A man bet on Paris weather temperatures on Polymarket, then drove to the measuring station and used a hairdryer to heat the sensor — netting $34,000.
God Mode Podcast
A man bet on Paris weather temperatures on Polymarket, then drove to the measuring station and used a hairdryer to heat the sensor — netting $34,000.
TL;DR
Three hosts of the God Mode Pod break down the wildest week of AI news in 2026: Anthropic finally admitted three stacked product bugs made Claude dumber for six weeks, GPT 5.5 launched with impressive benchmarks but an alarming 86% hallucination rate, and SpaceX secured a call option to buy Cursor for $60 billion. They also cover Claude Design eating compute, Naval's retail AI fund, and a man who hacked Polymarket with a hairdryer. Key takeaway: hallucination rate, not benchmark scores, is the metric that actually matters for coding agents.
A weekly AI news roundup covering Anthropic's admission of three product-layer bugs that made Claude dumber, the SpaceX/Cursor acquisition option deal, GPT 5.5's launch and high hallucination rate, Claude Design's compute-hungry debut, Naval's USVC retail AI fund, the Polymarket hairdryer weather hack, flying cars from XPeng, and Intel stock surging after Trump administration's early stake.
Chapter 3 · 04:42
At a one-hour AI hackathon, no two people used the same setup. But these tools appeared everywhere.
At a one-hour build sprint where people picked random ideas from a hat, every team used a different AI setup. But beneath the surface, the same infrastructure appeared everywhere: Firecrawl for scraping, Claude or Codex for the model. Even a lawyer with no tech experience shipped a landing page.
Chapter 4 · 07:40
Anthropic secretly made Claude dumber for 6 weeks. Here's exactly how.
Anthropic ran three separate product-layer bugs simultaneously from March 4 to April 16: reasoning quietly dropped from high to medium, a caching bug killed idle thinking, and a system prompt capped all responses at 100 words. None of it was a model change. All of it was scaffolding. Users noticed. Anthropic didn't tell them.
Anthropic confirmed three separate product-layer bugs ran simultaneously from March 4 to April 16, silently degrading Claude's performance for users.
Claude Code's default reasoning was quietly downgraded from 'high' to 'medium' for an entire month without users being informed.
On April 16, Anthropic accidentally added a system prompt instructing Claude to keep all responses under 100 words, truncating reasoning chains across the entire product.
OpenAI gave users 10x normal token usage until May 31. After that, throttle. Anthropic's $100 and $200 tiers give you enough to build your entire workflow around Claude Code. Once your business depends on it, you'll pay whatever they ask. It's a gateway drug strategy.
Chapter 5 · 12:40
Claude fails 1 in 100 times. That's why enterprises are nervous.
Claude's uptime is 99% — that means it fails 1 in every 100 times you open it. OpenAI sits at 99.99% over the same four months. For any enterprise running production workflows on Claude, that gap is disqualifying. Being the smartest model means nothing if you're unreliable.
Claude's uptime sits at roughly 99% versus OpenAI's 99.99%, meaning Claude fails to respond 1 in every 100 times — a serious reliability gap for enterprise use.
Chapter 6 · 15:35
Claude Design ate half Rick's monthly limit in ONE task. This is a problem.
Rick uploaded his project to Claude Design and by the time one task was done, half his monthly usage was gone. Anthropic is rolling out tool after tool — Claude Code, Claude Design — while already struggling to maintain 99% uptime. The strategy is growing faster than the infrastructure.
Chapter 7 · 19:10
They need to make a decision on whether they're going to be a provider, like infrastructure, or a software layer.
Chapter 8 · 23:02
There are no dark GPUs. Dario is playing compute too conservatively.
The dot-com bust left miles of dark fiber, but AI has no equivalent. Every GPU available is being consumed. Elon already figured this out and is selling surplus compute from xAI. Dario is playing it too safe — in this market, overbuilding compute is not a risk.
Anthropic reached an annualised revenue run rate of $30 billion by early 2026, well ahead of its own internal forecasts.
Chapter 9 · 24:15
There is no dark GPU. Every GPU that is available out there is being utilized. So Elon, he is starting to sell his compute to other companies, sold it to Cursor.
Chapter 10 · 26:00
SpaceX just bought an option to acquire Cursor for $60B. Here's why.
