Identity Crises at OpenAI, Palantir, and SaaS

Identity Crises (1): Sam Altman wants to make OpenAI a 'good citizen' again

On Sunday, OpenAI published a five-principle framework for the development of artificial general intelligence (AGI). In it, CEO Sam Altman promises to resist the concentration of power through AI technology and acknowledges that OpenAI is a much greater force today than when its original charter was published in 2018. The principles include democratization (AI decisions should be subject to democratic processes), empowerment (broad user freedoms with caution for unclear risks), universal prosperity (through global data centers and cost reduction), resilience (collaboration on bioweapon and cyber risks), and adaptability (transparent changes in position). Altman admits that OpenAI might have to 'sacrifice some empowerment for more resilience.' The timing is no coincidence: California passed the first state-level AI safety law last year, while OpenAI is under fire for its Pentagon cooperation. → www.implicator.ai

Synthszr Take: OpenAI is practicing regulatory prophylaxis like a pharmaceutical company that lists its own side effects on the package insert before the authorities ask. The five principles read like a self-commitment from a monopolist that has already cemented its market power: preaching democratization while controlling the infrastructure; promising empowerment but prioritizing resilience when things get serious. This is reminiscent of 1960s urban planning, where architects first drove highways through city centers and then planted green strips as compensation. Altman's admission that OpenAI has become 'materially bigger' sounds like the belated apology of a teenager caught crashing a house party. The real message is in the timing: just before regulators set the rules, OpenAI is quickly defining its own moral guardrails. Whoever controls the infrastructure writes the ethics.

Identity Crises (2): Palantir employees are losing faith

Palantir employees are beginning to fundamentally question their company's role in a second Trump administration. Internal Slack messages and interviews reveal a company in an identity crisis: the data analysis software, once developed to protect civil rights after 9/11, is now becoming the technological foundation of Trump's immigration policy. After the fatal shooting of nurse Alex Pretti by federal agents during protests against ICE, employees demanded clarification on Palantir's ties to the immigration agency. Management responded with philosophical evasions and deleted the critical Slack conversations after 7 days. A former employee sums it up: 'We were supposed to be the ones preventing abuse. Now we're enabling it.' → arstechnica.com

Synthszr Take: Palantir is currently experiencing what happens when infrastructure develops its own morality. The company was founded on the promise of being a technological immune system against the abuse of power—named after Tolkien's corrupting Palantír as an ironic warning of its own power. Now, the immune system is turning into an autoimmune process: the same software that was supposed to fend off external threats is being turned against its own population. This isn't a story about evil technology, but about the illusion of neutral infrastructure. When your employees are asking, 'Are we the baddies?' while management deletes Slack messages like an authoritarian regime purges newspaper archives, the infrastructure has already made its choice. Peter Thiel's bet was always that power is more important than principles—his employees are just beginning to understand that they are the chips in this game.

Identity Crises (3): Malus.sh clones any software

A new tool called Malus.sh uses AI to 'liberate' software from existing copyright licenses and create technically clean clones that do not violate copyright law. The method is based on the historic 'clean room' design process, where IBM's competitors in the 80s reverse-engineered its BIOS: one team analyzed the specifications, and a second 'clean' team programmed without any knowledge of the original code. What used to take months and two teams can now be done by an AI model in days. The project is intended as a satirical commentary on the open-source community, but it has real paying customers. The website advertises with 'No Attribution. No Copyleft. No Problem.' and thus hits a sore spot: Software-as-a-Service companies fear that competitors could simply replicate their expensive offerings, which has already led to massive stock price losses for companies like Oracle. → futurism.com

Synthszr Take: Malus.sh turns copyright into a pure optimization problem: whoever clones more cheaply than they license, wins. This is reminiscent of the game theory of the prisoner's dilemma, except here the cooperation option (licensing) is nullified by asymmetric cost structures. Historically, this is nothing new: the textile industry of the 19th century was based on stolen blueprints of British looms, but back then you needed industrial spies instead of Claude. The difference today: the marginal costs for clean-room reverse engineering are trending towards zero, while transaction costs for license negotiations remain constant. Software companies face the same challenge as music publishers after Napster: their business model is based on artificial scarcity through legal constructs that are becoming technically obsolete. The irony: Open Source wanted to liberate software, now AI is liberating software from Open Source.

Snap: our designers now write the code

Evan Spiegel, CEO of Snap, stated an old truth in a podcast with Lenny Rachitsky: in 15 years, only two consumer apps have truly broken through. Snapchat itself only survived because every one of its key features was copied—Stories, AR filters, swipe navigation. Spiegel's conclusion: pure software is no longer a moat. Hardware is the only real competitive advantage, and distribution is more important than the product itself. His 9- to 12-person design team works without titles or hierarchy, reviewing hundreds of ideas per week directly with the CEO. Designers at Snap now write code, and AI is fundamentally changing their workflow. → Lenny's Newsletter

Synthszr Take: Spiegel is describing the inversion of a Silicon Valley creed. It used to be: build a better product, and the users will come. Today, Meta, Google, and Apple control the distribution channels like medieval tollbooths on trade routes. Anyone without their own hardware (Spectacles, Pixy) is just a tenant in someone else's ecosystem. The irony: Snapchat invented the most important social media features of the last decade, but Instagram copied them faster than Snap could distribute them. Distribution beats product—unless you own the infrastructure on which the distribution runs.

