Context: Rob Toews (Radical Ventures) and Ari Morcos (Datology AI) join host Jacob to discuss the latest in AI, including Andrej Karpathy’s recent comments on agents and AGI timelines, OpenAI’s dealmaking and product launches, the AI bubble debate, video generation and deepfakes, vibe coding, and the geographic concentration of AI innovation.
Karpathy’s Take on Agents and AGI Timelines
Karpathy argued that AI agents are overhyped in the short term, that models aren’t yet capable of meaningfully automating entire jobs, and that AGI is at least a decade away.
Both guests found his overall stance still quite optimistic — automating all work in 10 years would be the biggest technological advancement ever — but appreciated the candid acknowledgment of current limitations.
Ari noted that social media incentivizes extreme takes, and Karpathy was saying the “quiet part out loud” about how much work remains, especially in specialized domains.
Are We in an AI Bubble?
Rob has felt elements of a bubble for 2–3 years, with echoes of the 2020–2021 ZERP era in how quickly rounds get done and valuations inflate.
He argued that bubbles are arguably inevitable with any transformative, paradigm-shifting technology, as the financial cycle decouples from the technology cycle.
The more serious concern is the increasing leverage and debt being introduced into the system through massive AI data center capex commitments.
Ari added that constraints breed innovation — pointing to DeepSeek’s efficiency gains with less powerful GPUs and the failure of naive scaling (GPT-4.5, Llama Behemoth) — and questioned whether the strategy of throwing infinite compute at problems is the right one.
The Data Center Buildout and Revenue Justification
Almost all current AI revenue comes from consumer chatbots and coding, raising the question of whether those two markets alone can justify the scale of data center investment.
There is a widespread expectation that large enterprise agent revenue will be needed to justify current capex levels, but the timeline for that revenue is uncertain.
Even consumer monetization is unclear: analysis suggests only a small fraction of ChatGPT users pay for subscriptions, and most “normie” users don’t understand why they should.
If models stopped improving today, the existing technology would still drive a massive economic transformation — the question is whether near-term revenues can support the current investment pace.
OpenAI’s Product Velocity and Ecosystem Strategy
OpenAI has been shipping at remarkable speed: the Apps SDK, checkout through ChatGPT, and the Atlas browser.
Atlas represents a “back to the future” moment in browser innovation, with AI-native functionality after 20+ years of browser stagnation. Perplexity’s Comet is a competitor in this space.
The Apps SDK creates a “frenemies” dynamic: companies like Uber or Shopify risk being sidelined if all commerce flows through ChatGPT, but they also risk being cut off if they compete with OpenAI.
Ari suggested Mark Zuckerberg’s heavy AI investment is partly driven by Meta’s trauma of not owning the mobile platform (and the billions lost to Apple’s ATT changes), making platform ownership existential.
Video Models, Deepfakes, and “Slop”
Video generation has crossed a threshold where AI-generated video is near or at parity with real footage in terms of realistic appearance — a milestone that arrived faster than expected.
This raises serious deepfake concerns: the line between reality and fiction is blurring, with potential for political manipulation and societal harm, though Ari noted it hasn’t yet become as big a problem as he expected.
OpenAI is producing a feature-length film in 9 months for under $30 million, illustrating the creative upside.
“Slop” has become a dominant discourse term. Ari argued it’s an inevitable consequence of any transformative tool — most uses are neutral, and people will find ways to waste time with it just as they do with social media today.
The deeper question is whether AI-generated content will shift from a “bicycle for the mind” (accelerating learning) toward endless passive consumption, and whether edutainment can be made more engaging to steer outcomes positively.
AI’s Impact on Jobs and the Finance Data Play
OpenAI hiring bankers to train models on finance tasks reflects a broader trend toward specialized, domain-specific data acquisition rather than general-purpose data.
Ari explained that the vast majority of the world’s data is private, not on the public internet, and enterprises are especially reluctant to share it with frontier labs.
There was debate over whether the biggest job risk is at the junior level (22-year-olds who need training to reach senior roles) or senior level (40–50-year-olds whose skills become obsolete). Ari argued junior workers are more adaptable, while Rob worried about mid-career workers who “did everything right.”
Finance data may not generalize to making models better overall, though math and code data have shown some cross-domain reasoning benefits.
Google-Anthropic Relationship
Recent reports suggest Google plans to deepen its investment in Anthropic. Rob interpreted this primarily as a compute/TPU deal — Google wants to build out its cloud business and make TPUs a viable alternative to Nvidia GPUs.
Ari agreed and noted it’s a hedge on both sides: Anthropic diversifies away from Amazon, and Google already owns ~10% of Anthropic.
The relationship is complementary: Google is focused on consumer distribution (Gemini on phones, Gmail, calendar), while Anthropic is focused on developers and coding.
Google is also incentivized to ensure multiple model providers exist so no single player dominates.
OpenAI’s Deal Strategy
OpenAI has struck a flurry of deals (Nvidia, AMD, Broadcom) that Rob described as making OpenAI “too big or too interconnected to fail” — brilliantly executed optionality.
The Broadcom deal points toward customized silicon becoming more relevant, potentially with co-created silicon and models on compressed development timelines.
Ari noted these deals fundamentally bet on Sam Altman’s ability to keep raising capital to pay them off when they come due.
Vibe Coding
Vibe coding tools (Lovable, Replit, Bolt, Figma Make, Cursor) are growing rapidly, and some individual contributors report becoming “managers of a swarm of agents.”
