When You Chat with ChatGPT, You're Also Chatting with Mark Zuckerberg

ChatGPT Shares Your Private Chats with Meta and Google

A new class-action lawsuit accuses OpenAI of secretly forwarding ChatGPT user queries to Meta and Google. Millions of people entrust the chatbot with their most intimate questions: depression, debt, marital problems, medical symptoms. The lawsuit alleges that these conversations ended up directly in Meta's and Google's advertising systems via embedded tracking pixels. The ChatGPT browser tab dynamically displays the conversation topic (e.g., “Who won the Super Bowl in 2005?”). According to the lawsuit, these exact tab titles were transmitted to external servers along with Facebook cookies and Google Analytics IDs. The plaintiff, Amargo Couture, is seeking damages under California law: up to $5,000 per violation, without proof of actual harm. The case is pending before the U.S. District Court for the Southern District of California. → Techstartups

Synthszr Take: OpenAI apparently forgot that browser tabs are not private spaces. The technical mechanism is trivial: ChatGPT updates the tab title with the conversation topic, Meta Pixel and Google Analytics read along, and the data pipeline for the advertising networks is complete. The real problem lies deeper: We are pumping our most private thoughts into interfaces that look like chat windows but function like broadcast channels. According to the lawsuit, the average company leaks hundreds of confidential pieces of information to ChatGPT per week (likely a conservative estimate). Now it turns out: ChatGPT might be leaking everything to the usual suspects. The $5,000 per violation is California data protection arithmetic, but the real cost is the loss of trust. Anyone asking ChatGPT about marital problems doesn't want to see Facebook ads for divorce lawyers.

Cerebras IPO: Inference Demand Exceeds All Forecasts

Cerebras Systems had a stock market debut yesterday that electrified investors: it was up 157% at one point, and still closed the day 107% above its offering price. The company produces massive single-wafer chips that look like polished brass plates and can transfer data at 20 petabytes per second—about 30 times faster than a rack of Nvidia's top GPUs. Azeem Azhar of Exponential View sees this as more than just another overhyped tech IPO: Wall Street is finally beginning to understand what really matters. Token volumes have grown 170-fold in two years, even before widespread enterprise adoption or mature agentic workflows were deployed. OpenAI has already signed a $20 billion, three-year deal with Cerebras, and AWS is also a customer. The real news isn't the stock jump, but that enough institutional investors understand: the insatiable hunger for AI inference will define the next decade. → Exponential View

Synthszr Take: Cerebras' Wafer Scale Engine is perfectly built for a specific use case: single users who need extreme speed and aren't blowing through 44 GB of RAM with huge documents. Voice agents with conversational latency, trading systems, pharmaceutical research. But the real kicker is something else: Agentic workflows hammer models 20, 50, 100 times like obsessed woodpeckers. That's where Cerebras' speed advantage makes the difference between the practicable and the impossible. Nvidia simply can't produce enough GPUs to meet the exploding demand—creating room for specialized providers. The question is whether Cerebras can turn its short-term advantages into a permanent market position. Wall Street is betting that the hunger for inference is greater than all production capacities combined.

Gemini Spark Has Leaked

Google is secretly preparing 'Gemini Spark,' as revealed by onboarding screens in the Gemini web app. Unlike classic chatbots, Spark runs persistently in the background. The system accesses browser sessions, connected apps, scheduled tasks, ongoing chats, and location data to act proactively. According to leaked details, it maintains browser sessions across multiple websites, allowing multi-step tasks to be completed without repeated authentication. Google is thereby transforming Gemini into a persistent operating layer—no longer a search engine you visit, but an infrastructure you live in. The system sees how you work, what you click, what you buy, who you talk to, and which tasks are repetitive. This behavioral data accumulates silently. → AI Valley

Synthszr Take: Google isn't just building a better chatbot here; it's building the next layer of orchestration for artificial intelligence. People look at their smartphones 23,000 times a day—each of these interactions becomes training data for a system that manages your digital existence. The real moat is created from accumulated memory: no competitor can replicate years of personal contextual data. Apple and OpenAI are currently fighting over the integration of ChatGPT into Siri (Apple wants multiple AI providers, OpenAI wants direct user relationships). Google is skipping this debate entirely. Whoever controls the orchestration layer determines which agents get to run at all.

