The Singularity vs. The Circularity — 2026-05-05#

Highlights#

Today’s discourse is dominated by the spectacular revelations from the Musk vs. OpenAI trial, exposing deep ethical questions around fiduciary duties and self-dealing among top AI executives. Meanwhile, the reality of deploying AI in enterprise is hitting hard—from EPFL’s alarming study on high hallucination rates in cutting-edge models to Coinbase laying off 14% of its staff to fundamentally restructure into an “AI-native” organization. It is a day of reckoning that sharply contrasts the soaring capabilities of new model drops, like OpenAI’s GPT-5.5, with the harsh realities of corporate governance, software reliability, and workforce displacement.

Top Stories#

• OpenAI Co-Founders Accused of Self-Dealing in Trial: Musk’s legal team revealed Greg Brockman’s diaries detailing early plans to oust Musk, and highlighted undisclosed personal investments in Cerebras by Brockman and Altman right before OpenAI committed $20+ billion to the hardware firm. Brockman admitted under oath to the overlap but lacked any evidence of disclosing this to Musk, raising serious questions about fiduciary duties under California charitable-trust law. (Source)
• Coinbase Slashes 14% of Staff in Pivot to “AI-Native” Pods: CEO Brian Armstrong announced a massive restructuring to rebuild Coinbase with “AI at our core,” eliminating pure management roles and moving to smaller, AI-native pods. The move reflects a broader industry shift where coding agents are drastically reducing the time required to ship production code, enabling non-technical teams to ship at an unprecedented pace. (Source)
• Top Models Fail Factual Hallucination Test Miserably: A new paper from EPFL and the Max Planck Institute demonstrates that cutting-edge models like GPT-5 and Claude Opus 4.5 still hallucinate up to 71.8% of the time on hard, domain-specific queries. The study shows that even with web search enabled, models routinely fabricate information, producing erroneous answers in 9 out of 10 medical guideline queries. (Source)
• GPT-5.5 Instant Rolls Out to ChatGPT: OpenAI shipped a major update to its default model, focusing on factuality, baseline intelligence, and thwarting hacks. Sam Altman and the OpenAI team highlighted its massive capability jump and drastically increased Codex rate limits, which prompted both excitement for the performance and skepticism from the developer community regarding future pricing rugs. (Source)
• Perplexity Computer Launches Professional Finance and Medical Tiers: Perplexity introduced dedicated workflows pulling from licensed financial databases like Morningstar and PitchBook, as well as premium medical journals like NEJM. The update emphasizes strict traceability to original sources for high-stakes professional research, attempting to bypass the hallucination issues plaguing raw LLMs. (Source)

Articles Worth Reading#

Andrew Ng on How Coding Agents Accelerate Different Software Work (Source) Ng breaks down the varying impact of coding agents across the software stack, creating a crucial mental model for engineering leaders. He argues frontend development sees the most dramatic speedup, followed by backend work, while infrastructure and core research see the least benefit from current AI. This nuanced take is essential for practically organizing software teams around AI tools rather than falling for generic productivity hype.

Alex Karp on AI “Slop” vs. Working Software (Source) Palantir’s CEO sharply contrasts genuine, battle-tested software engineering with the deceptive fluency of generative AI outputs. He warns that the appearance of software working is not actual functionality, highlighting how generative systems fail seductively at the edges of permissions, security, and accountability. It’s a much-needed reality check against the hyperbolic claims that AI slop will seamlessly replace all engineering jobs.

Claire Vo’s Playbook for Surviving AI-Driven Layoffs (Source) In response to the Coinbase restructuring, Claire Vo offers a highly tactical guide for individual contributors displaced by AI automation. She advocates aggressively reskilling by using tools like Claude Code and Cursor to automate one’s previous job functions, and then building a public portfolio of these AI-native skills. It is a pragmatic, contrarian take that embraces the “player-coach” model and urges workers to become AI builders before the adoption gap widens further.