Is AI a Bubble? Experts Debate the Future of AI w/ Dave, Salim, and AWG | EP #190
Summary
David Blundin rejects the idea that AI itself is a bubble, while conceding that plenty of AI investments will fail. His distinction is the investable one: charlatans and weak companies can collapse without invalidating “the biggest ship in human history.” OpenAI’s reported $1 billion in monthly revenue, roughly 50% week-over-week token growth, near-doubling of agentic usage, and 8x jump in reasoning usage suggest demand remains constrained more by compute than appetite.
The next edge-computing crossover could put frontier-grade intelligence inside robots, vehicles, phones, and industrial equipment within 6–12 months. The cited setup is a roughly $2,500 NVIDIA RTX 5090 running models comparable to today’s frontier, while Gemma 3’s 270-million-parameter model reportedly handles 25 chats on 1% of a Pixel 9 Pro’s battery. Alexander Wissner-Gross argues the durable thesis is latency, not merely privacy: humanoids need foundation models operating locally at “ultra-low latency.”
Model economics may improve far faster than conventional scaling curves imply. A 32-billion-parameter distillation result reportedly reached comparable capability with 1% of the training corpus—roughly a 100x difference—by using a teacher model, structured curriculum, and step-by-step explanations. Wissner-Gross sees “overhangs everywhere,” while Blundin argues gains across data selection, optimizers, software, and chips multiply, overwhelming diminishing returns on any single compute-scaling curve.
Scientific discovery is beginning to move from isolated demonstrations toward an automated production system. GPT-5 Pro reportedly improved a proof in convex optimization, while GPT-4b designed cellular reprogramming factors said to be 50x more effective; Wissner-Gross expects today’s “trickle” to become bulk proofs, discoveries, and inventions. His key mechanism is recursive optimization: if AI designs better optimizers—and optimizers better at designing optimizers—it reaches “the innermost loop of our civilization.”
Enterprise adoption, rather than model capability, is the near-term weak link. The cited MIT study says companies spent $30–40 billion on generative AI while 95% of pilots produced no financial return; buying existing products succeeded about two-thirds of the time, versus roughly one-third for internal builds. Salim Ismail’s prescription is an AI-native edge organization reporting directly to the CEO: “Do not try and transform the mother ship.”
Compute, energy, distribution, and talent—not benchmark leadership alone—will decide which frontier labs capture the economics. OpenAI is pursuing a Texas Stargate buildout of up to 5 GW and a 290 MW Norway center with 100,000 GPUs, while Google combines infrastructure, an $85 billion AI-capex figure, existing distribution, and a reported 14% Anthropic stake. Blundin’s formulation is stark: the industry now has “infinite appetite for compute,” creating a new market for allocating capacity by “productivity per token.” Diamandis also says GPT-5 has put 700 million people into free-model access, potentially creating a productivity-and-capital feedback loop.
BCIs and humanoid robotics are the panel’s bridge between accelerating machine intelligence and the human economy. Merge Labs is described as pursuing gene therapy plus ultrasound to read and write neurons, while Figure, 1X, Unitree, and others are advancing autonomous manipulation and locomotion. Wissner-Gross warns that high-bandwidth BCIs may need to arrive within a few years, before a “pure AI economy” decouples from humans; the panel’s closing horizon is a Star Trek-like convergence in the 2030s, but richer in AI and biotechnology.
Deep dive
1. Human benchmarks are saturating before model progress is
Peter Diamandis opens with GPT-5 Pro scoring around 148 on the Mensa Norway IQ test, versus an average human score of 100. Wissner-Gross’s response is that tests built around broad human populations are “running out of sigma”; as scores approach 200, the number becomes less meaningful than the model’s performance on specialist knowledge and unsolved problems.
His proposed replacement is an “abundant set of benchmarks” tied to major scientific and social challenges. The earlier jump from GPT-o3—described as scoring around 120 or 136—to GPT-5 Pro matters, but the stronger signal is that two-year-old barriers are becoming routine enough that “nobody noticed” when systems crossed them.
