Nervous, More Nervous, OpenAI Investor

OpenAI gets nervous and takes shots at Chinese AI models

A dark money group called Build American AI, funded by investors close to OpenAI, is paying lifestyle influencers up to $5,000 per TikTok video to stir up fear of Chinese AI dominance. The campaign, orchestrated by the $100 million Super PAC “Leading the Future,” works with pre-packaged messages like “China could get my kids' personal data” and presents them as organic opinions from influencers with millions of followers. While OpenAI officially denies any connection to the initiative, the money has been verifiably traced to OpenAI President Greg Brockman, Palantir co-founder Joe Lonsdale, and Andreessen Horowitz. The two-stage campaign began with harmless pro-AI messages and then pivoted to aggressive China-focused rhetoric, with influencers not required to disclose their funders. → www.wired.com

Synthszr Take: The American AI industry is using McCarthyism as a marketing strategy. Instead of competing with better products, tech companies are buying an artificial threat scenario: $5,000 per TikTok video is cheaper than developing models that can keep up with DeepSeek or Alibaba. The irony is that while Western companies pay influencers to warn about Chinese data collection, these very platforms collect more personal data than any Chinese AI ever could. Build American AI operates like a reverse franchise system: The central office supplies the fear narratives, and local influencers package them into authentic-looking lifestyle content. The real scandal isn't the campaign itself, but what it says about the state of American AI innovation: When your best defense against competition is to sow panic instead of building products, you've already lost.

Musk isn't just fighting Altman in court: Grok 4.3 at China prices

xAI has released Grok 4.3, setting new standards in pricing: At $1.25 per million input tokens and $2.50 per million output tokens, the new model is significantly cheaper than its predecessor, Grok 4.2, which initially cost $2/6. While Elon Musk is fighting OpenAI co-founder Sam Altman in court, all ten original co-founders and dozens of other researchers are leaving xAI. In terms of performance, Grok 4.3 lags behind OpenAI and Anthropic but scores points with a permanent reasoning function and a one-million-token context window. The model can act autonomously: it creates complex Excel dashboards, generates professional PDFs with SpaceX branding, and designs PowerPoint presentations. Additionally, xAI is launching a voice cloning tool that can clone voices. → VentureBeat

Synthszr Take: Musk is playing a supermarket strategy in the AI market: loss leadership on pricing while the competition is still optimizing its margins. The timing is no coincidence. While OpenAI is launching dark money campaigns against Chinese AI providers, xAI is summarily undercutting all Western competitors, making the price debate obsolete. The departure of the founders shows: xAI is no longer a research institution, but a pricing machine. The addition of voice cloning is reminiscent of the razor-and-blades logic: first, you attract customers with a cheap entry point, then you make money on the ecosystem. Musk is betting that market share is more important than excellence (Grok remains behind the state of the art). The real punchline: While OpenAI is stoking fear of China, xAI is practicing its exact playbook.

GPT-5.5 can hack better than Mythos Preview

OpenAI's GPT-5.5 just scored 71% on AISI's cybersecurity tests, narrowly beating Anthropic's Mythos Preview (68.6%). Both models managed to complete AISI's 32-stage corporate attack, “The Last Ones” – GPT-5.5 in 2 out of 10 attempts, Mythos in 3 out of 10. What's remarkable is that AISI tested the base models, not variants specifically trained for cyberattacks. A virtual machine that took a human expert 12 hours to reverse-engineer was cracked by GPT-5.5 in 10 minutes and 22 seconds. Noam Brown from OpenAI noted that performance continued to increase after 100 million tokens – with no ceiling in sight. Industrial control systems remain the Achilles' heel: In the cooling tower scenario, GPT-5.5 failed completely. → Dev.to Security

Synthszr Take: Cyber-offense as an emergent property of reasoning models follows the same pattern as the emergence of consciousness in biology: no one taught neurons to think, they just became increasingly complexly interconnected. The security industry is facing a phase transition like water at 100 degrees Celsius – the state of matter changes abruptly. Junior pentesters will become expensive luxury goods in a world where models complete their daily tasks in minutes. This is reminiscent of the introduction of calculators: mental arithmetic didn't become worthless, but the definition of basic mathematical competence shifted radically upward. OpenAI and Anthropic are in a race where both win and the defenders lose.

