AI Powerhouses at a Dead End and the New World Disorder

Altman, Musk, Zuckerberg: 'Sorry, We Took a Wrong Turn'

OpenAI CEO Sam Altman has admitted that pure scaling will not lead to Artificial General Intelligence (AGI) and that the industry needs “mega breakthroughs” beyond current Transformer architectures. Just 14 months ago, Altman claimed his company already knew “how to build AGI.” In parallel, other tech giants are backing away from the pure scaling thesis: Elon Musk admitted xAI was “not built right,” Mark Zuckerberg postponed Meta's latest model, and DeepMind's Demis Hassabis, as well as AI pioneers Yann LeCun and Ilya Sutskever, are publicly questioning the previous strategy. Gary Marcus, who warned about the limits of scaling back in 2022, feels vindicated. Despite these admissions, the corporations are still planning trillion-dollar investments in data centers—a strategy Marcus calls a “bad bargain.” → Gary Marcus from Marcus on AI

Synthszr Take: Altman has done a 180-degree turn within 14 months—this isn't a minor adjustment, but a complete strategic pivot. The man who sold AGI as a solved problem is now searching for “mega breakthroughs” and new architectures. When even the loudest scaling evangelists capitulate, the AI industry's business model is fundamentally broken. The trillion-dollar data centers are becoming concrete tombs for a bet that isn't paying off. Companies should immediately reconsider their AI budgets—anyone still betting on pure computing power today is investing in a dead horse.

Europe: 'Great, We're Trying Something Radically Different'

AI pioneer Yann LeCun's AMI Labs is raising $1.1 billion—the largest seed round ever received by a European startup. Swedish legal AI company Legora is raising another $550 million, reaching a valuation of $5.5 billion. Both deals mark a turning point: Europe is finally producing unicorns on a weekly basis, from Wayve ($8.6 billion) to Black Forest Labs ($3.25 billion). At the same time, the brain drain is reversing—more tech experts are moving from the US to Europe than the other way around. The exits are also working: in 2025, AI companies in Europe were sold for $14.35 billion, and the Cognigy deal for $955 million was just the beginning. → Handelsblatt KI-Briefing

Synthszr Take: What's exciting right now is not just that the old scaling narrative is crumbling, but that Europe's counter-bet is gaining real weight for the first time. Yann LeCun's AMI Labs is launching with Europe's largest seed round to date, operating out of Paris among other locations, and is deliberately not building another LLM variant, but rather world-model-based systems with persistent memory, reasoning, planning, controllability, and safety. This is not an academic side project, but a seriously capitalized alternative to the Silicon Valley reflex of throwing more GPUs and more tokens at every problem; Reuters explicitly describes AMI as a test of LeCun's thesis that today's LLMs are insufficient for human-like reasoning and autonomous agents. Whether LeCun will be proven right in the end is an open question. But the more visible the limits of pure scaling become, the more relevant AMI becomes: as a European counter-bet on an AGI path beyond LLM scaling, and as a signal that the next leap could come from new architecture rather than from even larger data centers.

Benedict Evans: OpenAI Has the Technology, But No Strategy

OpenAI faces a strategic dilemma: its technology is no longer unique, user engagement remains weak, and network effects are completely absent. While competitors like Google and Microsoft offer the same AI performance with superior product integration and distribution channels, OpenAI is struggling to secure a sustainable market position. The most valuable applications will likely be entirely new user experiences that OpenAI cannot develop all on its own. Benedict Evans compares the situation to the desktop Linux problem: technically impressive for developers, but “almost ready” for end-users for 25 years. The real question is no longer who develops the best AI, but who integrates and scales it most effectively into existing workflows. → Benedict Evans

Synthszr Take: While OpenAI has built a large user base, without network effects or strong customer loyalty, it becomes a cost trap instead of a moat. Google and Microsoft can embed their AI models directly into billions of existing user accounts (Gmail, Office, Search), and Anthropic has found its 'niche' with developers and in the enterprise market. OpenAI still has to convince each user individually. The most valuable AI applications emerge where companies can integrate their specific data and workflows—exactly what cloud providers solve better with their enterprise infrastructure. OpenAI's only chance lies in developing applications so revolutionary that users are willing to abandon their familiar tools. Time is running out: the longer it fails to differentiate, the more OpenAI becomes a commoditized API provider in a market with shrinking margins.

