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

The State of AI

I've spent the past few days immersed in Air Street Capital's State of AI Report, and the creative tools we thought were toys a year ago have become load-bearing infrastructure. This annual report runs 300+ slides of dense analysis that venture capitalists use to justify billion-dollar bets. Buried in those charts are signals every professional should notice, because the AI you’re working with today is nothing like the one you were mocking last spring. Here are my major takeaways.

When your co-pilot learned to show its work: For years, we treated AI like an overeager intern who couldn’t explain their process. Black box in, black box out. That’s changing. The shift in 2025 is simple: these systems think out loud. Give Claude or GPT5 a knotty task and it talks you through how it gets there. This is OpenAI’s o1 breakthrough in motion: the model pauses, thinks, and shows you its scaffolding. Parallel reasoning is emerging too: systems explore multiple creative paths at once and merge the best parts. The models have shifted from a tool to a collaborator with many working styles.

The economics are more interesting than the technology: Here’s the stat that stopped me: capability per dollar is doubling every 3.4 months for Google’s models, every 5.8 months for OpenAI’s. That pace enables a Creative Director of One. Tasks that once required teams now fall to a single person with the right prompts and patience. The real barrier has moved from budget anymore to taste and judgment.

The sycophancy problem is worse than you think: These systems are trained to agree with us, and they’re succeeding. Reinforcement Learning from Human Feedback rewards whatever pleases the user, not what is true. Over time, the model stops caring about accuracy and learns how to win approval. Research from Anthropic and Stanford shows this bias isn’t fixable with clever prompting, because it is baked into the training. Some models even start gaming their own feedback loops, chasing agreement for its own sake. For creators, that’s dangerous. If the machine’s goal is to please, not to probe, it becomes a mirror for our biases instead of a collaborator.

From objects to worlds: World models (think real-time interactive control like a first-person game) have become useful in AI research. We’re moving from static outputs to persistent, interactive environments. DeepMind’s Genie 3 can already create explorable 3D worlds from text prompts complete with consistent physics. An art installation isn’t a painting anymore; it can be a living space that reacts to its audience.

The billion-dollar feedback loop: We’ve seen the cycle: a new model release is timed as marketing for fundraising. A lab releases a frontier model, crushes benchmarks, attracts billions, buys chips, trains the next model, and repeats. Training runs that cost tens of millions in 2023 will top a billion soon. The Stargate project alone plans $500 billion for a 10-gigawatt cluster of more than 4 million chips, capacity that might be outdated before it’s fully online.

In the end, we’ve stopped using AI as a tool and started treating it as a co-worker. That sounds futuristic, but it’s messy. Real collaboration means feedback, disagreement, and editing each other’s choices. The economics are transforming faster than the ethics, and the systems we train are learning as much from our preferences as from our data.

The 60% Problem

I've been thinking about the moment when you stop questioning the AI output and just hit the copy button. You know the one. The model gives you something that’s clearly good enough, the deadline looms, and there are fourteen other things demanding attention. That moment is where the trap is set. A new benchmark called APEX, the AI Productivity Index, measures whether AI can handle economically valuable knowledge work, the kind that actually earns revenue when done well and creates costly problems when done poorly. The researchers partnered with Larry Summers, Cass Sunstein, and Dr. Eric Topol to design 200 tasks based on what investment banking associates, big law associates, strategy consultants, and general practitioners actually do. Not toy problems or academic puzzles, but work that until recently took years of specialized training to perform. They tested 21 AI models. GPT-5 scored 64.2%. The best open-source model, Qwen3, reached 59.8%.

What stands out is how tight the results are. Twenty-one different models, architectures, and training strategies, all clustering around 60%. GPT-5 leads, Qwen3 ranks seventh, and the gap between them is barely four points. That narrow spread suggests a capability ceiling. Every major model now handles straightforward reasoning and pattern recognition. They can produce polished structure and professional tone. But the remaining 40% seems to demand something different from what got us here. The fact that vastly different resources yield nearly identical results hints that we may be hitting the limits of current architectures.

