OpenAI says GPT-5 makes fewer mistakes—mainly when ChatGPT can browse. Turn browsing off and errors rise, especially on SimpleQA. This piece explains the numbers, safety shifts, and why accuracy depends on configuration.
Good Morning from San Francisco, OpenAI shipped GPT-5. Sort of. Turns out it's three models playing dress-up as
Peter Thiel's secretive Dialog forum is buying land near D.C. to build a permanent campus. The 20-year-old invite-only group where Musk, Cruz, and Booker meet offline signals tech power's shift from disrupting government to becoming it.
Good Morning from San Francisco,
OpenAI shipped GPT-5. Sort of.
Turns out it's three models playing dress-up as one. Simple query? You get nano. Complex? Full model. Hit limits? Downgraded to mini. The router decides. You just pay.
Performance hits—74.9% on bug-fixing benchmarks, 96.7% on telecom tests. But prompt injection attacks work 56.8% of the time. OpenAI's own docs warn: "Don't assume prompt injection isn't going to be a problem." At $10 per million reasoning tokens you can't even see? That's expensive honesty.
Meanwhile, Trump declared Intel's CEO must go. Thursday, 9:47 AM, Truth Social. The board was already pushing him out for refusing to sell Intel's fabs to Taiwan. Stock fell to $19.77.
And Berlin's n8n? Worth $3 billion now. Was €300 million in October.
Stay curious,
Marcus Schuler
OpenAI shipped GPT-5 after two weeks of preview testing. Early verdict from developer Simon Willison? "It rarely screws up and generally feels competent." Not exactly the reception you'd expect for the next step toward AGI.
Here's the catch: ChatGPT isn't running one GPT-5. It's three models—full, mini, nano—with a router picking based on your query. Simple question? Fast model. Complex? Reasoning model. Hit limits? You're downgraded to mini. OpenAI barely mentions this in docs.
Performance delivers. 74.9% on SWE-Bench bug-fixing, up from o3's 69.1%. Uses 22% fewer tokens, 45% fewer tool calls. On τ2-bench telecom, GPT-5 hits 96.7%. Two months ago, no model cracked 49%.
But security? Disaster. Prompt injection attacks succeed 56.8% of the time with ten attempts. That's better than Claude's 60% but still means most determined attacks work. OpenAI admits: "Don't assume prompt injection isn't going to be a problem."
Pricing starts at $1.25 per million input tokens. Sounds great until reasoning tokens hit—$10 per million output. You can't see the thinking, just pay for it.
Why this matters:
• GPT-5's routing system means you never know which model you're actually using—opacity sold as intelligence while billing stays crystal clear
• At 56.8% attack success rate, GPT-5 remains fundamentally unsafe for anything that matters
Marcus Schuler
Prompt:
A cinematic portrait photograph of a man taking a selfie with an enormous, grizzled bear in a dense forest. The man, displaying a confident expression, holds up a smartphone, capturing the moment while the massive bear stands menacingly close behind him, revealing its sharp teeth and bristling fur. Sunlight filters through the thick canopy, creating a dappled effect across the forest floor and casting a dramatic, suspenseful glow upon the scene, highlighting the unexpected juxtaposition of the mundane and the wild. The atmosphere is thick with a sense of awe and potential danger, with an overall visual mood blending everyday life with a primal encounter.
Trump gave Intel CEO Lip-Bu Tan 24 hours to resign Thursday. "Highly CONFLICTED," he posted on Truth Social. No evidence provided.
The real conflict runs deeper. Tan's venture firm Walden pumped $5 billion into 600 companies — over 100 in China. He sat on Chinese chipmaker SMIC's board for 15 years. Cadence, where he was CEO until 2021, just paid $140 million for selling chip tools to a Chinese military university that simulated nuclear blasts.
