei's vision extends far beyond health. He sees AI as a potential catalyst for rapid economic development. By optimizing economic policies and accelerating the spread of technological innovations, AI could help lift billions out of poverty at an unprecedented rate. It's an ambitious goal, that Amodei argues is within reach if we effectively utilize AI's potential. The essay also tackles thornier issues like governance and human rights. Amodei suggests that AI could strengthen democratic institutions, improve the delivery of government services, and even help counter authoritarianism on a global scale. It's a hopeful perspective that sees AI not just as a tool for efficiency but as a force for promoting human freedom and dignity.

Of course, Amodei is no stranger to the risks associated with AI development. As the leader of a company deeply involved in AI safety research, he acknowledges the potential dangers. However, he argues that focusing solely on risks misses the bigger picture. We need a positive vision of the future to strive for and a compelling reason to tackle those risks head-on. Amodei's essay challenges us to think bigger about AI's potential and to work towards realizing these ambitious goals. It's a future that will require technological breakthroughs, ethical foresight and unprecedented global cooperation. While we, too, believe in the power of AI, we will wait and see when Anthropic announces its next funding round before fully adopting these rose-colored glasses

Huang’s Law: NVIDIA’s Ambitious AI Roadmap

Nvidia's CEO, Jensen Huang, stands out as a visionary leader (and not just for his leather jackets) with a clear mission: to radically reshape computing for the age of AI. In a recent interview, Huang shared insights illuminating Nvidia's strategy and offering a strategic plan for the future of artificial intelligence and computing. Nvidia, once primarily known for its graphics processing units (GPUs), has evolved into the powerhouse leading the AI boom. At the heart of this shift is a staggering achievement: Nvidia has reduced computing costs by 100,000 times over the past decade. This leap in efficiency dwarfs Moore's Law, which would have predicted only a 100-fold improvement. For businesses in the AI space, this translates to unprecedented opportunities for scaling. Nvidia's ambitions extend far beyond mere cost reduction, as Huang emphasizes that the company is "reinventing the entire computing stack." This holistic approach involves breakthroughs across hardware and software, from developing tensor cores and high-bandwidth memory to creating comprehensive software libraries and tools.

One of the most compelling aspects of Huang's vision is his concept of the AI flywheel. He sees AI development not as a linear process but as a self-reinforcing cycle where each stage—from data curation to model training to inference—feeds into and improves the others. This perspective challenges organizations to think beyond isolated improvements and instead focus on optimizing the entire AI pipeline.

Huang also introduced the idea of "AI factories." In his vision, the next wave of technological infrastructure won't just be about data centers as we know them today. Instead, we're looking at the emergence of dedicated AI factories – massive computational hubs explicitly designed for creating, training, and deploying AI models. Unlike traditional data centers, which are optimized for storage and general-purpose computing, AI factories are designed to handle the enormous computational requirements of training large language models, processing vast amounts of video data, and running complex simulations. Huang predicts this new layer of AI-focused infrastructure will be a multi-trillion dollar opportunity.

Nvidia also has a unique approach to the competitive landscape, especially in an industry often characterized by cutthroat competition and battles over market share. Huang emphasizes that Nvidia doesn't see itself as fighting for slices of an existing pie. Instead, the company focuses on creating entirely new markets. This philosophy is evident in Nvidia's transformation from a graphics card company to an AI powerhouse. Rather than merely competing in the GPU market, Nvidia expanded the definition of what GPUs could do, opening up new applications in scientific computing, AI training, and now, inference at scale.

While Nvidia is known for its proprietary hardware and software, Huang also recognizes the critical role of open source in advancing AI technology. This is particularly evident in Nvidia's approach to synthetic data generation, as demonstrated by their release of large open-source models. The decision to open-source such powerful tools might seem counterintuitive for a for-profit company, but it aligns with Nvidia's market-making philosophy. By providing open-source models for synthetic data generation, Nvidia is creating new markets and use cases for AI. This approach has several benefits: i) it accelerates progress across the entire AI ecosystem, as researchers and developers around the world can build upon and improve these models; ii) it helps to democratize AI development, making advanced tools accessible to a wider range of organizations and individuals; iii) it positions Nvidia as a leader and key contributor to the AI community, reinforcing its role as more than just a hardware provider.