Blundin extends the point through the brain’s physical scale: if the neocortex is roughly “the size of a dinner napkin,” what happens when machine augmentation makes its functional equivalent the size of a tablecloth or football field? His answer is deliberately spare: “We listen carefully.”
Wissner-Gross’s broadest claim is that AI may already have crossed a practical singularity threshold—not because one benchmark declared AGI, but because “the pace of change is faster than we can process it.” The curve is outrunning the institutions and people attempting to measure it.
2. Local intelligence makes latency the edge-computing thesis
The panel discusses a forecast that one top-end gaming GPU—an NVIDIA RTX 5090 costing about $2,500—could run models matching current frontier performance locally within 6–12 months. Privacy and personal superintelligence are immediate benefits, but Wissner-Gross calls those the “superficial story.”
His real crossover is between compact frontier models and embodied machines: general-purpose humanoids running foundation models locally, interpreting video and acting in the physical world without cloud round trips. Cleaning a house or driving a car may not require the absolute best model; it requires a sufficiently capable one responding at “ultra-low latency.”
The 1X Robotics example makes the constraint tangible. CEO Bernt Øivind Børnich reportedly placed compute in the robot’s head because the delay from eyes to brain and brain to actuators was unacceptable. Sensitive law, accounting, and government work adds another reason for locality: even with MCP-style integrations, firms may not want their queries uploaded.
Google’s Gemma 3, a 270-million-parameter model, reportedly managed 25 chats using 1% of a Pixel 9 Pro’s battery. Wissner-Gross imagines the endpoint as a “small diamond nanokernel”: a multimodal reasoning core with perhaps only millions of parameters, calling external knowledge as needed instead of memorizing the world.
3. Distillation exposes a 100x training-efficiency overhang
A 32-billion-parameter model reportedly moved the Pareto frontier on AIME 2024 and 2025 through “data-efficient distillation.” Wissner-Gross explains the mechanism as a large teacher model educating a cheaper student through a structured curriculum, curated data, step-by-step explanations, and other training innovations.
Blundin emphasizes the logarithmic chart: the highlighted model used one-hundredth the training corpus of the comparison point to reach equivalent knowledge. The commercial implication is a roughly 100x training difference, potentially letting startups build specialist foundation models without matching OpenAI’s or Google’s billion-dollar-plus budgets.
His best use cases are narrow but valuable: reading X-rays or designing rocket components. A startup can pursue frontier-level intelligence within a specific domain, then differentiate through proprietary data and workflow integration rather than attempting to reproduce a general model’s entire knowledge base.
Wissner-Gross calls these latent opportunities “overhangs”—capabilities waiting to burst out once researchers locate the right technique. Distillation, dataset organization, unused device processors, and old scientific results all qualify as potential sources of major gains without simply scaling raw training compute.
4. Better optimizers create the recursive engine beneath AI progress
GPT-5 Pro reportedly received an unsolved problem from a convex-optimization paper and produced a new proof improving the authors’ attempt. Wissner-Gross characterizes it as a modest theorem-level improvement, yet also as the leading edge of AI beginning to “bulk solve math, science, and engineering.”
He expects the isolated proof to become a “tidal wave” over roughly the next year, possibly beginning by year-end: bulk mathematical proofs, scientific discoveries, and engineering inventions arriving simultaneously. The cultural problem is that research institutions have no precedent for absorbing discovery at that volume.
His recursive argument centers on computer science’s “innermost loop,” the critical loop whose optimization accelerates everything around it. Today that loop may itself be optimization: an AI that discovers an orders-of-magnitude-better optimizer—and then optimizers better at developing optimizers—would compound progress across the civilization’s technical stack.
Blundin says an earlier estimate put software-only improvement between 100x and 10,000x across roughly eight multiplicative dimensions. If data selection alone can contribute 100x, he argues that estimate may be conservative; diminishing returns in raw compute matter less when optimization, curricula, architectures, chips, and tools improve together.