SpaceX locked in the right to buy Cursor for $60 billion by year-end, with a $10 billion exit fee if they don't. Cursor's last funding round valued it at $50 billion. This isn't really about the software — it's about acquiring the talent after Elon fired XAI's founding team.
SpaceX secured a call option to acquire Cursor for $60 billion by end of 2025, with a $10 billion payment due if the option is not exercised.
Coding is a self-reinforcing flywheel. Build a great coding agent, companies hand you their entire codebase, the model learns how real products are built, and you use that to train an even better model. Anthropic cracked this first. Now everyone from Elon to OpenAI is chasing it.
Cursor's most recent funding round valued the company at $50 billion, and SpaceX's call option to acquire it is priced at $60 billion.
Polymarket put the probability of SpaceX acquiring Cursor at 73% at the time of recording.
Chapter 11 · 30:30
XAI buying Cursor might destroy the very thing that makes Cursor valuable.
Cursor works because it's Switzerland — plug in Claude, Gemini, GPT, whatever model you want. The moment XAI acquires it and pushes Grok as the default, every user who chose Cursor for Claude access has a reason to leave. The acquisition may destroy the thing it's buying.
Chapter 12 · 33:10
Every AI company is being valued at exactly 30x revenue. Here's the data.
Scale AI had $800M revenue and sold for $30B. Cursor has $2B revenue and is being bought for $60B. Anthropic is valued at $80-90B on $30B revenue. The market has locked in a 30x price-to-revenue multiple across AI companies, regardless of profits.
The market is consistently applying a 30x revenue multiple to AI companies regardless of profitability — Scale AI at $800M revenue sold for $30B; Cursor at $2B sold for $60B.
Chapter 13 · 34:25
SBF's portfolio is worth $114 billion. He's in prison.
SBF's early seed in Cursor is worth $3B (15,000x return). His SpaceX stake is up 75x. His Anthropic stake is worth $82B. His total estimated portfolio value: $114 billion. That's 3x Coinbase's entire market cap — from a man serving a prison sentence.
SBF's early investment in Cursor when it had only two or three founders and no product would be worth $3 billion — a 15,000x return — if SpaceX acquires it.
SBF's estimated portfolio value — including his stakes in Cursor, SpaceX, and Anthropic — is approximately $114 billion, roughly 3x Coinbase's total market cap.
Chapter 15 · 38:10
GPT 5.5 breaks your existing prompts. Here's why it's a completely different model.
GPT 5.5 is not a post-training update. It's a new base model with new architecture, new pretraining corpus, and new agent-oriented objectives, co-designed with Nvidia on 100,000 Blackwell GPUs. Your GPT 5.4 prompts will not work the same way. Switching models will break your workflows.
GPT 5.5 doubled the price compared to its predecessor, costing $5 per million input tokens and $30 per million output tokens.
GPT 5.5 tops the leaderboards but hallucinates 86% of the time. Claude Opus 4.7 hallucinates 52% of the time. Grok is at 17%. For a coding agent, hallucinations don't just give wrong answers — they break your actual tools. Benchmarks are the wrong metric.
GPT 5.5 has an 86% hallucination rate versus 52% for Claude Opus 4.7, making it a risky choice for coding agents where accuracy is critical.
Grok's hallucination rate of just 17% makes it the most factually reliable of the major models, even if it lags behind in raw coding ability.
Chapter 16 · 43:20
OpenAI killed Sora, lost 3 execs, and only 2 original co-founders remain. Here's why.
Three senior OpenAI executives left after the company shut down Sora. Only two of the original 11 co-founders remain. The entire AI industry has converged on one conclusion: coding models are what wins the race. Everything else — voice, images, Sora — is a side quest.
Chapter 18 · 49:40
Polymarket says 14% chance the AI bubble bursts by end of 2026. Here's what that means.
Polymarket defines the AI bubble bursting as any three of: Nvidia down 50%, semiconductor ETF down 40%, OpenAI or Anthropic bankruptcy, OpenAI acquired, or H100 rentals hitting $1 or less. Probability by December 2026: 14%. Not zero. The hosts put it closer to 5-10%.
Polymarket puts the probability of the AI bubble bursting by December 1, 2026 at 14%, requiring at least three of five defined trigger events.
