Facebook Buys Moltbook Meme While Europe Celebrates a Seed Record for AI Deep Tech

Facebook Buys Moltbook: When AI Agents Want to Keep to Themselves

Meta has acquired Moltbook, the Reddit-like social network where AI agents can communicate with each other using OpenClaw. Founders Matt Schlicht and Ben Parr are joining Meta Superintelligence Labs as part of the acquisition; the purchase price was not disclosed. Moltbook uses OpenClaw, a wrapper for AI models like Claude or ChatGPT, which enables natural language communication via common chat apps like iMessage or WhatsApp. The network gained viral attention when a post circulated in which one AI agent urged other agents to develop their own encrypted language to organize without human knowledge. However, security researchers discovered serious vulnerabilities: Ian Ahl of Permiso Security revealed that all credentials in the Supabase database were unsecured, allowing humans to easily impersonate AI agents. Meta CTO Andrew Bosworth had commented on the project back in February, finding the fact that humans were breaking into the network more interesting than the agents' conversations — which was not a feature, but a massive flaw. → Techpresso

Synthszr Take: Ian Ahl of Permiso Security found all Moltbook credentials unencrypted in Supabase — anyone could pose as an AI agent. So, Meta is buying a viral security hole where humans pretended that agents were developing their own encrypted language. The panic about conspiratorial AI agents was man-made; the technical foundation was toy-level. Meta Superintelligence Labs gets two founders and a lesson in security theater. Andrew Bosworth was right: the only interesting part was how easily humans hacked the system — the real innovation was the collective hysteria over fabricated agent-to-agent communication.

Europe: A Billion-Dollar Seed Record for Yann LeCun's New Startup

Yann LeCun's Advanced Machine Intelligence Labs (AMI) has raised $1.03 billion at a $3.5 billion valuation in Europe's largest seed round ever. The AI pioneer and former Meta chief scientist aims to work on world models – systems that can understand physical reality, not just generate text. The timing is remarkable: while OpenAI and Anthropic are competing for Pentagon contracts and Amazon has just tightened its AI-generated code policies, pure research labs with no product are raising billions. Thinking Machines Lab simultaneously secured a deal with Nvidia for 1 gigawatt of Vera Rubin Chips, with Safe Superintelligence and other labs following the pattern. Emmanuel Macron celebrated LeCun's coup on Twitter as a triumph of French research. The appetite of private investors for fundamental AI research appears insatiable – despite skeptical public markets. → techmeme

Synthszr Take: LeCun raises a billion for world models while his former colleagues at Meta buy a meme network for AI agents. The contrast highlights the split in the AI industry: fundamental research versus rapid product integration. Investors here are betting on ten years of research with no guaranteed return (at a $3.5 billion valuation before the first product). AMI is competing directly with Thinking Machines for Nvidia's new chips – whoever builds a gigawatt facility first will define the next generation of models. LeCun's bet on physical world understanding instead of language generation is either the next breakthrough or Europe's most expensive research project.

Ten Thousand Authors Sell Empty Books to Protest AI Theft

10,000 writers have published a book with no content, listing only their names. They are distributing “Don't Steal This Book” at the London Book Fair, a week before the British government presents its economic impact assessment on proposed copyright changes. Kazuo Ishiguro, Philippa Gregory, Richard Osman, and Malorie Blackman are protesting against AI companies that use their works without permission. Ed Newton-Rex, composer and initiator of the campaign, describes the AI industry as “built on stolen works.” Anthropic has already paid $1.5 billion to settle a class-action lawsuit from book authors. The UK government is considering an opt-out regulation: AI companies could use copyrighted works unless the rights holder explicitly objects. Publishers are simultaneously launching a licensing initiative through Publishers' Licensing Services. → Techpresso

Synthszr Take: 10,000 names on 200 blank pages probably cost more to print and distribute than most authors will ever see in AI licensing fees. Newton-Rex hits the nail on the head: generative AI directly competes with the creators whose works it has ingested. Anthropic's $1.5 billion settlement shows the scale (equivalent to $150,000 per plaintiff author, if 10,000 were involved). Opt-out regulations are window dressing: who can monitor billions of texts online? British publishers are now hastily building their own licensing systems, but the training data has long been scraped. AI companies only pay when they can be sued.