5. Forecasting becomes indistinguishable from invention
Asked whether GPT-5 can predict markets, Wissner-Gross separates finance from the rest of the world. Any reliable S&P 500 “crystal ball” would be adopted by quant funds and immediately priced in, limiting its visible edge; forecasting social, scientific, engineering, or physical systems offers a less reflexive test.
He then reverses the direction: if AI can predict the future, perhaps it can “retrodict” the past. With sparse evidence from ancient Greece or Rome, a model might reconstruct missing events at high fidelity—an aspiration sometimes described as “quantum archaeology”—and counter histories written primarily by victors.
Salim Ismail connects that possibility to historical archives and Will and Ariel Durant’s attempt to document civilization objectively. His proposed product is effectively a referenceable Wikipedia of what likely happened, balancing inherited narratives while retaining uncertainty where evidence cannot resolve competing reconstructions.
The forward benchmark Wissner-Gross proposes is the next 20 Nobel Prize-winning discoveries. Predicting them accurately would require making the discoveries first, collapsing the distinction between forecast and innovation: “The best way to predict the future is create it yourself.” Fusion milestones could provide a nearer-term, publicly trackable version of that benchmark.
6. OpenAI’s national expansion is a compute-and-distribution land grab
Sam Altman says India is OpenAI’s second-largest market and may become its largest, with users requesting better language support and cheaper access. Diamandis frames the move—alongside reported talks to provide ChatGPT Plus across the UK—as a land grab for students, entrepreneurs, factory workers, clinicians, and government employees.
Blundin sees more than customer acquisition in OpenAI’s decision to prioritize New Delhi while declining suggestions for Boston expansion. India combines the world’s largest population, a large 20-to-35-year-old cohort, and an “intellectual overhang”; he also points to Mercor, described as operating heavily there and valued around $10 billion.
Ismail argues that India’s underused talent resembles the Ramanujan story: exceptional intellectual capacity can remain invisible without infrastructure and institutional access. The binding constraints are energy, bandwidth, and compute; resolving them could let people “scaffold themselves” out of poverty through education, health services, and higher productivity.
Diamandis names this potential model “universal basic compute,” a supply-side counterpart to universal basic income. In his abundance framework, intelligence and energy are the feedstocks; once both are plentiful, many material and service scarcities become downstream engineering problems rather than fixed limits.
7. Compute demand, not weak usage, defines the bubble debate
OpenAI is described as building Texas Stargate capacity of up to 5 GW and a 290 MW Norway center with 100,000 GPUs powered by hydro. Diamandis’s formulation—power is the “pheromone that attracts the data centers”—turns national energy policy into an AI-industrial strategy.
OpenAI CFO Sarah Friar reports approximately $1 billion in monthly revenue, about 50% week-over-week growth in tokens after GPT-5’s release, agentic-token usage nearly doubling, and reasoning usage rising 8x. Despite launch disappointment among those expecting AGI or hard takeoff, actual use reportedly accelerated.
Blundin’s answer to “Is AI a bubble?” is categorical: “Absolutely not a bubble. It’s the biggest ship in human history.” Bad investments, opportunistic founders, and failed companies will exist, but he argues that conflating those failures with the underlying technology ignores a tailwind “like nothing we’ve ever seen.”
Ismail interprets Altman’s bubble warning as valuation management amid a reported secondary sale at a $500 billion valuation. Wissner-Gross supplies the condition for continued capex: as long as revenue keeps growing spectacularly, data-center spending can continue; the emerging allocation problem is a “microeconomics of productivity per token.”
Diamandis says GPT-5 has uplifted 700 million people into free-model access. Wissner-Gross describes a possible feedback loop in which their increased reasoning, productivity, and economic output generate more capital for data centers and broader access.