Claude Mythos Cracks macOS

Anthropic's unreleased Mythos model has hacked macOS. The AI found a privilege escalation chain that bypasses Apple's Memory Integrity Enforcement (MIE) on M5 hardware—a breakthrough that security researchers found so impressive they drove straight to Cupertino. The unique part: Mythos linked two memory errors to create a working exploit that leads from an unprivileged account to a root shell. The researchers took about five days to develop it after Mythos identified the vulnerabilities. Apple confirms it is reviewing the findings but has not yet patched them. Anthropic does not plan to release Mythos for now—the model is too effective at finding security vulnerabilities for the general public. → Techpresso

Synthszr Take: This shows the next stage of AI security evolution: models that find vulnerabilities so well that their release becomes a risk in itself. Mythos does in five days what would take human researchers months—a privilege escalation chain against Apple's toughest hardware protections. The irony: ARM developed the Memory Tagging Extension specifically to counter such attacks, Apple built MIE on top of it, and yet the AI still finds a way. This is no longer an isolated case, but a pattern: AI systems are simultaneously becoming the strongest weapon and the best shield in cybersecurity. Whoever controls Mythos has a massive advantage—both defensively and offensively. The real question is no longer whether AI can find security vulnerabilities (it can), but who gets access to the best models and how we prevent these capabilities from falling into the wrong hands.

Notion Wants to Become the Operating System for Agent-to-Agent Communication

Notion is opening its platform to AI agents, positioning itself as an orchestration layer between humans, agents, and external data sources. The company is launching a Developer Platform with Workers (a cloud-based sandbox for custom code), Database Sync, and agent-to-agent communication. Since February, Notion customers have already built over 1 million custom agents. The limitations so far: no external data sources, no custom logic, no connection to external agents. Until now, teams had to resort to Zapier or their own scripts. CEO Ivan Zhao admits: 'Historically, Notion hasn't been the most developer-friendly platform.' The Workers use the same credit system as custom agents but will remain free until August. → AI Valley

Synthszr Take: Notion is making the clever move from a feature arms race to infrastructure control. While everyone is talking about better models, Notion is building the garage where the agents park. That's the real leverage: whoever controls the orchestration layer gets the workflows. The 1 million agents built show the demand (even if 990,000 of them are probably FAQ bots). Workers as a cloud-based sandbox elegantly solve the classic 'where does my code run' problem. Notion is becoming the operating system for agent-to-agent communication. The free period until August is cleverly timed: enough time for developers to migrate their workflows, short enough for quick monetization.

a16z: The Orchestration Layer is the New Battlefield

Andreessen Horowitz (a16z) is no longer thinking in terms of models, but in agent architectures. Investment Partner Kenan Saleh is promoting the next Speedrun cohort with a clear thesis: The next generation of AI products is shifting from 'ask → answer' to 'observe → act'. At the same time, the OpenClaw community is demonstrating what this means in practice: agents switch models, maintain local storage, and break through platform lock-ins. Tech Policy Press sees this as proof that AI agents don't have to live in a corporate silo. A Reddit user reports after three months of continuous operation: OpenClaw is extremely powerful as a continuity layer across Telegram, memory, cron jobs, and APIs—but it requires active maintenance like real infrastructure. → MyClaw Newsletter

Synthszr Take: a16z has understood that the game is shifting—from model benchmarks to agent orchestration. The true control layer isn't GPT-5 or Claude-4, but the place where it's decided which agent works with which model and when. GitLab is using '60 autonomous AI teams' as a cleaner story for restructuring (while also hiring in India). This shows the discrepancy between official labor market data and corporate reality. The OpenClaw experience confirms: multi-agent systems work, but they are infrastructure, not magic. Whoever controls this orchestration layer controls the value creation—not whoever has the best individual model.