Finally, Huang's vision of the future is not one where AI replaces human workers but augments them. He envisions Nvidia growing to 50,000 employees working alongside 100 million AI assistants. This human-AI collaboration model offers a more optimistic and nuanced view of how artificial intelligence might reshape the workplace. As AI continues to evolve and reshape industries, Nvidia under Huang's leadership is not merely riding the wave but actively steering its direction. As we stand on the brink of this new era in computing, Jensen Huang provides both the vision and the tools to turn that vision into reality.

The Code Whisperer: Karpathy’s AI Roadmap

Andrej Karpathy, founding member of OpenAI, Tesla's former head of autopilot and current education disruptor, recently dished on the future of artificial intelligence. Although this podcast was taped before Tesla’s We, Robot event (aka the much anticipated, highly disappointing reveal of its robotaxis), Karpathy sat down to chat about everything from robo-cars to robot tutors. First up: self-driving cars. Remember when we thought we'd all be napping in autonomous vehicles by now? Karpathy suggests we pump the brakes on that fantasy, acknowledging that “Tesla has a software problem, and Waymo has a hardware problem.” (Although the streets of LA and SF would beg to differ.) Karpathy shifted to focus to the current buzz, which is all about humanoid robots. Apparently, Tesla's leap from cars to bipedal bots was strangely smooth, as early versions of Optimus thought it was a car. (We are unsure what curious scenarios that might have caused, but we can dream up a few.) Karpathy’s current passion project is disrupting education. His new venture, Eureka Labs, aims to democratize AI education. Imagine a world where everyone has a personal AI tutor, making learning as addictive as TikTok scrolling. Karpathy believes that the path to future success is paved with equations and algorithms — students should still study math, physics and computer science for future success. Don't worry, English majors - there's still hope for you. Maybe you can teach the robots Shakespeare? Whether you're an early adopter or a cautious observer, Karpathy's insights offer a compelling glimpse into a future that's rapidly becoming our present.

Back to Basics

Calculated Chaos: AI’s Math Problem

We won’t say we told you so, but ChatGPT is no math genius. A new study from Apple researchers has exposed some surprising weaknesses in even the most advanced AI language models when it comes to basic arithmetic and problem-solving. The report put several large language models through the paces on grade-school level math problems, and unfortunately, they won't be acing the SAT anytime soon. The researchers found that the models failed miserably if they added extra variables or extra steps to the math questions. Instead of just asking, "What's 2+2?", this test throws curveballs like “What’s 3+1"?” or "What's 2+2 if you're standing on one foot while patting your head?" Okay, not really, but it does mix up names, numbers, and complexity to stress-test the AI’s skills. The models were inconsistent; some models’ accuracy plummeted by up to 65% when extra steps were added. Adding extraneous facts caused the models to struggle, often trying to incorporate irrelevant data into their calculations. The study seems to underline that the models aren’t thinking or reasoning; they are engaging in pattern matching. This research serves as an important reminder of the current state of AI technology. As we continue integrating these systems into various aspects of our lives, understanding their strengths and weaknesses becomes increasingly vital. The path to AI that can genuinely reason like humans is still a long way away, although we’ll note that the study didn’t test the new o1 model, specially built to “reason.”

Tool Update

STORM: The New Scholar on the Block

What it is: STORM (short for Structured Task-Oriented Research Machine) was created by a team at Stanford, and acts like a research co-pilot. The chatbot generates Wikipedia-style articles with the precision of a seasoned researcher, using a multi-agent system that mimics a team of experts collaborating on a project. STORM doesn't just regurgitate information; it analyzes, synthesizes, and crafts content that's both comprehensive and meticulously cited.

How we use it: STORM is another research assistant in our toolbox - a study buddy. We like its analysis on complex topics, the detailed table of contents, and the ability to brainstorm and co-create with the chatbot. It reminds us of the customized reports we can create on Perplexity. We’d recommend STORM if you don’t have access to a paid Perplexity account or are working with academic research.

Perplexity: Gets Another Glow Up

What it is: Perplexity is our favorite tool for research, and it has quickly replaced Google search as our go-to for, well, google search. Perplexity has just rolled out a series of updates that make it even more indispensable. Perplexity's latest upgrades include: Charts, Spaces, and Reasoning. For the visually inclined, Perplexity now offers chart generation. Whether you're tracking stocks or analyzing trends, this tool transforms raw numbers into sleek visuals at a fraction of the cost of high-end financial terminals. Perplexity's updates to its Spaces feature extracts Collections, creating a dedicated home for your curated searches. We anticipate that Perplexity is adjusting the Spaces UX to make it more like a custom GPT or Claude Project, so users can organize and streamline their research process. Perplexity has also added o1 reasoning from OpenAI. This feature automatically kicks in for complex queries, breaking down multi-faceted questions and synthesizing coherent answers. While OpenAI has integrated its SearchGPT into its chatbot, our vote for the best results remains with Perplexity.