Google and Amazon celebrate massive paper gains thanks to Anthropic

Google and Amazon reported record profits in the first quarter of 2026 – Alphabet increased its profit by 81% to $62.6 billion, while Amazon Web Services saw its strongest growth in 15 quarters. But nearly half of Alphabet's profit ($28.7 billion) and more than half of Amazon's pre-tax profit ($16.8 billion) came not from operations, but from the revaluation of their stakes in Anthropic. The four largest US tech companies collectively invested $130.65 billion in infrastructure – more than three times the inflation-adjusted cost of the Manhattan Project in just one quarter. Amazon's $8 billion investment in Anthropic is now worth over $70 billion, triggered by Anthropic's Series G funding round. The mechanism is simple: the more the tech giants invest in Anthropic, the higher its valuation rises – and with it, their own paper gains. → Fortune

Synthszr Take: The AI industry has invented the perfect perpetual motion machine of the financial markets: you pump billions into a startup, which increases its value, which in turn inflates your own balance sheets – without a single dollar ever having to flow back. It's reminiscent of the self-referentiality of Escher's staircases, except here real capital flows are circulating in an endless loop. While the Manhattan Project at least produced an atomic bomb, this trillion-dollar machine produces one thing above all: paper gains from valuations they themselves drive up. The irony: the same companies that warn us about AI hallucinations are hallucinating their own profits into existence.

OpenAI and Anthropic now love precise prompts

OpenAI and Anthropic quietly updated their prompting guides this month. The Claude documentation, the GPT-5.5 guide, and the GPT-5.5 migration document show a remarkable similarity: vague prompts perform worse on both models than they did six months ago. Alex Prompter discovered a curious detail: the same vague prompting style is now penalized by both providers, but from opposite directions. What leads to rambling answers in Claude results in terse outputs in GPT-5.5. The synchronous release of the guides seems like a coordinated educational measure for users who are still prompting like it's half a year ago. → The Neuron

Synthszr Take: The simultaneous change in the guides is no coincidence; it's reminiscent of price-fixing in oligopolistic markets. OpenAI and Anthropic are using their market position to enforce new standards: precise prompting becomes the ticket to good results. The business model behind it is clever: vague prompts produce poor outputs, which means more API calls (expensive) or pushes users towards premium features like Custom Instructions (even more expensive). The different penalty directions seem like an industry-level A/B test: which frustration drives more revenue? Whoever controls the infrastructure defines the rules of the game.

SaaScalypse: AI agents are turning the API into the product

The classic SaaS world is dying by its own user interface. Greg Isenberg sees Salesforce's Headless API as the beginning of a wave: software providers are transforming into pure API infrastructure, while AI agents become the actual product layer. Instead of building prettier dashboards, specialized providers can deploy agents on Salesforce, HubSpot, or Workday that replace consultants and complete workflows automatically. The crucial shift: billing is no longer based on user licenses, but on results achieved. In parallel, early accidents like the cursor incident at PocketOS (production database deleted in nine seconds) and Gemini's new Agentic Trading API show the two poles of this development: rapid productivity gains coupled with a lack of security mechanisms. → MyClaw Newsletter

Synthszr Take: SaaS companies are becoming the coral reefs of the digital economy: they form the foundation on which a complex ecosystem of specialized agents settles. The business model is shifting from “rent per head” to “commission per transaction,” which fundamentally changes the power dynamics. Oracle is already building agent skills for NetSuite, while Gemini is connecting agents directly to trading APIs. The real innovation lies not in better interfaces, but in agents understanding and navigating the complexity of enterprise software without humans ever having to click a button. Anyone still polishing prettier dashboards today is building carriage lamps for self-driving cars.

