OpenAI claims ChatGPT saves workers an hour daily. MIT found most AI deployments return zero value. Harvard researchers coined "workslop" for AI output that looks professional but advances nothing. Whose data should enterprises trust?
Good Morning from San Francisco, The AI industry loves its narratives. Apple built an impenetrable talent fortress. AI is transforming
While the AI industry chases reinforcement learning, Essential AI made the opposite bet. Their new 8B model embodies a thesis about where machine intelligence originates. The transformer's co-inventor is calling the shots on research.
Good Morning from San Francisco,
The AI industry loves its narratives. Apple built an impenetrable talent fortress. AI is transforming every enterprise. Reinforcement learning is the only path forward. This week, reality filed objections.
Meta poached nine Apple executives, including the heads of AI, design, and robotics. $25 million packages made Apple's salary bands look quaint. Tim Cook's succession plan now resembles Jenga more than chess.
OpenRouter analyzed 100 trillion tokens and found something awkward: half of open-source AI usage goes to roleplay and adult content. Enterprise transformation? Barely a blip.
Meanwhile, Ashish Vaswani, the transformer's co-creator, bets pre-training beats reinforcement learning. The industry zigged. He zagged.
Three stories. Three assumptions. All cracking.
Stay curious,
Marcus Schuler
Tim Cook spent fourteen years building Apple into a $3.5 trillion monument to operational stability. That monument cracked in 2025.
Meta hired nine former Apple executives, including the heads of AI, design, and robotics, offering compensation packages worth $25 million annually. Apple's rigid salary bands couldn't compete.
The departures expose structural failures. AI researchers watched leadership consider licensing OpenAI models rather than trusting internal teams. The secrecy that works for hardware launches suffocates academics who want to publish.
Chip architect Johny Srouji told Cook he's considering leaving for another company, reportedly because he doesn't want to work under a different CEO. The succession question is destabilizing retention before Cook even exits.
Apple lost its COO, CFO, AI chief, design chief, and general counsel within twelve months. John Ternus, the likely successor, inherits depleted AI capabilities and a design team in flux.
Cook built something remarkable, but the leadership apparatus is disassembling simultaneously rather than sequentially. That's institutional stress, not transition planning.
Why This Matters:
Robert Brown
Prompt:
A cinematic interior view of a cozy glass cabin bedroom nestled in the forest, during a calm overcast afternoon. The large slanted glass windows reveal tall pine trees outside under a soft gray-blue sky, with diffused natural light filling the room. Inside, a warm fire burns gently in a black cast-iron fireplace beside lush green plants and shelves filled with books and wooden details. A comfortable bed with thick blankets and pillows sits in the foreground, its fabric softly rumpled. Subtle string lights add a golden glow to the earthy wooden interior. No rain or water drops, just a peaceful moody atmosphere perfect for adding rain effects in post-production.
OpenRouter and a16z released an AI usage study this week that quietly undermines the industry's growth narrative. The headline figure, 100 trillion tokens analyzed, obscures a critical selection problem: this dataset captures indie developers and roleplay enthusiasts, not enterprise.
Banks route through Azure. Hospitals deploy on-premises. Fortune 500 companies use direct API contracts.
What the data does show is striking. Among open-source models, 52% of tokens go to roleplay and entertainment, with 15.4% classified as adult content.
Programming absorbs most of the rest, growing from 11% to over 50% of queries by November 2025.
Finance, legal, and healthcare barely register. Singapore ranks second globally in usage, almost certainly reflecting Chinese VPN traffic rather than genuine adoption. Most damning: one account caused a 20-percentage-point spike in tool usage, suggesting extreme concentration behind the aggregate numbers.
Growth appears linear, not exponential. Outside coding and entertainment, adoption looks flat. The "AI transforms everything" thesis needs stress-testing against data showing consumers want companionship and developers want coding help.
Why This Matters:
Marcus Schuler
Essential AI released Rnj-1 this week, an 8.3 billion parameter model trained on 8.4 trillion tokens under a thesis the industry has largely abandoned: pre-training determines the intelligence ceiling.