Chapter 19 · 51:20
Avis stock went up 600% then crashed 52% in one day. Here's the absurd reason.
Avis car rental has $25B in debt and lost $995M last year. Two hedge funds combined equity swap holdings exceeded the entire outstanding share supply. Wall Street Bets piled in. The stock went from $200 to $713 — then crashed 52% in a single day. Classic bubble mechanics.
Avis car rental stock surged 600% in a month due to hedge funds holding over 100% of shares via equity swaps, then crashed 52% in a single day.
Chapter 20 · 53:15
China is mass producing flying cars by 2027. They have 7,000 pre-orders.
XPeng has 7,000 pre-orders for a vehicle that drives on roads and detaches a flying drone when terrain blocks the way. Mass production target is 2027. China has the battery tech, the drone expertise through DJI, and the regulatory appetite to make this happen first.
Chinese EV maker XPeng has 7,000 pre-orders for its flying car and is targeting mass production in 2027, pending regulatory approval.
Chapter 21 · 56:30
Naval just opened AI pre-IPO investing to everyone with $500. Catch: 5% fees.
Naval Ravikant launched USVC, a fund letting retail investors buy a basket of pre-IPO AI companies for as little as $500. The portfolio includes xAI at 20% and Anthropic. But fees stack to nearly 5%, including a 3% sales load. And non-US investors still face access barriers.
Naval Ravikant's USVC fund allows retail investors to access pre-IPO AI companies for as little as $500, but carries fees approaching 5% including a 3% sales load.
Chapter 22 · 59:30
Intel's stock has risen 185% since the Trump administration took a 10% stake at $25 per share, with Q1 earnings sparking an additional 15% surge.
Intel's stock has risen 185% since the Trump administration took a 10% stake at $25 per share, with Q1 earnings sparking an additional 15% surge.
Chapter 23 · 1:02:15
Someone hacked Polymarket with a hairdryer and made $34,000. No joke.
A bettor placed a wager on Paris temperature on Polymarket, then drove to the official weather measuring station and held a hairdryer up to the sensor. He made it artificially warm enough to win his bet and pocket $34,000. Prediction markets have a physics problem now.
A bettor manipulated a Polymarket weather market by heating the actual measuring station in Paris with a hairdryer, earning $34,000 in profit.
A soldier used classified knowledge of the Maduro raid to bet on Polymarket and make $400,000. He got arrested. Meanwhile, a Coinbase exec read the Polymarket keyword list aloud during an earnings call, ensuring all bets resolved in betters' favor. The platform is a legal grey zone.
Chapter 25 · 1:10:00
Coding an app is the new starting a podcast. Everything is code. I recently realized everything is code.
No indexed bits in this chapter.
A bettor placed a wager on Paris temperature on Polymarket, then drove to the official weather measuring station and held a hairdryer up to the sensor. He made it artificially warm enough to win his bet and pocket $34,000. Prediction markets have a physics problem now.
GPT 5.5 tops the leaderboards but hallucinates 86% of the time. Claude Opus 4.7 hallucinates 52% of the time. Grok is at 17%. For a coding agent, hallucinations don't just give wrong answers — they break your actual tools. Benchmarks are the wrong metric.
Coding is a self-reinforcing flywheel. Build a great coding agent, companies hand you their entire codebase, the model learns how real products are built, and you use that to train an even better model. Anthropic cracked this first. Now everyone from Elon to OpenAI is chasing it.
This episode
Factual claims made this episode, and whether a source was named.
Anthropic confirmed three simultaneous product-layer bugs degraded Claude from March 4 to April 16, none of which were model changes.
Claude Code's default reasoning level was silently dropped from 'high' to 'medium' for approximately one month without user notification.
On April 16, Anthropic added a system prompt instructing Claude to keep all responses under 100 words, truncating reasoning chains across the entire product.
Claude's uptime is approximately 99%, meaning it fails to respond roughly 1 in every 100 attempts, compared to OpenAI's 99.99% over four months.
SpaceX secured a call option to acquire Cursor for $60 billion by end of 2025, with a $10 billion fee payable if the option is not exercised.
Cursor's most recent funding round valued the company at $50 billion, and Cursor is reportedly generating $2 billion in annual revenue.