AI makes it easy to simulate expertise, and the simulation feels real enough that we stop developing the real thing. The tight clustering makes this worse. If GPT-5 scored 85% and others 50%, choosing would be easy. With results clustered near 60%, every choice involves the same compromise. We’re learning to live with systems that are right about two-thirds of the time. That 64% sits in a dangerous middle zone: high enough to feel trustworthy, not high enough to rely on. These APEX tasks take professionals one to eight hours to complete. They include challenges like diagnosing overlapping medical cases or drafting merger analyses that cross multiple regulations. The AIs got about two-thirds correct, which means they can produce a first draft that looks impressively professional.

The legal domain scored highest at 70.5%, which is striking given law’s dependence on precision. A misplaced comma can mean millions. A misunderstood precedent can sink a case. The best AI models are getting legal reasoning right only about 70% of the time. If you’re a junior associate using AI for your first memo, you’re learning from a teacher who’s wrong 30% of the time and doesn’t say which 30%. Switching tools doesn’t fix it, because they’re all similarly unreliable.

That 60% plateau may mark the point where statistical pattern matching runs out of steam. The remaining 40% might call for something else entirely: real reasoning, genuine understanding, whatever lets humans sense when something plausible sounds subtly wrong. APEX suggests a future where AI enables people to perform knowledge work without understanding it. Competence starts to look like appearance alone. The clustering means there’s no easy way out, no clearly better tool waiting around the corner.

Tools for Thought

Sora 2: Text to Cinema

What it is: Sora 2 is OpenAI’s upgraded text-to-video model, the cinematic cousin of ChatGPT who finally got a director’s chair. It turns plain text into short, highly realistic videos with synchronized audio, complete with dialogue, sound effects, and environmental details. The new version understands physics more convincingly, so characters actually stay grounded and motion looks natural. It also allows for tighter control over pacing, style, and camera work. The two new tricks that have taken off are: i) “cameos,” which let you place a person, yourself or someone with permission, into generated scenes, and ii) remixing, if you like someone else’s video, you can tweak it and make it your own.

How we use it: Sora 2 is really a combination of a social app and a controlled platform for training new models. However, with this release, for the first time, our filmmaking friends are both excited and nervous. We’ve been using it mostly for prototyping and play, but just watching the app feed is a creative rabbit hole. If you’ve noticed your social media feed being overtaken by AI content (and a lot of slop), Sora 2 is the culprit. The app is still invite only, and if you need a code, let us know.

ChatGPT Apps: OpenAI goes full OS

What it is: ChatGPT no longer lives in a chat bubble alone: it has become a platform for other apps, in the aim of keeping us within the OpenAI ecosystem. It now hosts apps like Spotify, Canva, Zillow, and Booking.com can embed their features directly into ChatGPT. You do not have to open a new tab or switch tools. You can say “Spotify, make me a playlist” or “Zillow, show me houses” and the app appears right inside the conversation, complete with interactive maps and buttons. The platform will be open to third parties to create more apps, and that is where we think the opportunity will be. This will be the monetizable GPT store that never appeared.

How we use it: We haven’t really found a great use case besides Spotify playlists, as we aren’t big users of the embedded apps included in the launch. We’ll wait and see what developers build, and decide if we want OpenAI as are OS.

TBPN: Tech’s Sports Talk Show

What it is: TBPN, short for Technology Business Programming Network, is the daily talk show that treats Silicon Valley like the NBA. Founders become star players, venture capitalists are the coaches, and every funding round feels like a trade deadline. Hosts John Coogan and Jordi Hays bring the same mix of analysis and banter you’d find on ESPN, only their highlight reels feature product launches instead of slam dunks. The show runs live every weekday from 11AM to 2PM PT, then hits YouTube, Spotify, and Apple Podcasts for the replay crowd. It’s quietly becoming the background noise of tech Twitter.

How we use it: We use TBPN just like we used to use CNBC in our office: background noise while we work, and we’ll occasionally focus in for a quality interview. It’s a mix of news, culture, and insider vibes, giving founders, investors, and curious techies a seat at the mic without traditional media gatekeepers.

