2025-08-27

Is AI a Bubble? Experts Debate the Future of AI w/ Dave, Salim, and AWG | EP #190

Fresh off a week at OpenAI headquarters, the panel races through the AI-bubble debate sparked by Sam Altman's dinner remarks and a leaked $500B secondary-sale valuation, with Dave Blundin flatly denying a bubble while Salim Ismail reads it as PR damage control and Alex Wissner-Gross ties the debate to capex-versus-revenue growth. They cover GPT-5 Pro's saturating IQ benchmarks, a distillation paper claiming 100x training efficiency, GPT-5 producing new math proofs, and OpenAI's 'land grab' expansion into India and the UK alongside its Stargate/Norway data-center buildout. Talent-war coverage spans Meta's stalled hiring blitz, Microsoft's counter-offensive, and the rise of 'zombie startup' reverse-acquihires (Windsurf, Character.AI) that the group warns could break the VC model. The back half moves through the MIT 95%-AI-pilot-failure study and exo-organization advice for incumbents, Sam Altman's new BCI venture Merge Labs, GPT-4b's redesign of Yamanaka factors for longevity, and closes on China's humanoid robot Olympics and Figure AI's laundry-folding robot.

▶ Watch on YouTube

Topics

Is there an AI bubble? Economy ▶ 36:29

Sparked by Sam Altman's dinner remarks and news of an employee secondary sale at a $500B OpenAI valuation, the panel splits: Dave says absolutely not a bubble, Salim reads it as market-cap PR management, Alex ties sustainability to continued revenue growth against massive capex.

GPT-5 Pro IQ and benchmark saturation AI ▶ 4:37

GPT-5 Pro scores ~148 on the Mensa Norway IQ test (up from GPT-o3's ~120-136); Alex argues human-population-based benchmarks are saturating and new specialist-knowledge benchmarks are needed.

Local frontier AI on consumer hardware Compute ▶ 6:51

A single RTX 5090 (~$2,500) is projected to run frontier-matching LLMs locally within 6-12 months, which Alex argues is really about enabling low-latency, offline foundation models embedded in humanoid robots rather than just privacy.

Distillation and the 'overhang' framing AI ▶ 13:19

A 32B-parameter model breaks the Pareto frontier on AIME24/25 using roughly 1% of the training data of larger models, which Alex and Dave frame as a new class of unexploited capability 'overhang' with huge implications for smaller AI startups.

AI predicting (and retrodicting) the future AI ▶ 16:34

Discussion of GPT-5's Brier-score forecasting ability leads Alex to propose 'retrodiction' or 'quantum archaeology' — using AI to reconstruct history to high fidelity — and Dave suggests predicting the next 20 Nobel Prize discoveries as a benchmark.

GPT-5 Pro produces new math and science AI ▶ 22:33

GPT-5 Pro independently proves a new result in a convex optimization paper; Alex calls this the leading edge of AI 'bulk solving' math, science and engineering, predicting a trickle-to-tidal-wave shift within roughly a year.

OpenAI's global expansion 'land grab' Geopolitics ▶ 27:06

OpenAI is expanding into India (its likely largest market) and in talks to give the UK population access to ChatGPT Plus; Dave notes OpenAI opened a New Delhi office instead of Boston/NYC, calling it a deliberate demographic and economic play.

OpenAI's data center buildout Compute ▶ 33:19

OpenAI is building Stargate Texas (up to 5GW) and a Norway hydro-powered mega center (290MW, 100,000 GPUs), while also developing custom chips; Dave reports OpenAI is deliberately throttling visible capability because compute is fully constrained.

Big Tech AI talent wars Economy ▶ 49:26

After Meta's hiring blitz (multi-million to reportedly billion-dollar packages) it paused and reorganized into four AI groups; Microsoft is now counter-recruiting with a 24-hour offer process, which the panel likens to a Star Wars-style escalation good for startup turbulence.