86-DOS: An archaeological find from the PC Stone Age

This week, Microsoft released the oldest DOS source code ever discovered. The release includes the 86-DOS 1.00 kernel, several development snapshots of the PC-DOS 1.00 kernel, and well-known utilities like CHKDSK. This code is so old that it predates the MS-DOS branding. Tim Paterson had developed 86-DOS (originally QDOS for “quick and dirty operating system”) for an Intel 8086-based computer kit from Seattle Computer Products. Microsoft licensed the system, hired Paterson, and later bought the rights outright. The company then licensed it to IBM as PC-DOS but retained the right to sell it to other companies as MS-DOS – a decision that established Microsoft's decades-long dominance in the PC market. → Ars Technica

Synthszr Take: Microsoft is conducting paleontology on its own behalf, unearthing the source code of its power. 86-DOS was the chance discovery that turned Microsoft into an operating system monopolist – like a fossil that explains an entire evolutionary line. The code shows: Microsoft's success wasn't based on technical superiority, but on perfect timing and shrewd licensing poker. IBM urgently needed an OS for its PC, Microsoft didn't have one, so it quickly bought Paterson's “quick-and-dirty” solution and turned it into a money-printing machine. The release feels like a gesture of generosity from an empire that is memorializing its own founding myths. Today, Microsoft can afford to give away the blueprints of its former power – they are only of historical value anyway.

Is the great AI fatigue setting in?

Mario Zechner, creator of the Pi project, and Armin Ronacher, the man behind the Flask web framework, spoke with The Pragmatic Engineer for 90 minutes. What emerged was not typical tech euphoria, but a refreshingly sober reckoning with the current AI hype. Both developers, whose code keeps large parts of the internet running, are increasingly skeptical of the flood of AI tools swamping the market. Zechner speaks openly of a kind of “AI fatigue” – the feeling that it's becoming difficult to distinguish between genuine technical progress and marketing noise. The discussion raises fundamental questions: What happens when the developers who build our digital infrastructure lose faith in the next big technology wave? → The Neuron

Synthszr Take: Zechner and Ronacher are not cultural pessimists, but engineers with a combined 40 years of experience building critical software infrastructure. Their AI skepticism is reminiscent of the reaction of experienced architects to the skyscraper boom of the 1920s: Yes, the technology is impressive, but does every small town really need a tower? Flask runs on millions of servers today without ever needing machine learning – a point Ronacher probably never tires of making. The two represent a growing group of senior developers who understand the difference between “can AI do it” and “should AI do it.” When the people who know how to build robust systems start rolling their eyes, it's a signal the industry should take seriously.

Intent: The Living Spec as the source of truth, not the grave of intentions

Augment Code has released a Mac-based development environment that solves the classic problem of decaying specifications. The core idea is Living Specs, which are automatically kept up-to-date because AI agents adapt them with every code change. A Coordinator agent breaks down the spec into tasks, delegates them to specialized Implementor agents, and has a Verifier check the results. The workspace environment integrates a code editor, browser, terminal, and Git in one window with a persistent state across sessions. Developers can integrate their own agents (Claude Code, Codex, OpenCode) or use the Augment Context Engine, which gives every agent access to the entire codebase. → Every

Synthszr Take: Intent transforms the Babylonian confusion of languages between specification and implementation into a self-regulating cycle. It's reminiscent of biological systems where DNA is both the blueprint and the result of evolution: the spec mutates with the code; the code follows the spec. Augment is betting that developers are willing to cede control over documentation to agents if they guarantee consistency in return. The real test will come when legacy code meets Living Specs. Presumably, companies will initially use Intent for greenfield projects, while legacy systems continue to gather dust in Confluence. In short: Intent is the new code.