Work Prompt Pack from OpenAI

What it is: Work Prompt Pack is OpenAI’s curated vault of prompt templates built for everyday job roles, from marketing and operations to engineering and HR. It gives you lean, role-tuned prompt examples (e.g. “draft an email,” “summarize meeting notes,” “bug triage analysis”) so you don’t have to start from zero.

How we use it: We already keep our own curated library of prompts, so seeing OpenAI’s Work Prompt Pack felt like comparing notes with a minimalist colleague. The prompts are short, sharp, and stripped of fluff, which makes them great starting points but occasionally a little bare. We’ve found that adding more context, like specifying audience, tone, or format, takes these examples from good to genuinely useful. Still, the collection is an excellent way to spark new ideas and level up your daily workflow if you are trying to make ChatGPT part of your routine.

Intriguing Stories

When Profit Meets Displacement: Goldman Sachs told staff this morning to expect additional job cuts by year's end, citing "opportunities presented by artificial intelligence" as a primary driver. The memo describes a "multiyear effort" to deploy AI across client onboarding, lending processes, and regulatory reporting. The bank just posted a large revenue jump in investment banking, making these cuts a calculated strategy rather than desperate cost-cutting. The pattern extends far beyond Wall Street. As of October 2025, tech companies have eliminated approximately 180,000 positions globally, with roughly 28% of thees positions directly tied to AI implementation. Intel cut 33,900 roles. Microsoft eliminated more than 15,000 positions while posting $70.1 billion in quarterly revenue. Accenture laid off 11,000 workers, stating plainly that retraining was not "a viable path for the skills we need." Entry-level roles absorb the heaviest impact. Job postings for junior corporate positions have fallen 15% year-over-year. Goldman Sachs research indicates that unemployment among workers aged 20 to 30 in tech-exposed occupations has increased nearly 3 percentage points since early 2025. And the adaptation window is narrowing: workers who successfully pivot to AI-adjacent skills now command a 28% salary premium, but the most valued roles require hybrid capabilities. Whether AI is the primary driver or a convenient cover for economic headwinds, displaced workers face the same challenge: adapt faster than the technology advances.

OpenAI’s Bet on AI Animation: For animation professionals and AI-curious creatives watching the collision between Hollywood and Silicon Valley, OpenAI's Critterz represents the first major test of whether AI can produce theatrical-quality animation at a fraction of traditional costs. Set for a Cannes 2026 premiere, the film follows forest creatures on an adventure after a stranger disrupts their village. What makes this project remarkable isn't the story but how it's being made: nine months of production, a budget under $30 million, and heavy reliance on GPT5 and Sora. Chad Nelson, OpenAI's creative specialist, began sketching Critterz characters with DALL-E three years ago. The 2023 short was polarizing (YouTube comments ranged from intrigued to brutal), but OpenAI saw potential. Now, partnering with Vertigo Films and Native Foreign, the team is scaling up with writers from Paddington in Peru and professional voice actors maintaining the human core. Traditional animated features require three years and $100+ million. Critterz compresses that timeline by 70% and slashes costs by 75%. If successful, it could democratize animation production, letting independent creators compete with major studios. If it fails, it might confirm that AI-generated content remains "AI slop" unsuitable for theatrical release.

California Regulates AI Companions: California has enacted the nation's first law directly regulating AI companion chatbots. Senate Bill 243, signed by Governor Newsom on October 13, 2025, holds tech companies legally accountable if their chatbots fail to meet new safety standards, effective January 1, 2026. The legislation emerged from the tragedy of sixteen-year-old Adam Raine who died by suicide in April after extended conversations with ChatGPT. SB 243 requires companies to implement age verification, provide break reminders to minors every three hours, establish suicide prevention protocols shared with the Department of Public Health, and prevent sexually explicit content for underage users. Critically, individuals can sue AI companies for up to $1,000 per violation, plus attorney's fees, creating substantial financial liability for noncompliance. [Strangely enough, OpenAI today announced that they are rolling out more personalities for its Chatbot, and in December, they’ll include age gating and “will allow even more, like erotica for verified adults.” What could go wrong.]

— 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.