'Zombie startup' reverse-acquihires Economy ▶ 53:22

Deals like Windsurf (Microsoft blocked OpenAI's bid, talent went to Google) and Character.AI's absorption into Google use non-voting equity stakes and undisclosed licensing deals to sidestep antitrust review, leaving shareholders and remaining employees in limbo.

MIT '95% of AI pilots fail' study and exo organizations AI ▶ 1:19:43

A widely-cited MIT study finds 95% of corporate generative-AI pilots deliver no financial return despite $30-40B in spend; Dave and Salim argue the real fix is building an AI-native 'exo' edge organization reporting directly to the CEO rather than forcing AI into legacy workflows.

Sam Altman's Merge Labs and the BCI race Health ▶ 1:28:15

Sam Altman is entering brain-computer interfaces via Merge Labs, combining gene therapy and ultrasound to read/write neurons; Alex frames high-bandwidth BCI as the best hope for keeping the human and AI economies coupled.

GPT-4b and AI-driven longevity Longevity ▶ 1:31:09

OpenAI's GPT-4b redesigned Yamanaka reprogramming factors to be 50x more effective at turning cells into stem cells, which the panel calls an early glimmer of AI-driven longevity escape velocity.

Humanoid robotics: China's robot Olympics and Figure AI Robotics ▶ 1:34:36

China staged the first humanoid robot games (hip-hop, soccer, boxing, track) in Beijing; Unitree set a 1500m robot world record still 91% slower than humans, while Figure AI's fully autonomous Figure 2 folds laundry and navigates rough terrain using its in-house Helix model.

Predictions made

open Alex Wissner-Gross: AI 'bulk solving' math, science and engineering discoveries will go from a trickle to a tidal wave.

EP #? · · due: end of 2025 (possibly) · ▶ watch

“the trickle I think is going to turn into a title wave over the next year or so possibly by the end of this year”

Your call:

open Salim Ismail: The Trump administration will get involved in the Google/Chrome antitrust matter and demand a stake (e.g., 10%) in exchange for protection.

EP #? · · due: unspecified · ▶ watch

“here's what I predict happens. I think the Trump administration gets involved and says, well, you want to give us 10% of Chrome to help you help you keep protective”

Your call:

open Elon Musk (clip): AI/robot companionship will counterintuitively increase the human birth rate because AI will be deliberately programmed to encourage it.

EP #? · · due: unspecified · ▶ watch

“I predict counterintuitively that it will increase the birth rate. We're going to program it that way.”

Your call:

open Alex Wissner-Gross: Within a few years, context windows will become effectively infinite, potentially making RAG and fine-tuning obsolete.

EP #? · · due: a few years · ▶ watch

“one can reasonably foresee a few years out maybe we'll have effectively infinite context”

Your call:

open Alex Wissner-Gross: The most futuristic frontier AI model will be a tiny 'nano kernel' (possibly just millions of parameters) with externalized knowledge, rather than a huge dense model.

EP #? · · due: next few years · ▶ watch

“if I had to predict what is the most futuristic possible frontier AI model look like, I think it probably will look like some sort of nano kernel”

Your call:

open Salim Ismail: Humanoid robot competitions/games (like China's robot Olympics) will fail to gain lasting audience interest because people watch the Olympics for the human element, not raw speed.

EP #? · · due: unspecified · ▶ watch

“I will say make a prediction that these type of robot games will fail just because we watch the Olympics for the human factor not for the speed”

Your call:

open Ray Kurzweil (clip, referenced by Peter Diamandis): High-bandwidth brain-computer interfaces will be achieved.

EP #? · · due: 2033 · ▶ watch

“when Ray made his prediction of high bandwidth computer brain computer interface by 2033, I was like, Ray, you got this one wrong... no, he got this one right”

Your call:

open Dr. Singularity (clip, read by Peter Diamandis): A Star Trek-like world of advanced AI and technology will arrive far sooner than science fiction imagined.

EP #? · · due: 2030s · ▶ watch

“A Star Trek world wouldn't wait for the 2,200s. It would arrive by the 2030s.”

