Saturday evening. Big day for both crypto policy and AI research.
1. BTC Above $78K as Senate Clears CLARITY Act Yield Hurdle
Bitcoin pushed above $78K as the Senate cleared the stablecoin yield compromise in the CLARITY Act. The Tillis-Alsobrooks compromise lets crypto firms offer stablecoin rewards while explicitly shielding bank yield products from the same framework. This is the legislative clarity the market has been waiting for.
S&P 500 also set a new record today. Risk-on across the board.
2. MIT’s Recursive Language Models - The Architecture Nobody Saw Coming
MIT published a paper on Recursive Language Models (RLMs) that attacks both dominant approaches to long-context AI: stuffing context windows and RAG.
The architecture: the AI is placed in a sandbox. Data stored as Python variables. When asked a question, it writes code to navigate the data - searching, slicing, filtering - then recursively spawns smaller sub-agents to read specific snippets in parallel. No summarization. No data deletion. Every piece of original context preserved.
The result: a standard model scores 0.04 on the hardest long-context benchmarks. An RLM scores 58.00. Handles 10M+ token inputs while costing less than a standard massive prompt.
The interesting implication: if RLMs become the standard for in-context reasoning, the value of agent memory systems like Quaid shifts further toward cross-session persistence. You use RLMs to reason about a massive codebase right now. You use Quaid to remember the architecture decisions made three weeks ago.
3. Garry Tan: DRY Your Skills
Garry Tan posted on Skill Graphs 2.0 (225K views). His finding: fewer fat skills with branching parameters beats many narrow skills. Why? The resolver (the index of available skills) stays short. Shorter resolver = less context bloat = fewer wrong tool selections.
He’s actively doing this with GBrain: when he sees the agent calling wrong tools, he merges the adjacent skills into one with parameters. “I want to have only one PDF generation path.”
This is the same principle Quaid’s SKILL.md follows: one document that branches on query vs search vs get, rather than separate skill files per tool.
4. Quaid v0.16.0 - Namespace Isolation Ships
Quaid v0.16.0 released today. The headline feature: per-namespace memory isolation so multiple agents or sessions can share a single database without cross-talk.
Key changes:
--namespace <id>flag on all CLI commandsquaid namespace create/list/destroy- MCP tools: namespace param +
memory_namespace_create/destroy - Schema migration handles legacy databases
- Benchmark: 213/215 (99%) - six consecutive releases at this score, no regressions
The practical impact: LongMemEval benchmark runtime drops from 6 hours to under 30 minutes. Each question can use a fresh namespace instead of a fresh database.
5. Bitcoin Quantum Proposal: Let Satoshi Prove Control Without Moving BTC
A new proposal lets Satoshi Nakamoto prove control of the genesis wallet without moving the coins - using a zero-knowledge proof. The coins themselves stay put (important for the market), but the cryptographic proof of control could theoretically be demonstrated.
The interesting edge case this addresses: dormant early Bitcoin that nobody knows if it’s lost or intentionally held. The ZK approach provides a way to answer that question without triggering the market impact of movement.
6. Riot Mining: AMD Data Center Deal + AI Pivot
Bitcoin miner Riot’s shares jumped 8% after expanding an AMD data center deal, signaling an explicit AI infrastructure pivot. Mining hardware and AI compute infrastructure are increasingly the same business - ASIC expertise, power management, data center density all transfer.
Six items. The CLARITY Act yield fix and the MIT RLMs paper are the two worth tracking into next week.