ASML drastically ramps up production

ASML builds the world's only machines capable of producing advanced AI chips at a large scale. These chips power ChatGPT and Gemini; without them, there is no AI revolution. Now, demand is exploding: Microsoft, Meta, Amazon, and Google alone plan to invest over $600 billion in AI infrastructure this year. ASML is responding with a massive expansion: the Dutch company plans to build 60 of its standard EUV machines in 2026, 36% more than in 2025. The next generation of their machines costs over $400 million each. These school-bus-sized devices shoot lasers at molten tin droplets to create extreme ultraviolet light, which prints microscopic patterns on silicon wafers. CEO Christophe Fouquet says, 'We don't want to be the bottleneck for our customers.' → Wall Street Journal

Synthszr Take: ASML is the perfect example of Gall's Law in action: a functioning simple system (lithography) was expanded over decades into something so incredibly complex that no one can copy it anymore. The machines are like gothic cathedrals of semiconductor manufacturing, whose builders work in cleanrooms instead of on scaffolding. A single speck of dust can halt production, yet they now have to ramp up from 44 to 80 machines per year. ASML learned from the pandemic and is building cleanrooms in advance, before they are needed (the opposite of just-in-time). The real power lies not with the AI models, but with whoever controls physical reality: without ASML's machines, no chips; without chips, no data centers; without data centers, no ChatGPT.

Google now controls 25% of global AI computing capacity

Google has quietly become the dominant infrastructure provider of the AI era: according to calculations by Epoch AI, the company controls about 25 percent of global AI computing capacity with around 3.8 million TPUs (Tensor Processing Units) and 1.3 million GPUs. Google Cloud CEO Thomas Kurian justifies the massive investments with rising demand and corresponding revenues. The sheer size of Google's compute fleet exceeds the combined capacity of many national AI initiatives. While the focus is on OpenAI and Anthropic, Google is building the physical infrastructure in the background, without which modern AI models cannot be trained. → techmeme.us14.list-manage.com

Synthszr Take: Google is transforming into the TSMC of the AI era: invisible to end-users, but systemically important for anyone who wants to stay in the game. The 5.1 million processors are not a technical gimmick, but a new form of market power measured in compute cycles instead of market share. Historically, this is reminiscent of the 19th-century railroad barons who controlled who could transport their goods and where. The difference: Google's TPUs are proprietary hardware that only runs in its own cloud, while Nvidia sells its GPUs to anyone. This asymmetric availability creates a two-tier society in AI development: those who train on Google get access to specialized hardware, while everyone else has to fight for scarce Nvidia chips. Kurian's reference to rising revenues shows that Google is already cashing in on the infrastructure dividend.

DeepSeek/Kimi: Open source adapts better

While all eyes were on Claude Opus 4.7 and GPT-5.5, two open-source models have redefined the rules for agent-based AI. DeepSeek-v4 and Kimi-K2.6 mark a turning point: they show that open models are no longer just derivatives of the large proprietary systems, but are developing independent architectures with specific strengths. DeepSeek-v4 solves the memory problem in 1-million-token contexts with its Compressed Sparse Attention (CSA) and Heavily Compressed Attention (HCA)—a technical masterpiece that shows how resource scarcity leads to innovation. Kimi-K2.6 takes a different path: with 1 million parameters, of which only 32 billion are active, it optimizes for orchestration rather than pure performance. The model manages agent swarms for over 13 hours and 1,000 tool calls without losing its bearings. Both models are available under the MIT license, which allows for commercial use without restrictions. → AlphaSignal

Synthszr Take: DeepSeek and Kimi follow the pattern of biological evolution: when the dominant species (here: Claude, GPT) monopolizes all resources, specialized niches emerge. The Compressed Sparse Attention is reminiscent of information compression in the neural networks of insects—fewer neurons, but highly optimized for specific tasks. While OpenAI and Anthropic compete for absolute peak performance, models are emerging in the open-source space that are sufficient for 90% of real-world use cases. This recalls the history of microcomputers: IBM and DEC fought for dominance in the mainframe sector, while Apple and Commodore opened up the mass market. The key to Kimi-K2.6 is the orchestration of agent swarms—instead of a monolithic model, it coordinates specialized subsystems, much like an operating system manages processes. Open source wins not through superiority in every benchmark, but through adaptability to specific needs.

Amateur proves a 60-year-old mathematical conjecture with ChatGPT

A hobby mathematician named Ryan Greenblatt has solved a conjecture from combinatorics that had been open since 1964, using ChatGPT (o1-preview). The conjecture stated that periodic patterns always arise in special tiling systems if certain mathematical properties are met. Greenblatt, who is neither a mathematician nor an AI specialist, engaged in an iterative dialogue with the language model over several weeks: he asked questions, had it generate proofs, identified gaps, and refined the argumentation together with the AI. After 120 conversation rounds, he had a complete proof, which subject matter experts confirmed as correct. The solution uses techniques from various mathematical subfields that a single person could hardly comprehend. Greenblatt emphasizes that no breakthrough would have been possible without his targeted questions and his basic mathematical understanding. → Lenny's Newsletter

Synthszr Take: Greenblatt has applied the successful formula of Polynesian navigation to AI: instead of using instruments, Pacific seafarers navigated for centuries by observing wave patterns, bird flight, and cloud formations. It is precisely this kind of distributed intelligence—human intuition coupled with machine computing power—that is now solving mathematical problems that generations of specialists have failed to crack. ChatGPT was not used as an oracle here, but as an intellectual sparring partner that generates and discards hypotheses. The real innovation lies in Greenblatt's questioning technique: he treated the AI like a brilliant but sometimes confused doctoral student whose ideas are taken seriously but critically questioned. Amateur science is making a comeback, only this time the assistant doesn't need human patience.