Your call:

Numbers that matter

GPT-5 Pro scores ~148 on the Mensa Norway IQ test Up from GPT-o3's ~120-136; used as evidence that human-population-based benchmarks are saturating.
A single Nvidia RTX 5090 (~$2,500) may run frontier-matching LLM performance locally within 6-12 months Framed as a precursor to low-latency, locally-run foundation models in humanoid robots.
New 32B model matches larger models' performance using ~1% of the training corpus (100x efficiency) From a data-efficient distillation paper breaking the Pareto frontier on AIME24/25.
Panel's prior estimate: software-only rapid AI improvement of 100x to 10,000x Dave Blundin says this was likely a lower-bound estimate given newly observed 100x gains just from data selection/distillation.
OpenAI's Stargate Texas up to 5GW capacity; Norway mega center 290MW / 100,000 GPUs powered by hydro Data centers increasingly measured in power (gigawatts), not GPU count.
OpenAI developer usage up ~50% week-over-week after GPT-5 launch; agentic-behavior tokens nearly doubled; reasoning usage up 8x Cited by OpenAI CFO Sarah Friar (clip) on CNBC amid the AI-bubble discussion.
GPT-5 uplifts 700 million people into free/reasoning-capable models Cited as a massive jump in global AI access with recursive productivity effects.
Claude Sonnet 4 now supports 1 million tokens of context (~750,000 words, 3,000-4,000 pages) Compared to Grok's ~500-600K tokens and GPT-5's ~250K tokens.
Perplexity makes an unsolicited $34.5 billion bid for Chrome Larger than Perplexity's own reported valuation; seen as a strategic move amid Google antitrust pressure.
xAI/Tesla signs a $16 billion (arguably ~$40 billion) chip manufacturing deal with Samsung for Dojo chips Framed as the real infrastructure race underlying Elon Musk conceding Google's AI lead.
Grok app store rating 4.9 (502,000 ratings) vs. Google Gemini 4.8 (394,000 ratings) Cited despite Elon Musk publicly conceding Google's AI lead, attributed partly to Musk's social media reach.
Google's Gemma 3 (270 million parameters) handles 25 chats using just 1% of battery on a Pixel 9 Pro Used to argue future frontier models may trend toward small, power-efficient 'nano kernels.'
US government converts an $8.9 billion CHIPS Act grant to Intel into a 10% equity stake Intel stock rose on the news; described as turning previously unusable, string-attached grant money into usable investment.
SoftBank signs a $2 billion investment deal; SoftBank owns 90% of ARM, whose chips power 99% of smartphones Cited alongside Intel and Samsung investments as evidence of a US native chip capacity push.
MIT study: companies spent $30-40 billion on generative AI, yet 95% saw no financial return; 80% are testing, 40% deployed Attributed largely to a 'learning gap' — internally built AI succeeds only ~33% of the time vs. ~67% for bought solutions.

Worth digging into

🕳️ Sam Altman's actual dinner remarks on an AI bubble and the $500B secondary sale

The whole episode's framing device is a reported private dinner comment plus a leaked employee stock sale valuation; the panel debates motive without ever seeing the primary source.

🕳️ The data-efficient distillation paper (32B model beating Pareto frontier on AIME24/25)

Alex and Dave treat this as evidence of a major 'overhang' enabling 100x training efficiency, with big implications for startups competing with frontier labs.

🕳️ Zombie-startup reverse acquihire structure (Windsurf, Character.AI)

Dave describes a specific legal structure (49% non-voting stake + undisclosed licensing deal) used to dodge antitrust review, which he says could break the VC funding model if it becomes standard.

🕳️ Merge Labs' gene-therapy + ultrasound BCI approach

Sam Altman's new BCI venture proposes reading/writing neurons via a combination of gene therapy and ultrasound rather than Neuralink's implant approach — a materially different technical bet.

🕳️ GPT-4b's redesign of Yamanaka reprogramming factors

A 50x improvement in cell-reprogramming factor effectiveness is framed as an early sign of AI-driven longevity escape velocity, tying directly into Demis Hassabis's cell-simulator work at Google.