8. Infinite context turns software creation into an industrial process
Claude Sonnet 4’s million-token context window is described as roughly 750,000 words or 3,000–4,000 pages, versus about 250,000 tokens for GPT-5 and 500,000–600,000 for Grok. Blundin says users underestimate context needs because AI creates material faster than they expect.
His personal example is intentionally extreme: more code written in two weeks than in the previous 40 years, with nearly a terabyte of code and text generated in one week. By day three, the model forgetting yesterday’s work becomes a material bottleneck: “At the rate that everything’s accumulating, it’s just never enough.”
Wissner-Gross sees no known theoretical ceiling and anticipates effectively infinite context within a few years. If a company can place all its code, documentation, emails, and institutional knowledge into context, retrieval-augmented generation and fine-tuning may become less important; the marginal intelligence comes from direct access to the entire corpus.
Blundin’s portfolio example is Blitzy, which he says can produce two, three, four, or five million fully debugged lines of code overnight. Its initial market is rewriting legacy mainframe systems into modern cloud software; greenfield projects now begin with AI-written specifications running 100–300 pages before autonomous building and debugging.
9. The AI war is breaking old alliances and venture conventions
Perplexity’s unsolicited $34.5 billion bid for Chrome exceeds its reported valuation. Ismail initially calls it a PR play, while Diamandis identifies the strategic purpose: rebut Google’s contention that Chrome is inseparable and unwanted, thereby keeping regulators engaged in a potential breakup.
Meta reportedly paused hiring after recruiting more than 50 researchers with packages ranging from tens of millions to a reported $1 billion, then reorganized around products, superintelligence, infrastructure, and fundamental research. Microsoft responded with a “most wanted” list, multimillion-dollar packages, and offers for critical talent within 24 hours.
Blundin says the prior détente among Apple, Microsoft, and Google lasted 10–15 years because each protected a cash cow and avoided destabilizing the others. AI has erased that equilibrium: “We’ve never seen all the tech giants going after one brass ring before,” creating turbulence that may advantage startups able to recruit quickly.
Acqui-hires create the darker edge. Microsoft reportedly blocked the Windsurf transaction, after which the talent moved into Google; shareholder outcomes remained unclear. Licensing the technology, buying 49% non-voting stakes, and transferring key employees can avoid the normal waiting period while leaving “zombie startups” and inviting lawsuits.
Blundin also says multiple companies in his portfolio reached multibillion-dollar valuations in under two years, a pace he describes as historically rare. He sees the resulting tailwind for venture as enormous even if the new speed forces investors to rethink startup endgames.
10. Distribution and chips may matter more than model rankings
Elon Musk is quoted as assigning Google the highest probability of leading AI, based on its infrastructure and data. The panel cites $85 billion in AI-related capex and a reported 14% Anthropic stake; Wissner-Gross says the winning combination is enormous compute plus default distribution to billions of users.
Ismail pushes back that small, focused teams repeatedly outperform big-company research groups, making external bets such as Anthropic strategically important. Blundin locates the underlying contest in Google TPUs versus Dojo chips associated with Tesla or xAI, alongside Musk’s stated $16 billion Samsung manufacturing agreement, which Blundin characterizes as economically closer to $40 billion.
Grok’s cited app rating—4.9 across 502,000 downloads and ratings—edges Gemini’s 4.8 across 394,000 downloads and ratings. The panel attributes part of that to Musk’s aggressive use of X as a distribution channel; Google could push Gemini through Search, but doing so too hard risks cannibalizing its existing revenue engine.
Blundin makes founder visibility part of the capital stack. Musk hosting Saturday Night Live was, in his telling, a strategic recruiting and fundraising move: a transformative purpose does not recruit by itself unless someone broadcasts it. In the current market, an outspoken CEO or co-founder can pull talent toward the mission at exceptionally low media cost.
