2025-08-09

AI Insiders Breakdown the GPT-5 Update & What it Means for the AI Race w/ Emad, AWG, Dave & Salim

The panel dissects GPT-5's low-key, 'folksy' launch versus the massive pre-event hype, concluding the real story isn't a jaw-dropping new capability but a collapse in the cost of frontier intelligence and the lifting of 700 million users to frontier-model quality. They tour the benchmark landscape (LM Arena, ARC-AGI, Frontier Math, HLE), debate whether labs are deliberately holding back their best internal models, and cover GPT-5's health and agentic demos, OpenAI's open-weight release, and the coding-tools realignment around Cursor. The back half surveys the wider AI war: Google's two-week blitz (Gemini 3, Genie 3, Alpha Earth), Meta's talent-poaching war and OpenAI's $1.5M employee bonuses, OpenAI's $500B valuation, and compute/chip geopolitics (Stargate Norway, Apple's US investment, the Intel CEO controversy, TSMC concentration).

▶ Watch on YouTube

Topics

GPT-5 launch reception AI ▶ 0:03

Panel reacts to GPT-5's low-key, 'folksy' launch event versus sky-high hype; agrees the real impact is lifting hundreds of millions of users to frontier-level AI and collapsing costs rather than a flashy new capability.

Benchmark landscape (LM Arena, ARC-AGI, Frontier Math, HLE) AI ▶ 16:41

GPT-5 debuts #1 on LM Arena and sets a new cost-performance frontier, but trails Grok on ARC-AGI and Humanity's Last Exam, suggesting labs are 'pulling punches' on their most advanced internal models.

Hyperdeflation of AI cost Compute ▶ 53:44

GPT-5 cut inference costs roughly in half to an order of magnitude versus GPT-4.5, driven by new Blackwell/GB200 chips, distillation, and architecture gains; panel argues labs are near break-even, not burning cash.

Vibe coding and the coding-tools war AI ▶ 33:52

A live GPT-5 'vibe coding' demo (building a French-learning app) triggers discussion of Cursor's new alliance with OpenAI after the Windsurf deal collapsed, and Anthropic's coding lead being directly challenged.

AI in healthcare and Fountain Life Health ▶ 26:17

Diagnostic AI now outperforms human physicians and even physician+AI teams; Peter and Salim discuss Fountain Life's full-body data uploads as the key unlock for proactive, personalized health AI.

Agentic executive-assistant demo AI ▶ 49:06

A ChatGPT demo managing email, calendar, and packing lists sparks debate over whether such tools are genuinely useful or just 'coolish'; panel agrees the real unlock is at the executive/strategic-planning level.

Open-weight models and collapsing training costs Compute ▶ 1:06:23

OpenAI's new open-weight model cost about $4M to train (a 20B-parameter laptop-runnable version ~10x cheaper); panel debates whether that reflects true from-scratch cost or distillation from a larger teacher model, and predicts sub-$1M frontier training within two years.

Google's two-week product blitz AI ▶ 1:17:37

Google shipped Gemini 3, Gemini 2.5 Pro Deep Think, Genie 3 (real-time generative world models), and Alpha Earth Foundations (planet-scale satellite mapping) within weeks, reportedly spurred by competitive pressure from GPT-5.

Meta's talent war and compensation shock AI ▶ 1:28:13

Zuckerberg's Meta Superintelligence Labs contacted 100+ OpenAI staff with huge offers and got only ~10% to move; meanwhile Altman gave every OpenAI employee $1.5M in bonuses over two years.

OpenAI valuation, IPO, and compute buildout Economy ▶ 1:03:59

OpenAI is eyeing a $500B valuation on ~$10B revenue versus Microsoft's $300B revenue; panel discusses Stargate Norway's hydro-powered data center, Apple's $600B US investment, and the Intel CEO/China controversy as parts of a broader AI infrastructure and chip-sovereignty race.

Will college survive the intelligence shock? Other ▶ 1:02:16

Prompted by Sam Altman's 'college in 2035, if it exists' remark, the panel (several with teenage sons) speculates the university system could implode within five years as AI commoditizes expertise.

Predictions made

open Alex Wissner-Gross: Extrapolating Frontier Math Tier 4 scores in a straight line, AI will solve about 70% of hard math problems by the end of 2027 (roughly 15-20% by end of 2025, 35-40% by end of 2026).

· due: 2027 · ▶ watch

“by again the law of straight lines... projected forward to 2027, end of year, you get to 70%”

Your call:

open Alex Wissner-Gross: A GPT-5-level model will be trainable end-to-end for under $1 million within two years.

· due: 2027 · ▶ watch

“I think you will have a GPT5 level model by 2 years max that will cost under a million dollars to train end to end.”