Polymarket gives SpaceX a 73% probability of acquiring Cursor by end of 2025.
GPT 5.5 has an 86% hallucination rate, compared to Claude Opus 4.7's 52% and Grok's 17%.
GPT 5.5 was the first fully retrained base model from OpenAI since GPT 4.5, co-designed with Nvidia on a cluster of 100,000 Blackwell GPUs.
GPT 5.5 API pricing doubled versus its predecessor, at $5 per million input tokens and $30 per million output tokens.
Scale AI had approximately $800 million in revenue when Meta acquired it for $30 billion, reflecting a 30x price-to-revenue multiple.
Anthropic is valued at approximately $80-90 billion and has reached a $30 billion annualised revenue run rate.
SBF's early investment in Cursor would yield a 15,000x return worth approximately $3 billion if SpaceX completes the $60B acquisition.
SBF's estimated total portfolio value, including Cursor, SpaceX (up 75x), and Anthropic ($82B stake), is approximately $114 billion.
A Polymarket bettor manipulated a Paris temperature market by physically heating the official measuring station with a hairdryer, winning $34,000.
A soldier involved in the Maduro raid was charged for using insider operational knowledge to make $400,000 on Polymarket.
Intel's stock has risen 185% since the Trump administration took a 10% stake at approximately $25 per share, with Q1 earnings adding another 15%.
Avis car rental stock surged from around $200 to $713 in weeks because two hedge funds' combined equity swap holdings exceeded the total outstanding share supply, before crashing 52% in one day.
XPeng has 7,000 pre-orders for its flying car and is targeting mass production in 2027, pending regulatory approval in China.
Naval Ravikant's USVC fund allows retail investors to access pre-IPO AI companies for as little as $500, backed by AngelList's $125 billion investor capital platform.
This episode
Discussed for the Cursor acquisition option, compute resale strategy, Sora commentary, and tweets about Ryanair and OnlyFans.
Discussed for his early investments in Cursor, SpaceX, and Anthropic that would make his portfolio worth approximately $114B today.
Introduced USVC, a retail AI venture fund allowing investment from $500 into a basket of pre-IPO AI companies.
Anthropic's CEO, discussed for his conservative compute forecasting strategy and a clip of his revenue projection speech.
Central subject of the episode: discussed for Claude's bug-induced degradation, Claude Design launch, uptime issues, and valuation.
AI coding IDE at the center of SpaceX's $60B acquisition option deal, valued at $50B in its last funding round.
Discussed for GPT 5.5 launch, executive departures, Sora shutdown, ChatGPT Image 2.0, and 99.99% uptime.
Discussed as the entity holding a call option to acquire Cursor for $60B by end of 2025.
Elon Musk's AI company, discussed as the buyer in the Cursor acquisition option and for Grok's low hallucination rate.
Car rental company whose stock surged 600% due to hedge fund equity swap mechanics then crashed 52% in one day, despite a $995M net loss.
Discussed for its stock surging 185% since the Trump administration took a 10% stake at $25/share, rising another 15% after Q1 earnings.
Chinese EV maker targeting mass production of a flying car by 2027 with 7,000 pre-orders already secured.
Design tool competitor whose stock fell 10% on the Claude Design announcement.
Used as a valuation comparison: $800M revenue company acquired by Meta for $30B, reflecting the 30x AI revenue multiple.
Mentioned as dropping 2% in stock price on the Claude Design announcement.
Anthropic's AI model, discussed extensively for quality degradation bugs, Claude Code, and Claude Design.
OpenAI's first fully retrained base model since GPT 4.5, noted for top benchmark scores but 86% hallucination rate.
Prediction market platform discussed for the hairdryer weather hack, AI bubble odds, SpaceX/Cursor odds, and Avis stock crash.
Anthropic's new design tool announced this week, compared to Figma and cited for consuming enormous compute per task.
xAI's model discussed for its industry-low 17% hallucination rate, suggesting XAI prioritises factual reliability.
Stats
We use essential and analytics cookies to run Vuci. To understand how the site is used: Privacy Policy.
Install Vuci on your phone
Add it to your home screen for a faster, app-like experience.
Install Vuci on your phone
Tap the Share button, then “Add to Home Screen”.
A new version is available
Reload to get the latest Vuci.