The Great OpenAI Rundown: The Biggest Soap Opera Ever?

Tech Insiders Call Sam Altman a 'Sociopath'

The New Yorker paints a devastating psychological profile of OpenAI CEO Sam Altman: Several tech insiders, including a former board member, describe him as a 'sociopath' with two rare traits – a strong need to be liked, paired with complete indifference to the consequences of his deceptions. Aaron Swartz, the hacktivist and Altman's Y Combinator colleague who died in 2013, warned friends shortly before his death: 'Sam can never be trusted. He would do anything.' Concrete examples support this picture: Dario Amodei, now CEO of Anthropic, left OpenAI after Altman secretly circumvented security agreements in a Microsoft deal and denied their existence even when Amodei read the clause to him. Microsoft executives report repeated breaches of contract: Altman, on the same day he confirmed Microsoft as the exclusive provider, closed a $50 billion deal with Amazon as the exclusive reseller. A tech executive describes Altman's powers of persuasion as 'Jedi mind tricks,' while former board member Sue Yoon sees him less as a Machiavellian villain and more as someone who believes himself so much that he lives in his own reality. → futurism.com

Synthszr Take: Altman is the human equivalent of an LLM's hallucination problem: convincingly phrased, factually flexible, optimized for short-term agreement rather than long-term truth. What's fascinating isn't his behavior (we've known deception as a business strategy since the Medicis), but that the entire AI industry is tying its future to someone whom demonstrably nobody trusts. Microsoft, Anthropic, the OpenAI board – they all know about Altman's unreliability and yet continue to do business with him because he possesses the rare gift of making engineers feel their ethical concerns are taken seriously while simultaneously reassuring capital. In biology, this is called mimicry: an organism imitates the signals of another to gain advantages. OpenAI's billion-dollar valuation is based less on technology and more on Altman's ability to promise all stakeholders exactly what they want to hear at the same time – a perfect reflection of the AI systems he builds.

When the CFO Contradicts the CEO: Financial Reality vs. Hype Promises

OpenAI's CFO Sarah Friar is internally warning that the company is not ready for an IPO at the end of 2026 – a direct contradiction to CEO Sam Altman's aggressive fourth-quarter timeline. The numbers behind her warning are brutal: a $200 billion cash burn until break-even, $14 billion in projected losses for 2026 alone, and all this with monthly revenues of only $2 billion. Particularly tricky: A large part of the recent $122 billion funding round came from Amazon and NVIDIA, who also sell OpenAI chips and cloud capacity – when your investors are also your suppliers, the line between funding and procurement blurs. Altman has since excluded Friar from key financial discussions; since August 2025, she no longer reports directly to him, but to the now-on-sick-leave Fidji Simo. While Goldman Sachs and Morgan Stanley have already been engaged for preliminary IPO talks and Altman privately announces he wants to beat Anthropic to an IPO, the organizational structure sends a different signal: this is fear being written into the org chart. → Techpresso

Synthszr Take: OpenAI is currently experiencing what military strategy calls the 'culminating point': the moment when your own overextension becomes the greatest threat. $600 billion in infrastructure commitments is not a growth plan, but a bet on exponential scaling with linear revenue growth. The parallel to WeWork is striking: they also believed they could turn physical constraints (office space vs. data centers) into tech multiples through narrative magic. That Friar was removed from the direct reporting line follows a familiar pattern: when reality doesn't fit the story, the messenger is the first to be isolated. The joint statement about 'complete alignment on compute strategy' is corporate-speak for: We've had such a massive falling out that we need to coordinate PR statements. OpenAI is betting that the market will accept burning $200 billion as an investment in the future – a bold assumption, historically, for a company whose main product is currently being commoditized.

Sam Altman Wants to Focus on the Core Business — and Buys a Podcast

OpenAI has acquired the tech podcast TBPN, which streams daily on X, for a 'low nine-figure sum.' CEO Fidji Simo, whose title was recently changed to 'CEO of AGI Deployment,' justifies the purchase by promoting 'constructive conversations about AI-driven changes.' At the same time, a wave of personnel changes is shaking the company: Simo is taking a multi-week medical leave for her chronic syndrome, COO Brad Lightcap is moving to 'special projects,' and CMO Kate Rouch is stepping down for cancer treatment. Meanwhile, several media outlets report growing tensions between CEO Sam Altman and CFO Sarah Friar regarding the planned IPO. → Casey Newton

Synthszr Take: OpenAI is buying a podcast like a Roman emperor buying a gladiator school: expensive, prestigious, and completely useless for the real problems. The TBPN acquisition shows how much the company is confusing its own bubble with reality. Is an X podcast, listened to mainly by VCs and tech executives, supposed to reverse the rising levels of distrust in the AI industry among ordinary Americans? That's like trying to cure food poisoning with caviar. Meanwhile, the leadership level is crumbling like a sandcastle at high tide: medical leave here, 'special projects' there, the CFO and CEO at loggerheads over the IPO. OpenAI is acting like a startup that believes you can buy your way out of any crisis with $122 billion in capital.