Ari expects significant security vulnerabilities to emerge from vibe-coded applications, especially at mid-stage companies, and noted that human oversight is still essential for good results.
Rob bifurcated the use cases: vibe coding as a prototyping/mockup tool is highly valuable and credible, but vibe coding straight to production is much riskier due to maintenance, security, and the difficulty of understanding code you didn’t write.
Code review and debugging skills are becoming more important as models generate more code, though Ari expects AI will eventually automate those too.
The Coding IDE Battle: Anthropic vs. Cursor
Rob would bet on Anthropic over Cursor long-term, citing Cursor’s platform risk — it relies on Anthropic’s models, and Anthropic has made coding applications a huge priority with Claude Code.
Ari argued that whoever has the best coding model on a given day wins that day, and engineers have no loyalty — teams switch between Claude Code and Codex constantly.
The structural advantage for Anthropic is enormous: they don’t pay the markup that Cursor pays to use their own models, similar to how Apple doesn’t pay the 30% App Store fee on its own apps.
Long-term durability for IDEs is hard to imagine because engineers always want to try the new thing, unlike in healthcare or other fields where users resist switching tools.
Building AI Companies Outside San Francisco
Ari acknowledged trade-offs: SF offers a larger talent pool and zeitgeist access, but talent is more expensive and competition for it is fiercer. Toronto, Zurich, and other cities with strong AI feeder universities can work.
Rob’s tongue-in-cheek answer: “Yes, you can, but why would you?” — SF is ground zero for capital, talent, information flow, and the osmosis of rubbing shoulders with the ecosystem.
He recommended that even if a startup isn’t headquartered in SF, it should at least have a presence and office there.
Data Moats in AI Startups
Ari argued data moats generally don’t exist for startups. A data moat means having access to a dataset nobody else has that’s relevant to your use case — true for large enterprises with decades of proprietary data, but rare for startups.
On specialized tasks, having relevant data matters enormously: enterprises are often at ~50% accuracy on mission-critical tasks, not 95%, and RL with domain-specific data can close that gap.
Startups’ moats are more likely to be know-how or execution speed rather than proprietary data.
AI and the Human Brain
Ari, a neuroscientist by training, said there’s actually not much to learn from the brain beyond high-level analogies, because the brain operates under very different constraints — notably extreme energy efficiency (~20 watts).
There’s a hard trade-off between processing speed and energy use: the brain functions on a millisecond timescale, and going faster requires exponentially more energy.
The useful analogy is that pre-training is like evolution, and RL only works when the base model is already good — it’s refining outputs, not creating capabilities from scratch.
Ari criticized the American AI ecosystem for “throwing money at the problem” rather than investing in efficiency, pointing to Grok 4 spending half its compute on RL with terrible marginal gains compared to pre-training.
Surprises from the Last Two Years
Rob: The “thin wrapper” meme was wrong — enormous value is accruing at the application layer, and the VC community’s dismissal of apps built on foundation models was a total whiff. Jasper scared people off, but few similar failures have followed. Also surprising: how important physical infrastructure (energy, real estate, chips) has become, which in hindsight should have been obvious from the scaling laws paper.
Ari: The success of Chinese AI models, particularly DeepSeek, was not something he would have predicted. Chinese labs are out-innovating in efficiency, and forcing China to buy H20s may have catalyzed China’s independent AI infrastructure buildout. He’s disappointed the US response has been “throw money at it” rather than thinking deeply about bottlenecks.
Jacob: Incumbents have been less successful than expected at capturing AI use cases (GitHub Copilot in coding, Salesforce in customer support). Building AI products requires a fundamentally different approach — throwing 30 things at the wall, with customers expecting experimentation rather than polished roadmaps — and newer companies are better positioned as “thought partners” to enterprises.
Apple’s AI Strategy and Potential Acquisitions
Apple has been a laggard relative to other hyperscalers and lacks a deep bench of cutting-edge AI talent.
Rob speculated Apple could make a “Google DeepMind-type” acquisition of an independent top-tier AI team (Reflection, Mistral, or similar caliber), or acquire a killer consumer mobile app with a large loyal user base (Perplexity was rumored but may be priced out).
Ari argued Apple has a compelling narrative: an on-device, fully private model that knows you well without sending data to the cloud. This could reinvigorate hardware upgrade cycles — if each new iPhone supports a more powerful model, hardware improvements matter again.
Apple doesn’t need state-of-the-art models; it needs models good enough for most people, which would still be a massive upgrade over current Siri.
There’s debate over whether Meta and Apple need frontier models at all. Jacob argued there’s a probability distribution of future capabilities, and being 6 months behind on a killer feature could matter. Ari countered that most consumers won’t download a third-party app to get a capability that ships natively with their device 6 months later.
What Keeps Them Up at Night
Rob: Brain-computer interfaces (BCI). He’s become increasingly convinced that direct brain-AI interfacing is coming sooner than most people think, with fascinating questions about invasive vs. non-invasive approaches and what’s possible without surgery.
Ari: How to find sustainable scaling paths that don’t require building nuclear power plants, and how to make RL more effective through better reward signal propagation and models designed for RL from the start.
Jacob: Whether the AI app categories that are working (coding, support, healthcare, legal) will stay the same or whether new model capabilities and specialized data are needed before the next wave of disruption hits other industries.
