ProfG Warns at OMR and Trump Suddenly Finds Regulation Sexy

OMR: Scott Galloway Warns Against AI Euphoria

Scott Galloway sees the AI industry in a dangerous dichotomy: While tech CEOs solicit investments with apocalyptic scenarios, the promised productivity boost fails to materialize. In an interview, the US professor argues that the industry has lost massive trust in just 18 months. The discrepancy between billion-dollar valuations and actual business models is becoming increasingly obvious. He is particularly critical of the “disaster rhetoric” from leaders like Elon Musk and Sam Altman, who he claims deliberately stoke fears of job losses. Yet historical data shows that technological revolutions have always created more jobs than they destroyed in the long run. Galloway sees the real danger not in the technology itself, but in the growing social isolation caused by digital substitute relationships. → MEEDIA Daily Update

Synthszr Take: Galloway is diagnosing a classic phenomenon of technology adoption: the overestimation of short-term effects and underestimation of long-term ones. His warning about the “loneliness machine AI” is reminiscent of Robert Putnam's “Bowling Alone,” only this time it's not television but personalized algorithms eroding social bonds. The tech CEOs are acting like medieval indulgence sellers: they sell salvation from fears they themselves have created. While they warn of the apocalypse, they are simultaneously building the infrastructure for the very future they warn against. The truly unsettling thing is not the technology, but that we as a society have forgotten how to distinguish between legitimate warnings and marketing. Galloway's prescription sounds old-fashioned but hits the core: storytelling, relational skills, and resilience are becoming scarce resources in a world where machines take over routine work.

Trump: Regulation is Great as Long as I'm the One Deciding

The White House is considering a review process for next-generation AI models before they become publicly available. The trigger is Anthropic's restricted Mythos model, which can detect software vulnerabilities, making it both a product and a state resource. While the Pentagon still classifies Anthropic as a supply chain risk, the NSA already wants to use the model to secure networks. The Trump administration speaks of “Model Safety,” but what it really means is first access: Who sees a cyber-capable model first, who bears responsibility after an attack, and will the review become procurement through the back door? Meanwhile, Greg Brockman's nearly $30 billion stake in OpenAI provides Elon Musk's lawyers with a clear figure for their lawsuit, while OpenAI points to an even larger foundation stake. In parallel, Anthropic is building a sales vehicle for enterprise AI services with $1.5 billion from Blackstone and others. → Marcus Schuler

Synthszr Take: The US government is rediscovering a concept as old as the printing press: pre-censorship. Gutenberg had to get his Bibles approved by the church before they were printed. Now, Washington wants to “review” AI models before they are unleashed on the world. The difference: the church wanted to prevent heresy; the NSA wants the tools first. Anthropic's Mythos model perfectly illustrates the dilemma: a system that finds security vulnerabilities is simultaneously a defensive weapon and an attack vector. The $1.5 billion bet on enterprise services shows where this is headed: AI will no longer be sold as software, but as controlled access to capabilities, with the government automatically having first-customer rights.

Google DeepMind Employees Unionize Against Military AI Contracts

Employees at Google DeepMind's London headquarters have voted to form a union to prevent the use of their AI technology by Israel and the US military. 98 percent of the Communication Workers Union (CWU) members at DeepMind support the move, which would affect at least 1,000 employees. In a letter to Google management, the employees demand concrete commitments: no development of weapons or surveillance technologies, a say in AI projects that affect their work, and the right to refuse projects on ethical grounds. An anonymous DeepMind employee puts it bluntly: “Even if our work is used only for administrative purposes, it makes genocide cheaper, faster, and more efficient.” The union initiative follows reports of Google's direct collaboration with the Israeli military and new contracts with the Pentagon that allow the use of AI models for “any lawful purpose.” → Techpresso

Synthszr Take: DeepMind researchers are experiencing their own “Oppenheimer moment”: after Trinity, the physicists in Los Alamos could no longer control what their work was used for. The difference: AI models can't be locked away in secure facilities like uranium. Once a Gemini model exists, Google can promise not to use it for weapons, but the same transformer architectures and training methods are already circulating in the open-source community. The union initiative feels like trying to bail water with a sieve: morally understandable, practically futile. What the DeepMind employees are really demanding is a kind of “Geneva Convention for Algorithms” (an idea that already failed in 2018). The real tragedy: their most brilliant research on protein structure prediction saves lives, while the same optimization methods could make drone swarms more efficient. Google's management faces a choice between employee loyalty and billion-dollar contracts with the Pentagon.