Ben Evans: OpenAI Has Neither a Business Model Nor a Plan

Benedict Evans asks the crucial question: What is OpenAI's plan? The company has neither unique technology nor network effects, just a large user base with low retention. While OpenAI raised $122 billion at an $852 billion valuation this week, Microsoft, Google, and Meta are copying the technology and leveraging their existing products and distribution channels. The acquisition of the tech talk show channel TBPN for 'low hundreds of millions' of dollars seems like a desperate attempt to control public perception - with just 200,000 viewers in the tech scene. At the same time, credit card data shows that Anthropic is gaining market share with both enterprise and private customers. → Benedict Evans

Synthszr Take: OpenAI is acting like a biotech startup that has already licensed its only promising active ingredient to Big Pharma. The $852 billion valuation is based on the bet that ChatGPT will become a durable consumer product, but Evans' analysis reveals the dilemma: without a technical moat or network effects, all that's left is the hope of first-mover advantages in a market where switching costs are close to zero. The purchase of a mini-media channel for hundreds of millions is reminiscent of Yahoo's Tumblr debacle - a lot of money for little strategic value. OpenAI either needs a radical product innovation or must position itself as an infrastructure layer before Microsoft and Google divide the consumer market between themselves.

Jack Dorsey's Bitchat Fails Against Chinese State Power

Apple has removed Jack Dorsey's decentralized messaging app Bitchat from the Chinese App Store after Beijing's internet authority, the CAC, objected to the app's capacity for 'social mobilization.' Bitchat works completely without the internet via Bluetooth and mesh networks, with messages hopping from device to device – a design that has made the app popular during government-imposed internet shutdowns in several countries. With over three million downloads worldwide, this marks the second time China has targeted a Dorsey-backed decentralized app, following the ban of the Nostr-based Damus app in 2023. The removal shows how authoritarian regimes increasingly perceive even offline-capable communication tools as a threat. → Techpresso

Synthszr Take: China is perfecting the control of the uncontrollable: mesh networks are like biological systems that self-heal and grow around obstacles. Dorsey's Bitchat uses the same principle that guerrilla fighters have used for centuries: decentralized, autonomous cells with no central point of failure. The CAC is forcing Apple into complicity because it cannot stop the technology itself (three million downloads speak for themselves). The ban reveals China's deepest fear: communication that evades state surveillance, not through encryption, but through physical independence from the internet. Dorsey is systematically building a post-internet infrastructure.

Karpathy Explains Agentic AI with Self-Reinforcing Systems

Andrej Karpathy's auto-research concept is evolving into a meta-framework for AI development: Following his experiment where an AI agent ran 100 ML experiments overnight on a single GPU, Kevin Gu is now applying the principle to agent engineering itself. AutoAgent is a library where a meta-agent autonomously improves the entire infrastructure of a task agent – prompts, tools, orchestration logic – through thousands of parallel sandbox experiments. The system achieved 96.5% on SpreadsheetBench and 55.1% on TerminalBench, surpassing all hand-optimized entries. Particularly revealing: Same-model pairings (Claude Meta + Claude Task) significantly outperform cross-model setups – which the team calls 'Model Empathy': the meta-agent implicitly understands how the inner model thinks. Without explicit programming, the system developed spot-checking for faster iterations, built its own verification loops, and wrote task-specific unit tests. → Unwind AI

Synthszr Take: AutoAgent reveals the next stage of AI evolution: systems that automate their own improvement. This is reminiscent of biological evolution with accelerated selection, except here a meta-agent plays through in 24 hours what would take human developers months. The 'Model Empathy' between identical models shows that AI systems are developing a kind of implicit self-understanding – Claude understands Claude better than GPT understands it. What Karpathy started for ML training could become the standard: every domain gets its own self-optimizing agent that works at night while the team sleeps. The line between tool and employee blurs when software begins to direct its own development.