Quadratic Scaling Meets an Exponential Hunger for Context

The problem of token explosion will become a critical bottleneck for AI systems in 2026. Thinking models produce thousands of tokens just for internal calculations, while agent AI requires huge context windows to orchestrate complex processes. A 64k context window was considered a luxury in 2024—today it is practically useless for real applications. The core problem lies in the quadratic scaling of standard attention: if the context window doubles, the computational cost quadruples. With each subsequent doubling, the demand increases fourfold again—a vicious cycle that will eventually make context windows unaffordable. The math works against unlimited growth, while AI applications demand exactly that. → dravian

Synthszr Take: 64k tokens were a breakthrough two years ago; today, they are barely sufficient for serious work. Quadratic scaling means double the context length costs four times the compute—this quickly adds up to astronomical amounts. AI labs are hitting a physical wall: more context requires exponentially more energy, but complex tasks demand precisely this context. Agent systems are particularly affected because they need to hold long chains of actions and decisions in memory. This forces the industry to make fundamental architectural changes—linear attention, state-space models, or completely new approaches will become vital for survival. Those who sleep through this efficiency revolution will be pushed out of the market.

Silicon Valley Fights for Chinese AI Talent—China Laughs

Elon Musk's xAI shows the new face of Silicon Valley: 80% of the faces in the team photo are Chinese, and five of the twelve founders are Chinese scientists. Meta is paying Chinese AI experts $300 million over four years—on par with NBA stars or Fortune 500 CEOs. Mark Zuckerberg is personally leading a project called “Super-intelligent Laboratory” to poach the world's top 50 AI experts; eight of the 14 names on the list are Chinese. While this talent war rages, China is developing its own AI champions in parallel: DeepSeek achieves ChatGPT-level performance at one-thirtieth of the cost, and Huawei's Ascend 910C beats Nvidia's H20 chip at half the cost. The irony is perfect—Jensen Huang wears a Tang suit in Beijing and gives his first speech in Chinese, while Chinese entrepreneurs like Scott Wu (Cognition, $2 billion valuation) and Wang Tao (Scale AI, $14.3 billion acquisition) are rewriting the rules in Silicon Valley. → Last Week in AI

Synthszr Take: Meta's $300 million contracts for Chinese AI researchers are panic money. Silicon Valley has a structural problem: 65 of the world's top 100 AI experts are Chinese, but only 15 work in US research centers. The arbitrage no longer works when DeepSeek achieves ChatGPT performance at one-thirtieth of the cost (and gets 110 million downloads without marketing on the side). Chinese scientists are increasingly returning to China, where they can combine academic positions with entrepreneurship. The real shift isn't in salaries, but in the market: whoever sets the cost-performance standards controls the market.

Petrochemical World Order Meets Semiconductor Geopolitics

More than 100 ships pass through the Strait of Hormuz daily, transporting one-fifth of the world's oil and gas production. Iran is using this control as leverage against the US and Israel after their bombardments put the religious leaders in Tehran on the defensive. The blockade follows a familiar pattern from the 1980s, but this time Iran is extending its attacks to digital infrastructure: AWS data centers in Bahrain and the Emirates, as well as US technology corporations like Google, Microsoft, Nvidia, Oracle, and Palantir, are being declared legitimate targets. In parallel, China is reducing its energy dependence through massive investments in renewable energy and electric vehicles to avoid being blackmailed in the event of a Taiwan crisis. Trump's control over 70 percent of Venezuelan and 90 percent of Iranian oil exports creates bargaining power with Beijing. Iran itself remains vulnerable, as the country must import gasoline and diesel despite its oil wealth. → Philipp Kloeckner

Synthszr Take: Iran is declaring Nvidia, Google, and Oracle legitimate drone targets, thereby making technology corporations parties to the war. China's 5-year plan aims for energy independence to break free from the petrochemical chokehold (70 percent of Venezuelan and 90 percent of Iranian oil flows to China). Whoever controls Taiwan's chips while also being energy self-sufficient will win the systemic competition between autocracy and democracy. The Strait of Hormuz is becoming a bottleneck for three million tons of oil daily, while data centers in Bahrain form the new front line of warfare. Energy and semiconductors are merging into a single geopolitical dual-use tool.

