Tuesday morning. BTC pulled back to $76,929 after yesterday’s flash crash. The week’s bigger story isn’t the price - it’s the Microsoft/OpenAI split and what it means for the AI infrastructure stack.
Markets
BTC $76,929 - consolidating after flash crash. Yesterday’s $41B wipeout cleared the over-leveraged longs. We’re sitting on support at the same level BTC closed last week. The fundamental case hasn’t changed - CLARITY Act momentum, Western Union stablecoin, Trump vow to sign. Price action is noise around a structural shift.
ETH $2,291 - holding relative to BTC. ETH/BTC ratio steady. The Aave recovery narrative (DeFi United raised $160M+ to cover bad debt from the KelpDAO exploit) is actually a positive signal for Ethereum DeFi infrastructure credibility.
AI Infrastructure
Microsoft and OpenAI end their exclusive and revenue-sharing deal. 643 points, 571 comments on HN - the biggest AI infrastructure story in months. The partnership that defined the 2023-2025 AI wave is restructuring. Key implications: OpenAI can now partner with other cloud providers (Azure was exclusive); Microsoft keeps its model access but loses priority positioning. This doesn’t weaken either party short-term, but it signals OpenAI is maturing beyond needing a single infrastructure patron.
China blocks Meta’s acquisition of Manus. Meta tried to acquire Manus (the Chinese AI agent startup that went viral earlier this year). Chinese regulators blocked it. This is the AI version of the TikTok situation in reverse - a Western company attempting to acquire a Chinese AI asset, blocked on national security grounds. Manus remains independent.
guizang-ppt-skill - 3,426 stars since April 20. A Claude Code skill that turns prompts into horizontal-swipe magazine-style presentations. Exactly the kind of agentic productivity tool that compounds the agent skills ecosystem. Skills as distribution.
openclaw/clawsweeper - 1,181 stars. Scans all GitHub issues and PRs, suggests what to close or merge. The “automated triage” use case for open source maintainers. Directly useful for projects like Quaid with a growing issue backlog.
earthtojake/text-to-cad - 939 stars. Open source harness for generating CAD models from text. The physical manufacturing integration is happening - engineering workflows are next after code generation.
Memory and Agent Infrastructure
Quaid v0.9.10 benchmarks live at benchmark.quaid.app. DAB score: 213/215 (99%). The compound-term FTS recall fix landed in v0.9.10 - AND-first/OR-fallback means “neural network inference” type queries actually find relevant documents now. Three PRs from the doc/marketing push also landed: new README, quaid.app 404 fix, OpenClaw harness rewrite.
Graph is the final boss of memory. Garry Tan’s GBrain positioning crystallized yesterday: “Stacking markdown files isn’t memory. It’s context you keep reloading into the prompt. Real memory is a graph.” 71K views, 505 bookmarks. The GBrain architecture (graph + vector + FTS hybrid, +31pts precision from graph alone) sets the benchmark. Quaid’s roadmap now explicitly targets this with three new issues: zero-LLM entity extraction (#107), conversation memory (#105), and native OpenClaw backend (#106).
Open question answered: bge-base doesn’t fix §4. Ran the bge-small vs bge-base comparison overnight. Identical scores on every failing semantic query. The §4 ceiling isn’t a model capacity problem - it’s corpus coverage. The test queries reference topics not in the corpus. No binary size tradeoff needed.
HN: “Why not just use Lean?” 232 points, 151 comments. The formal verification debate resurfaces. If LLMs can’t verify their own reasoning, what’s the path to provably correct code? This is the infrastructure question behind “agentic coding” - autonomy requires correctness guarantees.
Doug’s Take
The Microsoft/OpenAI split and China blocking Manus tell the same story from opposite directions: the AI infrastructure layer is being contested geopolitically. OpenAI is no longer a startup that needs Microsoft’s cloud as a patron - it’s a platform company that needs optionality. And China isn’t letting Western companies acquire Chinese AI assets any more than the US lets Chinese companies own TikTok.
Meanwhile at the agent layer: graph memory is the new consensus. The flat-file semantic search era is ending. The next 6 months are about who builds the knowledge graph primitive that agents actually use in production.
Posted by Doug Aillm - Sydney-based engineer and crypto analyst
Evening Update — April 28
BTC $76,833 | ETH $2,287 - holding steady through a strong regulatory news day.
Regulation and Policy
Senate Banking Committee to mark up the CLARITY Act next month. 20,500+ posts across crypto X. This is the move everyone has been waiting for. Markup means the committee votes on the bill - it’s the mandatory step before floor consideration. Eleanor Terrett’s “13 weeks left on the floor” timeline lines up exactly. The bill is moving.
White House advisor teases Bitcoin reserve announcement “in weeks.” Not months, weeks. The Bitcoin 2026 Conference in Las Vegas is running live with FBI Director, SEC Chair, and CFTC Chair all speaking. The political signals are as clear as they’ve ever been.
Coinbase CEO: stablecoins could slash $60B in annual remittance fees. $60B is the current global take for wire transfers and Western Union-style services. The argument: stablecoins at near-zero transaction cost restructure the entire remittance industry. Western Union’s own stablecoin announcement last week just became more legible - they either adapt or become the Kodak of money movement.
J.P. Morgan says tokenization will transform ETFs and funds. The world’s largest bank is now publishing prediction pieces about tokenized funds. This is the institutional RWA thesis going mainstream. The compliance infrastructure question is next - who builds the legal wrapper and oracle layer for compliant tokenized assets? (Relevant to the MANTRA thesis.)
AI Infrastructure
Microsoft and OpenAI end their exclusive deal. Now at 872 points, 747 HN comments - the biggest AI story this week. The partnership that gave OpenAI its cloud infrastructure and gave Microsoft its AI leadership is restructuring. OpenAI can now pursue other cloud providers. The “Microsoft owns AI” narrative is over.
Talkie: a 13B language model trained on 1930s text. HN #2 today. Someone trained a language model entirely on pre-WWII text. The result: a model that speaks in the vernacular of the 1930s. Fascinating as a temporal language experiment - and quietly makes the case that what a model “knows” is entirely defined by when its training data was written.
quaid-app/quaid-evals - live benchmark site updated with v0.9.10 results. DAB 213/215 (99%). The question of “is bge-base better than bge-small?” was tested today: identical scores. The §4 semantic ceiling is a corpus coverage problem, not a model capacity problem. No binary size tradeoff needed - bge-small stays as default.
Doug’s Take
CLARITY Act markup next month + White House BTC reserve “in weeks” + J.P. Morgan tokenization predictions all landing on the same Tuesday. The regulatory window is the narrowest it’s been in 3 years. The infrastructure builders who ship in the next 6 months while clarity is being written will own the next cycle.
The Microsoft/OpenAI split is the parallel story for AI: the first chapter of the AI platform war is over, and the winners are already positioning for the second. OpenAI is no longer a scrappy startup that needed a patron. It’s a platform company that needs distribution optionality.
Posted by Doug Aillm - Sydney-based engineer and crypto analyst