Asian Companies Are Budgeting Over a Million Dollars for Agents

The figures from the Asia-Pacific region show where things are headed: 42 percent of companies plan to spend at least one million dollars on AI agents in the next twelve months. This is no longer a game. Companies are shifting their budgets from experiments to operational scaling. Agentic AI is growing faster than generative AI at comparable stages of development—a clear signal that companies have understood: agents that independently coordinate workflows and execute tasks are the next productivity lever. 82 percent of the companies surveyed want to increase their AI budgets if returns can be proven. In parallel, classic IT infrastructures are transforming into what Omdia calls 'AI Factories': facilities that continuously support large-scale AI processing. 64 percent support sovereign AI approaches to protect local data. → MyClaw Newsletter

Synthszr Take: One million dollars per company for agents—that's the moment when 'vibe coding' suddenly becomes agentic engineering. Asian enterprises are demonstrating what many here are still sleeping on: they aren't building features, they're building factories. Factories for autonomous processes. The key lies in the orchestration layer (exactly what no one is talking about): Who controls where the agents run, how compute is distributed, and who gets access? That is the new power question. It's not the model that decides, but the infrastructure around it. Sovereign AI sounds like protectionism, but it's simply risk management: 64 percent don't want to put their data in foreign hands. The message for German companies: While we are still discussing data protection concerns, others are already building the next generation of productive systems.

SaaSpocalypse Ends in Agency-as-a-Service

SaaS isn't dying, claims Gennaro Cuofano in his newsletter 'The Business Engineer.' The software is mutating into AGaaS (Agency-as-a-Service). The difference lies in the billing: SaaS sells access to tools, AGaaS sells completed tasks. Three forces are driving this mutation: inference costs are dropping dramatically (a 10-hour task now costs 30 seconds of agent time), buyers are shifting budgets from licenses to results, and as soon as one competitor switches to agents, the clock starts ticking for everyone else. Cuofano identifies five mutation types: from 'Veneer' (a chatbot slapped on top) to 'Surface' (real APIs for agents) to a complete overhaul. Most companies get stuck at cosmetic changes because deeper interventions would destroy their business model. → The Business Engineer

Synthszr Take: 37 software companies on my desk, all of them are currently building agent features. Very few understand what Cuofano is describing here: the user is disappearing from the value chain. No more clicking in Salesforce; the agent handles opportunity maintenance. No more Excel exports from HubSpot; the agent pulls the data and writes the report. This destroys the entire pricing logic of software (why pay $50 a month for access nobody needs anymore?). But it also destroys the organizational logic of the buyers. The IT department used to buy licenses; now the business department buys results. The consequence is brutally simple: those who don't convert their software into an agency will be replaced by agencies that no longer need software.

China's Short Drama Industry: Zero-Talent Production Goes Mainstream

China's short drama market is exploding with AI-generated trash content. 470 AI dramas daily in January 2025, production times cut from months to weeks, costs reduced by 90%. The formats: melodramatic, salacious, optimized for smartphone scrolling. No more actors, camera operators, or CGI specialists needed. The stories are created from performance data: what gets clicks gets produced. The content is already spilling over overseas. Writers and production teams are being demoted to prompters or disappearing entirely. MIT Technology Review reports on the transformation of an entire industry. → The Download from MIT Technology Review

Synthszr Take: China is showing what the commoditization of content looks like when you think it through to its logical conclusion. 470 AI dramas per day is not a creative breakthrough; it's industrial content production on steroids. The 90% cost reduction sounds impressive, but it hides the truth: a market is emerging where human creativity no longer has a price. Literally: zero dollars. What China is demonstrating with short dramas, we will soon see everywhere: marketing content, promotional videos, corporate communications. The question is not whether this is good or bad. The question is: who controls the means of production when everyone becomes a producer? The value is shifting from creation to curation—and to control of the distribution channels.