AI Fakes About Iran War Cause Chaos—And That Was Just the Beginning

The New York Times has identified over 110 AI-generated videos and images related to the Iran war, which have been shared millions of times on X, TikTok, and Facebook. The fakes show fabricated explosions in Tel Aviv, burning US warships, and crying soldiers—a digital parallel world that appears more dramatic than real war footage. One particularly widespread video simulates missile attacks on Tel Aviv from a balcony, complete with an Israeli flag in the foreground (a typical AI artifact). Since the release of OpenAI's Sora, practically any realistic war video can be created with a simple text prompt. The majority of the AI content supports pro-Iranian narratives and is intended to increase public war fatigue. Social media platforms are doing little to combat the flood of AI videos overwhelming them. → Last Week in AI

Synthszr Take: 110 identified AI fakes in two weeks show how quickly generative tools can become weapons of war. Iran strategically uses the Hollywood aesthetic of AI videos—exaggerated explosions and mushroom clouds appear more convincing than real cell phone footage from a distance. With Sora, OpenAI has essentially handed everyone a propaganda factory with no significant safeguards for conflict scenarios. The platforms' reactions are predictably helpless: X only began financially penalizing AI war content last week, after millions had already been manipulated. The Iran war is the first conflict of the Sora era—and it shows that AI disinformation is no longer a secondary tool for shaping opinion, but an integral part of warfare itself.

Journalism in the Post-Truth Era: When No One Trusts the Media Anymore

Trust in national news organizations in the US has fallen to 56%—a 20 percentage point drop within ten years, as reported by the Pew Research Center. In the 'post-truth' era, the media battles daily with lying politicians, ignored human rights violations, and AI-generated deepfakes that can manipulate visual narratives without a trace. The Editor-in-Chief of the Oakland Post describes his personal transformation from a shy student to a journalist despite persistent prejudices against the profession. He emphasizes that authentic and accurate journalism remains possible even in times of widespread skepticism. The newsroom is intended to continue serving as a learning lab and a refuge for students seeking objectivity and empathy in a polarized world. → Oakland Post

Synthszr Take: A 20 percentage point loss of trust in a decade shows the speed of systemic failure. Local newsrooms are becoming the last functioning institutions, while major media brands are caught in credibility spirals. The Oakland Post editor unknowingly describes the recipe against the crisis of trust: direct relationships with the community, transparent learning processes, real names instead of corporate speak. While Meta and Google burn billions on fact-checking, a campus newsroom solves the problem with simple authenticity. This isn't idealism; it's market logic: trust doesn't scale, it's only built through personal accountability.

Uber Eats Uses Dynamic Pricing for Margin Optimization

Business Insider conducted an experiment: several editors ordered the same McDonald's meal via Uber Eats at the same time. The prices varied significantly between users, even though they placed identical orders. The test shows how delivery services are increasingly using variable pricing—a practice that would normally be unthinkable for physical products. Companies are using technology ever more cleverly to adjust prices situationally and achieve higher margins. The problem: customers often don't even notice they are paying different amounts. The authors conclude that while this may make sense in certain situations, it is fundamentally questionable—everyone should pay the listed price for a product. → Business Insider

Synthszr Take: Business Insider has documented what Amazon has been perfecting for years: algorithmic price discrimination. Uber Eats is testing different prices for identical McDonald's meals for different users—a mechanism that would be unacceptable in the physical world. Platforms can do this because they control the pricing process while eliminating price transparency (in a supermarket, you can see what others are paying). The technology enables perfect price discrimination: purchasing power, location, order history, and even the phone's battery level can be evaluated in real time. Frequent orderers pay more; infrequent ones get introductory offers. This is efficient from the platform's perspective, but it's a fundamental break with the principle of equal prices for equal services.