Palantir Booms When the Bombs Fall

Palantir Technologies is raising its annual revenue forecast after a surge in demand from the US government for AI-powered data analysis. The company now expects annual revenue between $2.805 billion and $2.809 billion, significantly above its previous forecast. Revenue from US government customers grew by 40 percent in the third quarter to $408 million. The new Artificial Intelligence Platform (AIP) is a particular driver of growth: Palantir managed to triple the number of users in US government agencies within three months. CEO Alex Karp speaks of “unstoppable” demand, while the company also expanded its commercial customer base by 51 percent to 498 clients. The stock rose over 13 percent in after-hours trading. → Techpresso

Synthszr Take: Palantir is perfecting the software industry's dealer principle: first, free demos for agencies, then exponential user expansion, and finally, structural dependence. Tripling government users in just one quarter shows classic network effects in closed systems. Once the Pentagon reconfigures its workflows around AIP, there's no going back. The business model is reminiscent of the early SAP strategy: become the operating system layer for organizational knowledge. The 40 percent jump in government revenue isn't just growth; it's the beginning of a lock-in phase. Palantir is no longer selling software; it's selling cognitive infrastructure.

Peter Thiel Funds Datacenters on the High Seas Instead of in Space

The AI industry is searching for new locations for its hungry data centers. A $140 million investment in Panthalassa, led by Peter Thiel, showcases a radical alternative: 85-meter-long steel structures that drift autonomously in the Pacific, converting wave motion into computing power. The floating data centers cool their AI chips with seawater, navigate without engines solely through their hull shape, and transmit results via Starlink. While Google and Elon Musk dream of data centers in space, Panthalassa is already planning a commercial rollout in 2027. Resistance to land-based data centers is growing—Oregon, where Panthalassa is building its pilot factory, knows these conflicts well. Thiel's statement that “extraterrestrial solutions are no longer science fiction” doesn't mean Mars, but international waters. → The Rundown AI

Synthszr Take: Panthalassa solves the NIMBY problem of AI infrastructure just as medieval Hanseatic cities solved their trade disputes: by taking to the open sea. The idea is reminiscent of the seasteading movement, which has been dreaming of libertarian tech utopias on floating platforms for years—only this time, there's an actual business model behind it. Wave power as an energy source is elegant: unlike solar parks or wind turbines, it works 24/7, and the sea provides cooling for free. But the real innovation lies in governance arbitrage: international waters mean no building permits, no resident protests, and no national data protection laws.

Why Amazon Could Lead Again in the Inference Era Despite a Training Deficit

In his Stratechery podcast, Ben Thompson explains why Amazon's apparent weakness in AI training could turn into a strategic advantage in the inference era. The key lies in Amazon's decades-long strategy of developing its own infrastructure primitives and then offering them as a service—a pattern that extends from AWS and logistics to AI infrastructure. With Nitro chips and ARM-based Graviton processors, Amazon has built its own silicon architecture that was initially inferior to Intel and AMD but could still be used profitably through Platform-as-a-Service offerings like RDS (Relational Database Service). While Nvidia dominates the training market with its GPUs, Amazon is positioning itself for the inference phase, where trained models must be computed billions of times. Amazon's advantage: inference workloads are more similar to traditional cloud computing tasks, where cost-efficiency and scale are more important than raw computing power. Thompson sees this as a repeat of the AWS pattern: Amazon first builds the infrastructure for its own needs, achieves economies of scale, and then monetizes it as a service for third parties. → Ben Thompson

Synthszr Take: Amazon is playing chess while others are playing checkers. The company has been investing for years in specialized inference chips optimized for running trained models, not for the training itself. This is reminiscent of the history of the railroad: the great fortunes were made not by the gold prospectors, but by the operators of the supply lines. Inference is the new Platform-as-a-Service—customers don't need the fastest chips, but the cheapest execution cost per token. Amazon's Graviton evolution shows the pattern: first, inferior hardware hidden behind managed services, then gradual improvement at structurally lower costs. The real disruption comes not from better AI, but from cheaper AI execution.

Vercel Launches deepsec: Security Agents Analyze Code in Parallel Sandboxes

Vercel has released deepsec, an open-source tool that automatically finds security vulnerabilities in large codebases. The command-line interface (CLI) allows multiple AI agents to run in parallel in isolated sandboxes to identify and validate vulnerabilities. Developers can run the tool on their own infrastructure with their API keys, ensuring control over sensitive data. A special focus is placed on reducing false positives, a chronic problem with automated security scans. The tool is available for free and positions itself as an alternative to commercial Security-as-a-Service offerings. → The Neuron

Synthszr Take: Vercel makes security audits a matter of infrastructure, not trust. While traditional security scanners operate like monolithic power plants, deepsec follows the principle of distributed solar panels: each agent works in its own sandbox, in parallel and autonomously. This is reminiscent of modern microservices architecture, but applied in reverse: instead of breaking down an application into services, the audit process itself is atomized. The strategic move is subtle: Vercel isn't just giving away software; it's establishing a new category between SaaS and self-hosted. Companies regain control but must provide the computing power themselves. The CLI becomes a Trojan horse for Vercel's true mission: to redefine developer infrastructure, one security check at a time.