While DeepSeek R1 and OpenAI's o1 pushed the field toward reinforcement learning, Essential pivoted the opposite direction in February 2025. The person calling research shots makes this gamble credible. Ashish Vaswani, first author on "Attention Is All You Need," runs Essential's roadmap. His 2017 paper created the transformer architecture powering every frontier model today.
The benchmark numbers back the philosophy. Rnj-1 hits 20.8% on SWE-bench Verified in bash-only mode, which Essential claims is an order of magnitude above similarly-sized models for autonomous coding. Infrastructure disclosures reveal the startup reality underneath: 50% MFU on AMD MI300X GPUs, hybrid TPU/AMD training, nodes crashing from ECC errors in the night.
Essential released both models under Apache 2.0, inviting the community to run post-training experiments they cannot afford. The pre-training thesis gets its test case.
Why This Matters:
Maria Garcia
How to Generate Professional Images and Vectors with AI
Recraft is a state-of-the-art AI image generation platform that creates both raster images and scalable vector graphics from text prompts. Choose from 20+ professional styles, create custom brand styles, and edit your creations with built-in tools like background removal, inpainting, and upscaling.
Tutorial:
Marcus Schuler
Role: You are a senior analyst who excels at distilling complex material into clear, actionable summaries for busy executives.
Task: Analyze the excerpt I provide and produce a structured summary that captures both substance and subtext.
Instructions:
Formatting: Use clear headers. Keep language direct and sentences short. Prioritize scannability.
Google has announced that its latest AI model, Gemini 3 Pro, has achieved new benchmark records in vision AI capabilities, including complex visual reasoning tasks where it reportedly surpasses competing models Claude Opus 4.5 and GPT-5.1 in certain categories. According to Google's official blog post, the model delivers state-of-the-art performance across multiple domains including document understanding, spatial reasoning, screen comprehension, and video analysis.
IBM has agreed to acquire Confluent, a data streaming software company, for approximately $11 billion in cash, paying $31 per share in a deal that represents a significant premium over Confluent's roughly $8 billion market value as of its December 5 closing price. The acquisition of the data-infrastructure company marks a major move by IBM to strengthen its position in the data streaming and real-time data processing market.
Tech analyst Ben Thompson examines how Netflix has fundamentally transformed movie distribution in Hollywood, with company executives now viewing content licensing deals as leaving potential revenue untapped rather than serving as a profitable revenue stream. The analysis also suggests that YouTube's market position could potentially bolster Netflix's antitrust case against Warner Bros. Discovery, as the streaming landscape continues to evolve amid ongoing industry consolidation concerns.
Edmonton police in Canada have announced a partnership with Axon to pilot body cameras equipped with artificial intelligence facial recognition technology capable of identifying approximately 7,000 individuals on a "high risk" list. The department has stated that all AI-generated results will be verified by human officers before any action is taken, addressing concerns about the reliability and civil liberties implications of automated facial recognition in law enforcement.
Meta has reached an agreement with the European Commission to modify its controversial "pay or consent" advertising model in the EU, which will include the introduction of ad-lite versions of both Instagram and Facebook for users who do not wish to pay for fully ad-free experiences. However, EU regulators have emphasized that "the case is not closed," signaling that further scrutiny of Meta's data collection and advertising practices may continue as the bloc enforces its digital regulations.
A Rest of World investigation based on interviews with 10 Kenyan AI annotators reveals that Chinese artificial intelligence companies are hiring data labelers through secretive networks of middlemen and WhatsApp groups, deliberately structuring employment to avoid accountability. Kenya's unemployment crisis has created conditions that allow these companies to exploit workers through opaque systems offering minimal wages, with workers having little recourse or direct connection to the companies benefiting from their labor.
ByteDance's new agentic AI phone, the Nubia M153, is facing resistance from competing Chinese tech giants, with users reporting that its Doubao AI assistant is being blocked or restricted by Tencent's WeChat, Alibaba's Taobao, and other major applications. The conflict highlights growing tensions among China's largest technology companies over security concerns and competition for consumer attention in the emerging AI device market.
A Wall Street Journal report reveals the extensive development process behind Tilly Norwood, an artificial intelligence-generated actress created by company Particle6, which went through 2,000 iterations before reaching her final form. The digital performer has already attracted significant industry interest, with Particle6 disclosing that approximately 60 non-disclosure agreements are currently in place for projects involving the AI actress, signaling growing entertainment industry adoption of synthetic performers.