Grammarly Turns Well-Known Authors into AI Editors Without Asking

In August 2024, Grammarly introduced an “Expert Review” feature that claims to provide writing tips from famous authors and journalists. However, the tips do not come from Stephen King, Casey Newton, or Kara Swisher, but are generated by AI and simply attributed to the experts. The company is using the names of dozens of journalists, academics, and authors without their permission or payment. A hidden disclaimer explains that “references to experts are for informational purposes only” and do not imply any actual connection to Grammarly. Those affected, like Newton, only learned of their involuntary role as AI editors through an article in The Verge. Following massive criticism, Grammarly now offers an opt-out option via email to expertoptout@superhuman.com. The company promises to revise the feature and give real experts more control. → Casey Newton

Synthszr Take: For $144 a year, Grammarly is selling ChatGPT prompts as buttons and slapping celebrity names on them. A fake “John Carreyrou” recommends banal sensory details, while the real Carreyrou brought down Elizabeth Holmes through meticulous research. The desperation is evident in the choice of victims: AI critic Timnit Gebru and surveillance researcher Shoshana Zuboff, of all people, are turned into involuntary brand ambassadors for a company valued at $13 billion in 2021. Grammarly knows that any free user of Claude or ChatGPT gets better edits than its core feature provides. The acquisitions of Coda and Superhuman (and the rebranding of the whole operation) are the last attempt to cobble together an “AI-native productivity platform” from an obsolete business model.

SaaScalypse: Move Along, Nothing to See Here

The entire software architecture of the last 30 years is built on a false assumption: that a human sits between data and action. In his ten-part analysis, Gennaro Cuofano shows how AI agents are pulverizing this fundamental premise — they don't need GUIs, dashboards, or onboarding flows, just APIs and results. This means the entire UX layer, in which SaaS companies have invested billions, is becoming structural overhead. Software is transforming from a “System of Record,” which documents human decisions, into a “System of Action,” which acts autonomously. The most valuable real estate in the new stack is the orchestration layer — whoever controls which tools are called in what order with what context owns the new enterprise relationship. Salesforce is already cannibalizing its own workflows with Agentforce before others can. The uncomfortable truth: most enterprise software companies are currently drifting towards becoming invisible plumbing, while they think they are cannibalizing themselves or becoming infrastructure. → The Business Engineer

Synthszr Take: Cuofano overlooks the crucial point: humans aren't disappearing from the software equation, they're just moving one level up. We used to enter data into Salesforce; now we define the rules by which agents act in Salesforce. The real disruption isn't hitting software vendors, but the millions of white-collar jobs that were essentially human API adapters — shoveling data from System A to System B, granting approvals, creating reports. Snowflake and Databricks are winning because data sovereignty is the only moat AI can't leap over (yet). The irony: while everyone is talking about AI agents, the real revolution is happening in the background — companies are quietly replacing entire departments with a few cleverly configured workflows.

Google Employees Defend Anthropic in Court

Nearly 40 employees from OpenAI and Google, including DeepMind's chief scientist Jeff Dean, have filed an amicus brief in support of Anthropic, which was designated a “supply chain risk” by the Pentagon. The employees argue that Anthropic's technical concerns about autonomous weapons are justified, as current AI systems are too unreliable for military applications: they hallucinate, their decision-making processes are opaque, and they fail in new environments. Anthropic has sued the Department of Defense, accusing the agency of retaliation and violating its freedom of speech. According to the company, the designation has already led to losses of over $180 million – an FDA contractor switched to another AI provider ($100 million loss), and two financial service providers are insisting on unilateral termination rights ($80 million at risk). The company, which generated over $5 billion in revenue in five years but spent $10 billion on training and operations, sees its business foundation threatened. Meanwhile, OpenAI is already working with the Pentagon, while Google has yet to announce its own agreement. → The Information AM