Your call:

open Emad Mostaque: Training a GPT-OSS-class open model (currently about $4M) will drop to roughly $400,000 next year.

· due: 2026 · ▶ watch

“Next year it will cost $400,000 to train a model”

Your call:

open Elon Musk (quoted tweet): Grok 5 will ship before the end of 2025 and will be 'crushingly good.'

· due: 2025 · ▶ watch

“Grock five will be out before the end of the year, and it will be crushingly good.”

Your call:

open Emad Mostaque: Crypto x AI will produce the biggest startup bubble ever next year, as small teams get massive traction using frontier models and become instantly tradeable.

· due: 2026 · ▶ watch

“next year Crypto XAI is going to be the most ridiculous thing you've ever seen”

Your call:

open Alex Wissner-Gross: OpenAI's $1.5M-per-employee bonus windfall will spark an enormous wave of new startup seed funding over the next one to two years.

· due: 2027 · ▶ watch

“I would expect this to create a a bloom of seed funding of startups in in the next year or two.”

Your call:

open Dave Blundin: OpenAI will reach $100-150 billion in annual revenue within two years, per Sam Altman's own projection that Dave relayed and said he doesn't doubt.

· due: 2027 · ▶ watch

“Sam's Sam's projection is 100 to 150 billion in revenue in what is it two years from today”

Your call:

open Emad Mostaque: By next year, continuous personal health data (e.g. Fountain Life-style full-body uploads) combined with AI will let AI proactively flag health issues before a diagnosis is even sought.

· due: 2026 · ▶ watch

“by next year I think we'll be at the point where the key thing is you get the right data”

Your call:

Numbers that matter

GPT-5 cut AI inference cost by at least half, possibly more vs GPT-4.5 and prior frontier pricing
GPT-4.5 cost $75/million input tokens and $150/million output tokens compared to GPT-5's roughly $1.25 input / $10 output per million tokens
OpenAI and Anthropic each have about $3B in API revenue; $1.4B of Anthropic's is API, about half from Cursor and Microsoft Copilot frames GPT-5's ~40% lower price than Claude Sonnet as a direct attack on Anthropic's revenue
OpenAI's new open-weight model cost about $4M to train (2M H100 hours, 80 trillion tokens); the 20B-parameter laptop version was ~10x cheaper released model described as better than any model from a year earlier
Frontier Math Tier 4 solve rate projected at 15-20% by end of 2025, 35-40% by end of 2026, 70% by end of 2027 Alex Wissner-Gross's straight-line extrapolation of hard-math benchmark progress
On OpenAI's Healthbench, doctors scored about 20% while the latest models score 60-70% evaluation built with 250 physicians on real-world health tasks
Physicians alone diagnose correctly about 80% of the time, physicians+AI together about 90%, AI alone about 93% study cited showing human involvement can reduce accuracy by reintroducing bias
OpenAI is eyeing a $500B valuation on about $10B/year revenue, versus Microsoft's roughly $300B/year revenue panel calls the ratio 'overpriced' given the revenue gap
Sam Altman gave every OpenAI employee $1.5M in bonuses over two years; 78% of Nvidia employees are already millionaires response to Meta's poaching attempts
Zuckerberg's Meta contacted 100+ OpenAI employees with huge offers; about 90% turned him down attributed to belief that OpenAI is closer to AGI than Meta
Stargate Norway: a $2B data center with 100,000 Nvidia GB300 chips, 230MW capacity expandable to 520MW, powered 100% by renewable hydro OpenAI/Nvidia's first European data center
TSMC holds roughly two-thirds (66%) market share of the chips driving the AI revolution cited as a dangerous single point of failure, raising Intel's national-priority status
AI data-center capex is 1.2% of US GDP in 2025, versus railroads at 6% of GDP in the 1880s and telecom at 1% around 2000 used to argue the AI infrastructure buildout is still early

Worth digging into

🕳️ Frontier Math Tier 4 extrapolation to 70% by 2027

Alex Wissner-Gross's straight-line projection is a specific, falsifiable claim about AI solving professional-level math within two years, a clean marker for tracking.

🕳️ Genie 3 and decoder-only Earth models

Alex frames Alpha Earth's encoder-only model as a precursor to decoder-only models that could simulate future land-use outcomes via tree search, a potentially huge but under-discussed application.

🕳️ True cost of OpenAI's open-weight model

Emad cites a $4M 'from scratch' training cost, but Alex disputes this, suspecting synthetic data distilled from a far more expensive teacher model, a distinction that matters for judging real frontier-training-cost trends.

🕳️ OpenAI's $1-per-agency-per-year federal rollout

A land-grab-priced push of ChatGPT into every US federal agency raises questions about government AI dependency, procurement precedent, and competitive response from Google/Anthropic.