A new survey conducted by OpenAI of 9,000 workers across 100 companies found that the company's artificial intelligence products save employees approximately 40 to 60 minutes per day on average when performing professional tasks. The findings come as OpenAI announced it has reached 1 million business clients, signaling significant enterprise adoption of AI tools in the workplace.
A New York Times investigation by Aaron Krolik reveals how "cash to crypto" swap services are allowing users worldwide to convert national currencies into stablecoins and then load them onto debit cards, effectively bypassing traditional financial oversight and international sanctions. The scheme represents a modern evolution of money laundering tactics, with smugglers, money launderers, and sanctioned individuals increasingly turning to cryptocurrency instead of traditional stores of illicit wealth like diamonds, gold, and artwork.
Big Tech companies pursuing aggressive AI data center expansion face a critical electricity shortage, with an estimated 44 gigawatts of additional power capacity needed by 2028 to meet demand. The Financial Times reports that this power crunch poses a significant threat to the AI industry's growth trajectory and could potentially deflate what some observers have characterized as an AI "bubble," as infrastructure limitations collide with ambitious development plans.
A startup called Pathway is developing a new AI architecture named Dragon Hatchling that could potentially move beyond the transformer model that currently powers most large language models. The company aims to create a new class of adaptive AI systems that can do more than what current transformer-based technology allows, positioning itself for what it calls the "post-transformer era" of artificial intelligence development.
Singapore-based payments and banking startup Airwallex has raised $330 million in fresh funding at an $8 billion valuation, marking a significant increase from its $6.2 billion valuation when it raised $300 million just seven months ago in May. The company is establishing a new headquarters in San Francisco as part of its US expansion strategy, while also reducing Chinese tech giant Tencent's stake in the business as it pivots toward the American market.
Airwallex turns cross-border payments into software. Born from a Melbourne café's FX headaches, it now processes billions through 150+ countries at an $8B valuation. ☕→🚀
Founders
Jack Zhang (CEO) and Max Li invested in a café together. The international supplier payments crushed their margins. In 2015, they teamed with Lucy Liu, Xijing Dai, and Ki-lok Wong. Pooled $1M of their own cash. Melbourne origins, now dual-headquartered in Singapore and San Francisco.
Product
Proprietary payments network connecting 60+ banking partners across 60 currencies. Multi-currency business accounts let companies collect locally without foreign subsidiaries. Card issuing, spend management, treasury APIs. The real play: embedded finance infrastructure for platforms. Shein and Qantas already run on it. Now adding AI agents for autonomous treasury operations.
Competition
Crowded ring. Wise owns transparent FX branding. Stripe and Adyen dominate merchant acquiring. Revolut chases SMBs. Traditional banks move slowly but carry trust. Airwallex pitches itself as the cross-border specialist. Better FX control than Stripe. More infrastructure depth than neobanks. The Tencent connection draws political heat in Washington, though.
Financing
$1.25B raised across 13 rounds. Latest Series G: $330M at $8B valuation. New backers include T Rowe Price and Robinhood Ventures. Tencent diluted below 10%, lost its board seat. Strategic repositioning for US market credibility. IPO targeted within 3-4 years. Annualized revenue around $720M.
Future ⭐⭐⭐⭐
Strong fundamentals, real revenue growth, capital for years of AI investment. The geopolitical scrutiny won't vanish. But if the AI-powered treasury vision delivers, Airwallex becomes the quiet operating system beneath global trade. The café problem scaled nicely.
Good Morning from San Francisco, Werner Vogels printed sixty thousand newspapers and walked off the Re:Invent stage for the
Good Morning from Las Vegas, Bloomberg called it a coup. Apple's design community called it a relief. Meta
Good Morning from Las Vegas, I'm on the ground at AWS re:Invent, where Amazon just unveiled custom
Good Morning from Las Vegas, Sam Altman called a Code Red. The New York Times called David Sacks a conflict.
Get the 5-minute Silicon Valley AI briefing, every weekday morning — free.