How we use it: Perplexity has become our Swiss Army knife for research and search. Some of our not-so-obvious use cases: we've found it invaluable for price comparisons when shopping online (goodbye, endless tab hopping). It's our secret weapon for pre-meeting intel, digging up far more intriguing tidbits than LinkedIn profiles ever could. And let's not forget those late-night learning binges – Perplexity turns our curiosity into full-blown intellectual adventures. It's like having a research assistant who never sleeps and always has one more fascinating fact up its sleeve.

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

Wiki Watchdogs: You may have noticed that the internet is getting bombarded with AI slop, and unfortunately, our favorite source of supposedly documented information, Wikipedia, is getting similar treatment. A crack team of Wikipedia editors, called the WikiProject AI Cleanup, have taken it upon themselves to protect the internet's favorite fact repository from AI-generated misinformation. Ilyas Lebleu, one of the group's founding members, noticed an uptick in articles that read like they'd been penned by a particularly eloquent robot. "We managed to replicate similar 'styles' using ChatGPT," Lebleu explains, inadvertently turning AI against itself in a clever bit of digital judo. Armed with this knowledge, the team has combed through Wikipedia's vast archives, developing a sixth sense for AI-authored entries. They're compiling a phrase book of AI tells – linguistic quirks that betray a non-human origin. As AI-generated content floods everything from Google search results to academic journals, Wikipedia is a potential oasis of human-verified information. While AI might be able to string together impressively coherent paragraphs, it takes a human touch to separate fact from fiction, context from confusion. So the next time you fall down a Wikipedia rabbit hole at midnight, spare a thought for the WikiProject AI Cleanup team. They're out there, tirelessly ensuring that facts, not fabrications fuel your late-night knowledge binge.

The State of AI: Nathan Benaich dropped his annual "State of AI Report 2024" and the TLDR: artificial intelligence, once the darling of sci-fi and tech enthusiasts, has finally stepped out of the realm of speculation and into the harsh light of reality. Once dominated by OpenAI's GPT-4, the AI race has seen competitors closing the gap with impressive speed, turning what was once a solo performance into a symphony of innovation. And yet, the real story of 2024 isn't just about language models getting smarter. AI has broken free from its linguistic roots, branching into disciplines as diverse as genetics, protein folding, and climate science. This expansion marks a significant shift, transforming AI from a specialized tool into a universal problem-solving platform. The economic impact of AI in 2024 is also quite impressive. With a collective value of $9 trillion, AI companies have become a dominant force in the global economy. Public companies in the AI sector are experiencing a bull run reminiscent of the early days of the internet, while even private startups are securing funding rounds that would have been unthinkable just a few years ago. This influx of capital is fueling rapid advancement, and raising questions about sustainability and market dynamics. One prediction from the report caught our eye: a video game centered around AI-generated elements is set to achieve break-out status in the coming year. This isn't just about better graphics or smarter NPCs – we're talking about a fundamental shift in gaming. With the release of Google’s GameNGen AI model for gaming, we think this is highly likely (and highly entertaining).

To the Max: Adobe is in the middle of its annual showcase, Adobe Max, and our socks have already been blown off. Among many innovative tools, two standouts are stealing the show: the Firefly Video Model and Project Concept. These offerings aren't just incremental updates—they're quantum leaps in creative technology. The Firefly Video Model, now in limited public beta, introduces "Generative Extend," allowing editors to seamlessly add frames to existing clips—ideal for perfecting transitions or extending shots. The model can also generate entire video clips from text prompts and even animate still images, bringing Adobe into fierce competition with Runway, Kling, and Luma, among others. Equally intriguing is Project Concept, Adobe's foray into collaborative ideation — it is a virtual, AI-powered moldboard. This platform enables multiple users to remix images in real-time on a shared canvas. It's a digital reimagining of the traditional brainstorming session, designed to streamline the conceptual phase of projects. These tools will speed up workflows and expand the boundaries of what's possible, and we can’t wait to dive in.

— Lauren Eve Cantor

thanks for reading!

if someone sent this to you or you haven’t done so yet, please sign up below 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.