11. Chip sovereignty turns governments into direct shareholders
The US government converted $8.9 billion of Intel support into a 10% equity stake. Blundin says prior CHIPS Act funding carried enough restrictions to be “unusable money”; restructuring it as equity could give Intel more freedom while recognizing that domestic chip capacity is “too strategic to fail.”
The trade around the policy shift was dramatic as told: short-dated Intel call options priced around three cents reportedly rose to roughly $3, a 100x move. Blundin says the signal came when CEO Lip-Bu Tan emerged from a White House meeting still in place and receiving favorable comments rather than entering a public fight.
The surrounding capital confirms the strategic direction: SoftBank’s Masayoshi Son is described as investing $2 billion in Intel while SoftBank owns 90% of Arm, whose chips power 99% of smartphones. Meanwhile, Apple’s US-bound iPhone 17 production is reportedly moving to India through Foxconn, potentially breaking the perception that India cannot manufacture high-quality electronics at scale.
12. Failed enterprise pilots indict organizational design, not AI demand
The cited MIT study says businesses spent $30–40 billion on generative AI while 95% of pilots delivered no financial return. Adoption was broad—80% testing and 40% deploying—but firms struggled to scale; purchased tools succeeded roughly two-thirds of the time, while internal builds succeeded only about one-third.
Ismail identifies two failure modes: reckless use, illustrated by a medical CEO uploading sensitive patient data into ChatGPT, and cultural resistance from employees who fear displacement. Large firms also force AI into established workflows, while startups redesign the business around AI and face fewer bureaucratic vetoes.
His prescription is an independent, AI-native “edge organization” that replicates core functionality from scratch, automates use cases bottom-up, uses younger entrepreneurial teams, and reports directly to the CEO. The mother ship and edge team effectively run an A/B test until the new operation becomes the center of gravity: “Do not try and transform the mother ship.”
Blundin adds a corporate-venture channel: restart internal funds, invest in AI startups, become their first or second customer, and connect them to the edge organization. His objection to the study’s bearish reading is that pilots fail when reluctant, inexperienced teams receive tools without incentives or redesign—not because the underlying models lack economic capability.
The panel also discusses Albania’s Prime Minister Edi Rama advocating AI ministers and prime ministers as a way to address entrenched corruption. Ismail argues that AI oversight could at least monitor ministers and expose conflicts of interest, including cases involving land purchases around public infrastructure.
13. Medicine, BCIs, and robots move AI into the physical economy
OpenEvidence reportedly scored 100% on the US medical licensing exam, versus GPT-5 at 97%, and is already used by 40% of US doctors. Diamandis’s conclusion is blunt: clinicians will use systems like this or eventually cease to be competitive practicing physicians.
OpenAI’s GPT-4b reportedly designed transcription factors for reprogramming cells into stem cells that were 50x more effective. Wissner-Gross calls this an early glimpse of AI-driven longevity escape velocity, while Diamandis argues that AI is the force most likely to move health-span research at the necessary speed.
Merge Labs is described as combining gene therapy and ultrasound to read and write neurons. Wissner-Gross hopes it and roughly 20 BCI competitors can deliver high-bandwidth interfaces within a few years; otherwise, machine productivity risks forming a “pure AI economy” decoupled from humans. Altman’s preferred outcome is “a merge,” not merely friendship with a chatbot.
In robotics, Unitree’s 1,500-meter record remained 91% slower than a human, but public games create benchmarks that accelerate iteration. Figure 02 autonomously folded laundry, while a Figure/Helix demonstration showed a humanoid navigating rough debris; the robots may be reaching movements that outpace teleoperation as a debugging method.
Apple is also described as exploring AI-enabled devices including a tabletop robot, smart speakers, and AI-enhanced home-security cameras. China’s reported artificial-womb robot, priced around $14,000, is discussed alongside stationary artificial-womb work relevant to Colossal’s de-extinction efforts.
Salim notes that Star Trek’s imagined future is biotech-poor and AI-poor. The panel’s alternative 2030s horizon combines abundant intelligence, energy, and biology simultaneously.