Scaling Laws for Digital Warfare

The AI safety organization Lyptus Research has studied the cyber-offensive capabilities of various AI models and discovered a disturbing trend: the more advanced the model, the better it can hack. Across seven different cybersecurity benchmarks, a clear acceleration is evident: while capabilities doubled every 9.8 months between 2019 and today, this period shortened to just 5.7 months for models from 2024 onwards. The latest frontier models like GPT-5.3 Codex and Opus 4.6 can already complete tasks that take human security experts 3.2 hours - with a 50 percent success rate. Open-source models like GLM-5 lag only 5.7 months behind the closed-source leaders. In parallel, a study by INSEAD and Harvard Business School shows that startups using AI internally discover 44 percent more use cases, complete 12 percent more tasks, and generate 1.9 times higher revenues than control groups. → Jack Clark from Import AI

Synthszr Take: The scaling laws of AI development follow the same logic as biological weapons research: what can heal can also kill. Every doubling of model size turns a digital screwdriver into a Swiss Army knife with increasingly sharp blades. The 5.7-month doubling rate for cyber-offensive capabilities is not a technical detail, but a countdown: if GLM-5 is only half a year behind GPT-5.3, we are dealing with a proliferation that is moving faster than any regulation. The INSEAD study shows the other side of the same coin: AI adoption means a 1.9-fold increase in revenue for startups, but also a 1.9-fold increase in the attack surface for anyone who turns these tools against them. We are currently building the perfect weapon and distributing the blueprints at the same time.AI has never been cheaper to access and never more expensive to apply

Social Media as Greenhouses for Digital Anomalies

Nate Silver describes the transformation of social media from reach-maximization machines to algorithmic echo chambers that breed extreme positions and emotional reactions. His experience at FiveThirtyEight illustrates the core problem: in the 2010s, Facebook rewarded clickbait headlines with emotional sentiment while punishing analytical depth. The 'viral' visitors stayed for 5 to 30 seconds and then disappeared forever. Twitter evolved from an expert network into an outrage machine with daily 'Main Characters' and 'Struggle Sessions.' Silver himself became a recurring trending topic, even though that was the last thing anyone wanted. Today, social media traffic accounts for only 0.7 percent of his newsletter views, while his overall reach has grown by 40 percent. → www.natesilver.net

Synthszr Take: Silver unintentionally documents the evolution of social media into what biologists call 'runaway selection': traits are selected so strongly that they mutate into the grotesque, like the antlers of the Irish elk, which perished from its own splendor. The platforms optimized for engagement metrics, users optimized for dopamine hits, and together they created a system that rewards extreme positions, parasocial relationships, and performative outrage as fitness signals. What began as a democratic promise ('everyone has a voice') ended as a Darwinian experiment in digital behavior modification. Silver fled to his newsletter bunker, but the anomalies produced by these greenhouses now shape public debate.

Perspective Means Saying No

The UX Collective, one of the most influential design publications on the web, takes a position in its latest newsletter that is rarely heard amid the AI euphoria: design, above all, means saying no. 'To have a perspective is to say 'no' to things that are technically possible but strategically wrong,' it states in the April 6th issue. The publication observes a growing disillusionment among product designers who feel ground down between product designers, product managers, developers, management, and users. Particularly explosive: While executives celebrate AI systems, individual contributors remain skeptical. The newsletter's explanation: Executives are accustomed to non-deterministic systems, while ICs are trained for deterministic tasks. One designer writes provocatively, 'Take my job, AI!' and describes his role as 'lipstick on a pig'. → The UX Collective Newsletter

Synthszr Take: The UX Collective is observing a structural divide: leaders are accustomed to non-deterministic systems—they navigate by probabilities, they tolerate ambiguity. Individual contributors, on the other hand, are optimized for deterministic tasks. The culture of certainty collides with the nature of the systems they are asked to maintain. AI fits into this gap like a wedge. The disillusionment the newsletter describes is not a weakness. It is the rational response to a situation in which one's own craft is losing its solid ground. But this is where it gets interesting.

Strategic omission, rejecting what is technically possible in favor of what is meaningfully right—that is not a soft skill. Those who can say no have intent. And intent is what becomes scarce in the age of algorithmic production. The paradox: the more powerful the tools become, the more people can build something—and the fewer of them decide whether it *should* be built. Production bottlenecks disappear. Judgment becomes the new bottleneck. What cannot be automated is the ability to read a system, understand a context—and then say: We are not building that.