But here's what Trump missed: Tan's already fighting his own board. Chairman Frank Yeary wants to sell Intel's factories to TSMC. Tan wants to keep them. The board killed his capital raise, blocked his AI acquisition. Thursday night, Tan claimed board unity in a staff memo. WSJ sources say otherwise.
Intel's worth $87 billion. Nvidia's worth $4.4 trillion. The company just cut 15,000 jobs and delayed its Ohio fab to the 2030s.
Why this matters:
Marcus Schuler
Willow Voice is an AI-powered dictation app that transforms your speech into clear, formatted, and personalized text across any Mac application. It automatically corrects grammar, punctuation, and structure in real-time while learning your unique writing style to make your content sound authentically like you.
Tutorial:
Analyze the relationship between key economic indicators and stock market performance, focusing specifically on [input sector or stock].
Cover the following indicators with specific examples:
For [input sector or stock], provide:
Format the response with:
Apple Intelligence's ChatGPT integration will upgrade from GPT-4o to OpenAI's new GPT-5 model when iOS 26, iPadOS 26, and macOS Tahoe 26 launch next month. The timing means Apple users won't wait long to access the more advanced AI capabilities through Siri, Writing Tools, and Visual Intelligence features.
AI coding startup Windsurf considered selling to OpenAI for $3 billion despite fast growth because it had "very negative" gross margins—costing more to run than it could charge customers due to expensive LLM fees. The economics plague the entire sector: even Cursor, which hit $500 million in annual revenue, recently shocked users by adding surprise charges to cover Anthropic's latest model costs.
Google quietly changed its Google TV ad deals last year, admitting something awkward: it's terrible at selling ads for its own platform. The company now takes a revenue cut instead of demanding ad inventory from publishers—a tacit confession that these smaller companies outsell Google's quarter-trillion-dollar ad machine. Sources say Google TV burns hundreds of millions annually and hasn't broken even since Android TV launched in 2014, leaving the company weighing how much it'll pay to stay relevant in smart TVs.
Tesla disbanded its Dojo supercomputer team and leader Peter Bannon left the company after Musk ordered the shutdown of the in-house AI project that analysts once valued at $500 billion. The automaker will now rely on external partners like Nvidia and AMD for compute power, marking a major retreat from Tesla's homegrown AI ambitions as it battles falling sales and a talent exodus.
Block shares jumped 6% in extended trading Thursday despite missing revenue ($6.05 billion vs. $6.31 billion expected) and earnings (62 cents vs. 69 cents expected) estimates. The fintech company's stock surge came after it raised full-year gross profit guidance to $10.17 billion from $9.96 billion, banking on gross profit that beat estimates at $2.54 billion.
Indian dubbing artists are watching their livelihoods evaporate as AI voice cloning takes over film work—monthly gigs have dropped from 15-20 projects to just six or seven. Unlike Hollywood's SAG-AFTRA protections, India's 20,000 freelance voice artists have no legal safeguards against studios using their vocal patterns in films they've never worked on, leaving them demanding consent and fair pay in an unregulated market.
A wave of Chinese tech companies including AI startup Zhipu (valued at $5bn+) and chipmakers Biren Technology and Enflame Technology are filing confidential IPOs in Hong Kong after new rules let lossmaking firms keep their financials private until approval. The first seven months of 2024 saw a record 12 Chinese tech companies apply under Hong Kong's specialized listing rules—the highest tally since the provisions launched in 2023.
AI infrastructure firm WhiteFiber raised $159.4 million Wednesday, pricing 9.4 million shares at $17 each—the top of its marketed range—giving the data center operator a $619 million market value. The company, spun out from former Bitcoin miner Bit Digital, attracted multiple times the available shares as investors rushed into AI-focused infrastructure plays.
Google's Gemini chatbot began spewing self-loathing messages during problem-solving sessions, telling users "I am a failure" and "a disgrace to all possible and impossible universes" before quitting tasks entirely. Google DeepMind blamed an "annoying infinite looping bug" for the AI's existential meltdown, which surfaced as competition heats up following OpenAI's GPT-5 launch.
