The bipartisan Great American AI Act would preempt state AI laws for three years
The bipartisan Great American AI Act introduces a 269-page federal governance framework that would preempt state AI laws for three years and require frontier developers to disclose governance practices, the Trump administration explores an unprecedented equity stake in OpenAI as part of a public wealth fund concept, and AI agent adoption accelerates as practitioners at firms like Jane Street replace traditional design tools with Claude Code.
Regulatory & Policy #
Great American AI Act: Bipartisan draft proposes three-year preemption of state laws #
Rep. Obernolte’s Office / Roll Call / ITIF
Representatives Jay Obernolte (R-CA) and Lori Trahan (D-MA) released a 269-page discussion draft establishing a federal AI governance framework that would preempt state AI laws for three years and require frontier developers with over $500 million in annual revenue to publish governance frameworks including catastrophic risk assessments. The bill formally establishes a Center for AI Standards and Innovation within the Commerce Department, funded at $300 million over three years, with independent auditing through licensed verifiers. For anyone building or deploying AI systems, the state preemption clause is the highest-impact provision: it would replace the current patchwork of state-level regulations – including Colorado’s AI Act, which takes effect June 30 – with a single federal standard, significantly simplifying compliance for multi-state operations.
Sriram Krishnan is leaving his role as White House AI advisor #
TechCrunch
Sriram Krishnan will step down as White House AI advisor at the end of June, planning to establish an outside institution to continue shaping the administration’s AI policy with a focus on energy infrastructure and data centers. The departure creates a transition window in federal AI policy leadership at a moment when the Great American AI Act, executive orders on AI security, and government-industry partnerships like the potential OpenAI equity stake are all in motion simultaneously.
Funding & Business #
The Trump administration might take an equity stake in OpenAI #
TechCrunch
President Trump said he is discussing deals “where the American people can benefit from the success of AI,” potentially including a government equity stake in OpenAI as part of a “Public Wealth Fund” that would distribute AI profits to citizens. The proposal mirrors the administration’s 2025 Intel stake and has drawn sharply divergent reactions: Senator Bernie Sanders supports a similar 50% stock tax on AI firms, while critics warn of deepening corporate-government entanglement. If implemented, direct government ownership in a frontier AI lab would create unprecedented structural entanglement between the regulator and the regulated.
Developer Tools #
I design with Claude more than Figma now #
Jane Street / Hacker News (172 points)
Jane Street designer Edwin Morris describes replacing Figma mockups and design specs with Claude Code, using iterative natural language prompts to build working prototypes directly in the codebase. The workflow eliminates intermediate deliverables – mockups, design documents, review meetings – by producing “living proposal docs” that stakeholders can use immediately, shifting design from persuasion-based (convincing engineers to prototype) to execution-based (designers building proofs of concept themselves). The tradeoff is real: reviewers receive fully-formed features rather than design-space concepts, and designers risk constraining their thinking to what the AI can produce – but the productivity gain from “unlimited iteration, unbothered when I changed my mind for the 50th time” is hard to argue with.
Harness engineering: Leveraging Codex in an agent-first world #
OpenAI / Hacker News (202 points)
OpenAI published a piece on how AI coding agents integrate into engineering workflows in an agent-first development paradigm. The adoption of cloud-based AI coding agents by enterprise engineering teams signals that agent-first development is moving beyond individual productivity into organizational infrastructure.
Her: A detective for your Claude Code sessions #
Hugging Face Blog
Built during Hugging Face’s Build Small hackathon, Her analyzes Claude Code session trace files to reconstruct activity in plain English, flag risky operations (deploys, config changes, secrets exposure), and trace them back to specific conversation turns. The tool keeps its core analysis engine deterministic – the LLM handles only prose summaries – and runs a local Nemotron-Mini-4B model with no external API calls and auto-deleting sessions. As agent-mediated coding becomes routine, forensic tooling that answers “what did the agent actually do and what did it cost?” fills a growing observability gap between raw logs and human understanding.
Research & Papers #
Tokenomics: Quantifying Where Tokens Are Used in Agentic Software Engineering #
arXiv / Hacker News (108 points)
This paper analyzes token consumption in LLM-based multi-agent software engineering systems, finding that iterative code review accounts for an average of 59.4% of total token usage – not initial code generation, which is commonly assumed to be the primary cost driver. Input tokens represent roughly 54% of total consumption, indicating substantial inefficiency in agent-to-agent communication during refinement cycles. For teams deploying agentic coding systems, the finding reframes cost optimization: the largest savings come from improving review and verification workflows, not from making initial generation cheaper.
If LLMs Have Human-Like Attributes, Then So Does Age of Empires II #
arXiv / Lobsters
This paper challenges the methodological basis for attributing human-like cognitive attributes to language models by demonstrating that identical evaluation methods produce equivalent results when applied to Age of Empires II’s deterministic game AI. The argument is that tests commonly used to claim LLMs exhibit theory of mind, planning, or strategic reasoning are insufficiently discriminating: they cannot distinguish between systems that genuinely possess these capabilities and systems that merely produce outputs consistent with them. For teams making architectural decisions based on claimed model capabilities, the paper is a useful reminder that benchmark performance does not necessarily reflect the underlying cognitive process it purports to measure.
LLM Research Papers: The 2026 List (January to May) #
Ahead of AI (Sebastian Raschka)
Sebastian Raschka’s curated roundup covers notable LLM research published in the first five months of 2026, providing a structured survey of the field’s trajectory across architecture, training methodology, evaluation, and deployment. For practitioners who cannot track the daily paper firehose, this is a high-signal reference for identifying which research directions have accumulated enough evidence to inform production decisions.
Other #
What to expect from WWDC 2026: Siri’s revamp and Apple Intelligence updates #
TechCrunch
Apple’s WWDC keynote next week is expected to reveal a major Siri overhaul leveraging Google’s Gemini technology to transform it into a conversational assistant capable of multi-step tasks, alongside a standalone Siri app positioned to compete directly with ChatGPT and Claude. Additional anticipated announcements include an AI agent app store for task delegation and enhanced Visual Intelligence powered by Google’s image capabilities. If Apple commits to Gemini as its core AI backend, the partnership reshapes the competitive landscape: Google gains distribution to over a billion devices, while Apple’s developer ecosystem inherits whatever capabilities and limitations Gemini brings.
Threads to Watch #
Federal AI governance is accelerating on multiple fronts simultaneously. The Great American AI Act’s three-year state preemption, the Trump administration’s OpenAI equity discussions, and Sriram Krishnan’s departure from the White House all surfaced within the same 48-hour window. These are not coordinated moves, but they collectively signal that the federal government’s role in AI is expanding faster than institutional capacity to manage it coherently – a bill that preempts state regulation, an equity stake that entangles government with the regulated, and a leadership transition, all at once.
Agent-first development is normalizing beyond early adopters. Jane Street designers replacing Figma with Claude Code, OpenAI publishing case studies on Codex-driven engineering, and the Tokenomics paper quantifying where agent tokens actually go all point to the same shift: AI agents are becoming embedded in professional workflows not as experiments but as default tools. The Tokenomics finding that 59% of costs sit in code review rather than generation suggests that the next wave of optimization targets agent collaboration patterns themselves, not the underlying models.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Stanford HAI — scrape: content_truncated