David Blundin

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.

🔗 Original source & video: Is AI a Bubble? Experts Debate the Future of AI w/ Dave, Salim, and AWG | EP #190

GPT-5 was presented as OpenAI’s effort to combine frontier capabilities into one broadly useful product, while the interview left AGI unresolved. Kevin Weil called it OpenAI’s smartest model and emphasized coding, health, complex instruction-following, tool use, and agentic work; pricing came in at “less than half” the prior level. Yet OpenAI still cannot reliably predict what a model will unlock: capabilities appear “through the mist,” sometimes only after release.

Compute, not customer demand, is OpenAI’s binding constraint. Weil said the company uses essentially the same model settings as customers, remains “completely maxed out at all times,” and finds an immediate use for every new GPU—lower latency, faster tokens, wider product access, or more research experiments. Stargate’s more than $500 billion of planned infrastructure is therefore a capacity expansion into what he called “basically infinite demand for GPUs within these walls.”

OpenAI’s distribution strategy reverses conventional software monetization: expensive capabilities begin in paid tiers, then migrate toward free. Deep Research moved from Pro to Plus and eventually limited free access; India received a heavily discounted plan with roughly 10x the usage of a free account. Diamandis said models are now about 100x cheaper than GPT-4 was at launch even as intelligence increased, while Weil said paid tiers will retain the most computationally intensive work.

The startup test is whether better foundation models strengthen the product or erase it. Weil advised founders to build where current models show “little glimmers of hope,” so the next release makes the application “sing,” rather than patching a limitation that OpenAI may soon remove. His premise is sweeping: practically every scaled product, service, and device predates AI and “they’re all going to be reinvented.”

OpenAI’s envisioned AGI product is ambient, proactive, and capable of generating disposable software—not merely a smarter chat window. Weil expects interfaces to be created in real time, routine work to be completed before the user asks, and an assistant that can “see what you see” and keep “chugging away behind the scenes.” The Jony Ive question remained unanswered when the interview was cut.

Reasoning adds a second scaling axis beyond pretraining: how long a model can work while staying on track. ChatGPT reasoning may run for roughly 60 seconds and Deep Research for 20–30 minutes, but Weil sees no reason models could not work for days, months, or years; “the longer the models think, the smarter they get.” Well-specified problems such as chip layout can then turn compute into iterative gains because every candidate design can be scored.

AGI may arrive as a rising capability gradient rather than a clean threshold event. Weil argued that today’s models already outperform him in some domains and remain clearly inferior in others, while “the level of water is rising”; the hosts compared that transition with society barely noticing that AI had passed the Turing test. The measurement challenge is shifting from saturated, easily graded benchmarks toward ambiguous but economically valuable work.

The human hedge in an increasingly automated economy is purpose, personal connection, and deployment into consequential institutions. Weil rejected futures where people merely “eat grapes and write poetry and receive our UBI,” arguing that saved time gets redirected toward larger goals and that in-person connection will matter more. His parallel role as an Army lieutenant colonel reflects the episode’s broader deployment thesis: superior models provide little advantage “if they’re sitting on the shelf” while rivals integrate weaker systems everywhere.

🔗 Original source & video: Open AI Insider on GPT-5, AGI & the Great AI Race w/ Kevin Weil & Dave Blundin

1X’s central thesis is that the home is the fastest route to consumer scale and embodied intelligence, not merely another market for automation. Its earlier EVE robots plateaued after roughly 20–40 hours on repetitive guarding or logistics tasks, while homes have not yet shown a diversity ceiling. At 10,000 deployed robots, Bernt Bornich estimates the fleet could generate more non-duplicated useful daily data than YouTube: “The internet isn’t actually that big.”

The near-term commercial proposition could already be useful before full autonomy arrives. Peter Diamandis floated a roughly $30,000 purchase price or $300 monthly lease—$10 a day and about $0.40 an hour—and Bornich replied, “I think we could do better,” while declining to announce the actual price. The factory’s end-2026 annual run rate is planned at north of 20,000 units, although the ramp means 2026 production itself will be lower.

NEO Gamma’s industrial advantage is a deliberately simple, lightweight architecture rather than car-like complexity. The 5-foot-4-inch, 66-pound robot can reportedly lift about 150 pounds, carry roughly 50, run for four hours, and recharge from empty in about two; it contains hundreds of components versus roughly 50,000 in a car. Bornich’s manufacturing frame: “It’s closer to a refrigerator than a car.”

1X is betting that physical intelligence must be spatial, temporal, tactile, and interactive—not language-first. Internet video provides observations but not an agent’s goal, chosen action, or observed consequence; robots can instead execute the scientific-method loop of hypothesis, action, feedback, and revision. Bornich would not claim embodiment is theoretically indispensable, only that it is “a way shorter path” than text or low-fidelity simulation.

Teleoperation is part of the product and training stack, with transparency carrying much of the trust burden. Early customers will receive a mix of best-effort autonomy and scheduled, operator-assisted work; users must approve teleoperation, the robot visibly signals when a person is present, and operators see filtered scenes. Private data has a 24-hour pre-training deletion window, while human review requires approval and a user-supplied decryption key.

Safety is being bounded both physically and through action-conditioned world models. NEO Gamma is soft and intrinsically designed so an accidental strike might hurt but is unlikely to cause severe injury; cooking and other dangerous-object tasks will initially remain disabled because “once you pick up a kettle of boiling water, there’s no more guarantee that you are safe.” For model evaluation, 1X places the controller inside a simulated world where “the robot’s in the Matrix” and tests performance, red-team cases, and unsafe behavior.

The upside case is a labor-and-infrastructure flywheel, but its constraints are intensely physical. Bornich envisions the “hard takeoff moment” as robots building robots, chip fabs, data centers, energy systems, laboratories, and specialized automation; Peter framed labor as roughly half of $110 trillion in global GDP. Bornich considers 10 billion humanoids by 2040 “probably roughly correct” and possibly early, conditional on permitting, power, aluminum, rare-earth processing, magnets, fabs, and enough robots to bootstrap the required labor.

🔗 Original source & video: Robotics CEO: The Humanoid Robot Revolution Is Real & It Starts Now w/ Bernt Bornich & David Blundin | EP #188