Open AI Insider on GPT-5, AGI & the Great AI Race w/ Kevin Weil & Dave Blundin
Peter Diamandis and Dave Blundin sit down with OpenAI chief product officer Kevin Weil at OpenAI's San Francisco HQ to unpack the GPT-5 launch, OpenAI's 'iterative deployment' philosophy, and why the company believes GPUs will never commoditize given Stargate's roughly $500B buildout. Weil describes OpenAI's unusual habit of pushing paid features down to free over time, a steep new India-only discount plan meant to turn latent coding talent into builders, and his vision of AGI as a proactive, UI-generating assistant rather than just a smarter chatbot. The group also digs into saturating AI benchmarks, the two axes of scaling (pretraining vs. test-time compute), and AI's growing role in designing its own chips. Weil reveals he's joined a new Army program alongside Palantir's and Meta's CTOs to connect tech executives directly with the DoD, and the conversation closes on education, brain-computer interfaces, and Dave Blundin's timeline for Moon, Mars, and fusion milestones.
Kevin Weil reflects on launching GPT-5 as OpenAI's most anticipated release, highlighting its coding strength, agentic tool-use, and new health-data features meant to help with everyday medical questions.
OpenAI's strategy is to ship capable-but-imperfect AI to the public as soon as it's ready rather than hold back a polished, fully-formed product, using real-world feedback to improve faster.
OpenAI says it is permanently GPU-constrained and is building out Stargate (reportedly a $500B infrastructure push) because demand for compute inside the company is effectively infinite.
Weil argues AI coding tools can grow the world's effective developer population from 30 million toward billions, and describes a steeply discounted India-only ChatGPT plan meant to spread that leverage to underserved talent.
Unlike typical SaaS, OpenAI moves features down from Pro to Plus to free over time (e.g. deep research), aiming to make its most powerful product available to as many people as possible.
Weil's advice to startups building on OpenAI's models: build at the bleeding edge of what today's model can barely do, not around its current limits, so the next model release makes the product 'sing' instead of obsoleting it.
Weil says AGI won't just be a smarter chat window; it will generate custom interfaces on the fly and become anticipatory, acting on context it gathers about the user rather than waiting for prompts.
With many classic AI benchmarks now saturated, OpenAI is building harder and 'softer' evals (health, financial modeling, creative writing) tied to its definition of AGI as the ability to do economically valuable human work.
Weil and Blundin discuss how scaling now has two axes, pretraining and test-time 'thinking' time, with no observed ceiling yet on how much smarter models get the longer they're allowed to reason.
OpenAI is using AI to help design and lay out its own custom chips, echoing Google's public TPU self-design work, closing the historic gap between algorithm teams and hardware teams.
Blundin lays out a rough timeline for human expansion into space: recurring Starship Moon trips by 2026-27, robotic (then human) Mars landings around 2030, and fusion plants from Helion (~2028) and Commonwealth Fusion (~2030) that could power it.
Diamandis criticizes schools that ban AI outright, arguing kids should instead be given much harder, AI-assisted assignments (per Ethan Mollick's 'Co-Intelligence') so they go deeper than a classroom alone allows.
Predictions made
openDave Blundin: Starship will carry humans to the Moon on a recurring basis.
openDave Blundin: AI will soon generate full chip designs directly from a software/algorithmic specification, becoming the next big unlock in hardware.
“it'll do the chip design right off your software design. That's going to be the big unlock.”
Your call:
openKevin Weil: AGI will likely be recognized as having arrived only in hindsight, the same way society passed the Turing test without really noticing.
“I think I really think so cuz arguably AGI is here.”
Your call:
Numbers that matter
GPT-5 was rolled out to 700 million ChatGPT usersWeil describes how shipping to that scale surfaces usage patterns internal testing never would, which then feeds back into rapid fixes.
New India-only ChatGPT plan gives ~10x the usage of the free tierA heavily discounted plan launched the same day as the interview, aimed at spreading AI access to latent talent in India.
Stargate infrastructure buildout reportedly has a $500 billion budgetCited from public statements (including at the White House) about OpenAI's compute buildout with partner companies.
Models have gotten roughly 100x cheaper even as they've become far more capableComparing current GPT-5 pricing to GPT-4-era pricing despite the newer model being far more intelligent.
Only about 30 million people worldwide can code todayWeil's starting point for arguing AI coding tools could expand that to 300 million or even 3 billion effective coders.
37 venture-backed fusion-energy startups now existKevin Weil's surprised reaction to Dave Blundin's count, cited as evidence deep tech investment has had a renaissance.
GPT-5 Pro measured at roughly IQ 148 on a Mensa-style scaleRaised as one marker of how far model capability has climbed.
The US scored 32-36 of 42 points at the 2025 International Math Olympiad, placing second; China's team had a perfect score across all six problemsUsed to set up Alex Wissner-Gross's practice of tracking IMO performance as a self-improvement benchmark.
Algorithmic efficiency gains were previously 'swagged' on the podcast at 100x-1,000x, with new evidence suggesting it could be even higherDistinguished from raw compute/hardware scaling as a separate, less-discussed driver of AI capability growth.
Worth digging into
🕳️ OpenAI's proposed 'abundance benchmarks'
Diamandis and Weil discuss replacing saturated academic benchmarks with a proof-of-work-style measure of AI's ability to solve grand challenges (health, poverty, etc.), which could reshape how AGI progress gets tracked publicly.
🕳️ Tech executives joining a new Army program
Weil, Palantir CTO Sham Sankar, Meta CTO Andrew Bosworth, and ex-OpenAI research lead Bob McGrew all joined the Army as officers with individual focus areas (Weil's is AI for physical performance), a notable deepening of Silicon Valley-DoD ties.
🕳️ Merge Labs' brain-computer interface
Weil says he'd personally get a BCI 'when it's safe' and references a friend involved with the newly leaked/announced Merge Labs project, without confirming whether its approach is invasive, subdural, or external.
Blundin references a prior podcast 'swag' estimate that software/algorithmic improvements (separate from raw compute scaling) have delivered 100x-1,000x gains, with new evidence suggesting it could be higher still -- a huge and under-discussed driver of capability growth.