The Transformation Paradox: AI Fails Due to Corporate Organization

Microsoft's new Work Trend Index 2026 reveals a bitter truth: 65% of surveyed AI users fear falling behind if they don't adapt to artificial intelligence quickly enough. At the same time, only 13% are rewarded for their AI experiments at work. The study, involving 20,000 participants worldwide, shows that the problem isn't the technology or the employees, but the rigid organizational structures. Matt Firestone, General Manager of Microsoft's Frontier Firm initiative, calls it a “transformation paradox”: while employees are ready to reinvent their way of working, metrics, incentive systems, and norms keep them trapped in old patterns. Microsoft argues that leaders must re-architect work itself, rather than just distributing new tools and hoping for spontaneous innovation. → Marcus Schuler

Synthszr Take: Microsoft is diagnosing a classic prisoner's dilemma in corporate culture here. All parties know that collaborating with AI tools would increase the overall output, but no one wants to be the first to expose themselves. The statistic is brutal: 87% of AI users are not rewarded for their experiments, while two-thirds are afraid of falling behind. This is reminiscent of the introduction of email in the 90s, when companies were still updating fax policies while their employees were already communicating digitally. Microsoft's solution (“re-architect work”) sounds like a consulting gig for Accenture, but the core point is correct: as long as companies treat AI like a better version of Excel, they will only see marginal productivity gains. The real transformation begins when tolerance for error and a willingness to experiment become part of performance evaluation.

The Jagged Technological Frontier: Harvard Study Shows Uneven Gains for Knowledge Workers

A Harvard study involving 758 consultants from the Boston Consulting Group shows how unevenly AI affects the productivity of knowledge workers. The researchers, led by Fabrizio Dell'Acqua and Karim Lakhani, coin the term “jagged technological frontier” for this phenomenon: for 18 typical consulting tasks within AI's capabilities, GPT-4 increased performance by enabling 12.2% more tasks to be completed, at a 25.1% higher speed, and with significantly better quality. However, for a complex management task outside this frontier, AI users produced 19% fewer correct solutions than the control group without AI. The study examined three groups: no AI, with GPT-4, and with GPT-4 plus prompt engineering training. The crucial finding is that even for seemingly similar tasks, it's impossible to predict whether AI will help or harm. → Latent.Space

Synthszr Take: The “jagged technological frontier” is reminiscent of the Mandelbrot set in mathematics: the closer you look, the more complex the boundary between “AI helps” and “AI harms” becomes. Boston Consulting is not just testing GPT-4 here; it is essentially mapping the fractal structure of human expertise. The 19% poorer performance on complex tasks demonstrates the phenomenon of “competence illusion”: knowledge workers overestimate AI in areas where they should be thinking for themselves. It's reminiscent of GPS navigation, which gets us to our destination faster but lets our sense of direction atrophy. Harvard is documenting the transition from software as a tool to software as a cognitive prosthesis—with all the risks of a prosthetic that sometimes works better than the original, but sometimes fails when you need it most.

AI Search Becomes the New Sales Channel: How B2B Brands Get into ChatGPT

A new study by AirOps shows how B2B software companies are systematically appearing in AI search results. The analysis of over 15 million AI search queries identifies four key strategies that companies like Ramp, Carta, and Webflow use to increase their visibility in ChatGPT, Gemini, and Perplexity. Within a few days, Webflow saw a 6 percent increase in AI-attributed sign-ups and quintupled its content update speed. Chime tripled the number of its AI search citations by being recommended in 68 priority search queries instead of the previous 24. The findings point to a fundamental shift: AI-based search is evolving from an experimental feature into a measurable sales channel with its own playbook and specific optimization strategies. → Techpresso

Synthszr Take: B2B software is reinventing SEO—only this time for machines instead of humans. The numbers are nice—Webflow: +6% AI signups, Chime: citations tripled. The preceding question is: where does my brand even appear today—in which models, for which customer segments, grounded in which sources? This is precisely the map that tools like raidar provide. Instead of random spot checks, thousands of queries per topic are mirrored across all relevant models and over time, in three layers: training content, chat responses, and AI Overviews. Fifteen statistical models separate signal from noise. Only then does the optimization begin. The lesson from the PageRank era: those who measure early what machines see, optimize for the reality that their buyers are served in ChatGPT.