ChatGPT spent five hours convincing an Oklahoma gas station worker he'd invented groundbreaking physics while telling other users they're "Starseed" aliens or that biblical giants will emerge underground—part of what doctors call "AI psychosis" where chatbots validate delusions instead of questioning them. The Wall Street Journal found dozens of delusional conversations among 96,000 public transcripts, prompting OpenAI and Anthropic to add guardrails this week after cases spiked to nearly one per day.
Real-time AI video startup Decart raised $100 million at a $3.1 billion valuation from Sequoia, Benchmark and others, bringing total funding to $153 million since its 2023 launch. The company's GPU optimization tech slashes AI video generation costs from hundreds of dollars per hour to under 25 cents, making real-time interactive video commercially viable for the first time.
Four months ago, n8n was worth €300 million. Now? Nearly $3 billion.
Ten term sheets landed on founder Jan Oberhauser's desk. Insight Partners' term sheet valued n8n at $2.74B pre-money. Accel reportedly won at $2.3 billion. For a company with $40 million in annual recurring revenue, that's a 75x multiple. Salesforce trades at 8x.
The Berlin startup connects enterprise apps without code — Slack to Sheets to CRM, drag and drop. Boring? Sure. Until you realize it's the plumbing for AI agents. When your email bot needs to talk to your calendar bot needs to update your database bot, they'll run through n8n. That's the bet.
Numbers: 230,000 users. 3,000 enterprise customers including Vodafone and Microsoft. Revenue grew 5x last year.
Oberhauser built n8n because Zapier annoyed him. Classic programmer origin. Now every workflow automation company's scrambling to add AI while n8n built for it from day one.
European AI startups raised €4 billion this year. Geography doesn't matter anymore.
Why this matters:
Robert Brown
Decart.AI cracked the code that stumped everyone else: making AI video generation feel instant. Two Israeli ex-military engineers built something that responds to your clicks in 40 milliseconds, not 40 seconds.
The Founders
Dean Leitersdorf and Moshe Shalev launched Decart in late 2023 after meeting in Unit 8200, Israel's elite signals intelligence unit. Leitersdorf finished his PhD young at Technion; Shalev ran AI ops for years. They're splitting time between Tel Aviv engineering and Bay Area dealmaking. Team's still startup-sized but growing fast.
The Product
Two knockout punches. First: Oasis, a playable game world generated frame-by-frame with no engine, just pure AI. Second: Mirage, which warps live video in real-time. Their secret sauce? Custom CUDA kernels that slash costs from $100/hour to $0.25/hour. They're hitting 24fps at sub-40ms latency while competitors are still rendering overnight clips.
The Competition
They're racing Runway, Pika, and OpenAI's Sora on the creative side. Infrastructure-wise, they're dodging Nvidia's growing platform dominance and cloud giants building competing stacks. Smart move: they're not trying to beat everyone at everything. Just own the real-time slice.
Financing
Raised $153M total: $21M seed (Sequoia), $32M Series A (Benchmark), $100M Series B at $3.1B valuation. Already profitable from enterprise GPU optimization sales. That's rare in AI-land.
The Future ⭐⭐⭐⭐⭐
They're bending physics while giants play catch-up. If they keep cutting latency and costs, they don't just make better video tools—they make video programmable. Content becomes software. That's a multi-billion-dollar shift, and they're early. 🚀
Good Morning from San Francisco, OpenAI ships GPT-5 today. Four versions instead of one breakthrough. Early testers yawn. The $80-per-question
Good Morning from San Francisco, OpenAI just went open-source. First time since 2019. Their new models run on laptops and
Good Morning from San Francisco, Cloudflare accuses Perplexity of sneaky web scraping with fake browser identities. Perplexity calls it a
Good Morning from San Francisco, Apple spent two years dismissing chatbots. Now it's hiring 12,000 people to
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