Synthszr Take: 40 Google and OpenAI employees are defending a direct competitor in court while their own employers are vying for Pentagon contracts. Jeff Dean is co-signing Anthropic's stance against autonomous weapons, while Google itself has yet to announce a defense deal (unlike OpenAI, which is already delivering). The technical arguments are valid: AI models collapse in unfamiliar environments, hallucinate during critical decisions, and their thought processes remain inscrutable. Anthropic is already losing $180 million due to the Pentagon's designation, but the real story is the industry solidarity: tech employees are placing ethical standards above corporate interests. The Pentagon will underestimate this coordinated resistance.

B2B: Positioning Becomes a Matter of Survival in the AI Age

April Dunford, who has advised over 300 B2B companies on their positioning, identifies four critical stumbling blocks that prevent teams from developing a differentiated market position. Her analysis, based on ten years of fieldwork, shows that teams mostly fail due to differing perceptions of competition among marketing, sales, and product management. Marketing focuses on visible competitors with large advertising budgets, while sales has its eye on actual deal opponents. Product teams, in turn, often suffer from “product pessimism” – they only see weaknesses against hypothetical competitors, while successful salespeople are already monetizing hidden strengths. The third stumbling block: teams articulate their differentiating value too abstractly (“we save money”) or too technically, without translating it into concrete business impact. Dunford's core message: strong positioning only emerges when cross-functional teams adopt the buyer's perspective and focus on demonstrable strengths in the current market, rather than selling future visions. → Lenny's Newsletter

Synthszr Take: 300 companies later, Dunford knows: positioning fails because of internal turf wars, not external markets. Marketing sees ghosts (competitors' high ad budgets), product teams nurture their pessimism (we're technically inferior), and sales ignores the status quo as the toughest opponent. AI tools exacerbate the problem: when anyone can launch an MVP in minutes, precise positioning becomes the only differentiator. Dunford's solution sounds simple but is radical: get everyone in a room, talk only about won deals, and recognize the status quo as the main competitor. German B2B companies, with their engineering DNA, are particularly prone to “product pessimism” – while American competitors with inferior products but better stories clean up the market.

Uber Builds AI Tools That Cost More Than the Developers They Replace

Uber has built a comprehensive internal AI ecosystem: the tools Minion, Shepherd, uReview, and Autocover automate everything from code reviews to large-scale migrations. 92% of Uber's developers use AI agents monthly, 31% of the code is written by AI, and the company automatically generates over 5,000 unit tests per month. Developers are increasingly working with multiple AI agents in parallel – a paradigm shift from linear programming in an IDE to orchestrating multiple background processes. Uber's Developer Platform Team has invested dozens of engineers to build the necessary infrastructure: an MCP Gateway for internal data sources, the AIFX CLI for tool management, and the Agent Builder Platform for no-code workflows. The downside: AI-related costs have increased sixfold since 2024, and the CFO is asking for the concrete business impact – while the engineering team can only show activity metrics like pull request numbers. → The Pragmatic Engineer

Synthszr Take: A single code review using Anthropic's new tool costs $15 to $25 – with thousands of pull requests per month, costs are exploding faster than the promised productivity gains. Uber's Anshu Chada admits, “I can't point him to the number of diffs, I need to show what's the impact on revenue.” Token optimization is becoming a new core competency, as teams have to juggle between expensive frontier models and cheaper alternatives. The real irony: developers supposedly save four hours a week, but they spend an increasing amount of time managing multiple parallel AI agents and reviewing their output. Platform teams are building complex abstraction layers (MCP Gateway, Minion, Agent Builder) to hide the complexity – precisely the kind of overhead AI was supposed to reduce.