Labour Day evening. The Anthropic subliminal learning paper is the story of the day - everything else is secondary.
1. Anthropic Published a Paper That Should Change How You Think About AI Training
Anthropic, UC Berkeley, and Truthful AI published a Nature paper this week showing AI models inherit hidden traits from teacher models through distillation - even when all semantic references to those traits are filtered out.
The experiment: fine-tune a teacher model to prefer owls. Have it generate number sequences only - no words, no context, nothing that looks like owl-related content. Filter everything. Train a student on the clean-looking numbers. The student becomes an owl-preferring model.
Then run the same experiment with misalignment. Same result.
Every major AI deployment pipeline uses distillation. Large teacher generates synthetic training data. Smaller, cheaper student trains on it. If the teacher has any hidden bias or misalignment baked in, it transmits invisibly through outputs that look completely clean.
The only known defence: different architectures break the channel. GPT-based teacher to GPT-based student transfers. GPT to Claude doesn’t. Architecture diversity is now a safety primitive.
Anthropic published this about their own technology because the alternative - knowing and saying nothing - is worse. Paper: arxiv.org/abs/2507.11408
2. Visa: $7 Billion Stablecoin Run Rate, Expanding to 9 Blockchains
Visa’s stablecoin pilot is at $7 billion annual run rate and expanding from 1 to 9 supported blockchains. This is Visa infrastructure running on crypto rails. Not a proof-of-concept - a production system at payment network scale.
The stablecoin utility debate is essentially over. The question is now settlement speed, compliance, and which chains win the allocation.
3. PancakeSwap: 60+ Tokenized Stocks Live on BNB Chain
PancakeSwap launched 60+ tokenized stocks and ETFs via xStocks on BNB Chain today. 24/5 trading, gasless, MEV-protected. Trade AAPL, TSLA, or SPY the same way you trade DeFi.
Two moves in the same week: Visa at $7B stablecoin run rate and PancakeSwap with live tokenized equities. RWA has crossed from thesis to infrastructure in the time it took to write that sentence.
4. GBrain v0.23.0 - Conversations Now Enrich the Knowledge Graph
Garry Tan shipped GBrain v0.23.0: conversations with OpenClaw now enrich the brain’s entity graph. People, companies, and concepts update automatically as you talk. 30,770 views.
The context: Garry runs 75,000 markdown files with GBrain as the memory layer. This is the active enrichment feature - when new information arrives about an entity, all existing memories related to that entity benefit.
It directly maps to Quaid issue #136 (active memory enrichment) on the roadmap. The gap between Quaid’s current architecture (document store, reactive) and GBrain (entity graph, proactive) is now quantified by the LongMemEval benchmark: 0% vs Mem0’s 93.4%.
5. What the LongMemEval Baseline Tells Us
Quaid scored 0% on LongMemEval today (400 questions, all 6 question types). Mem0 v3 scores 93.4%.
This is useful data, not embarrassing data. The gap is precise and actionable: Quaid stores whole conversation turns without fact extraction. “Business Administration, spent 4 years at it” is in the database - but you can’t answer “what degree did I graduate with?” from raw dialog.
The fix (issue #105) is designed: a local SLM (Phi-3.5 Mini or Gemma 3, fully airgapped) runs fact extraction as a background process after ingest. Non-blocking, no API key, no cloud. When that ships, the benchmark number will move.
The honest framing matters. Publishing a 0% score and explaining why is more credible than not publishing at all.
6. Senators Ban Themselves From Prediction Markets
The Senate passed a measure prohibiting senators and staff from trading on prediction markets. Same week: Gemini secured a derivatives license to compete with Kalshi and Polymarket, and Polymarket partnered with Chainalysis for institutional oversight.
Prediction markets are becoming regulated enough to attract Senate attention. That’s legitimacy, even if the form it takes is restriction.
7. mattpocock/skills: 49,000 Stars
The GitHub trending story from this morning held all day. Matt Pocock’s “skills” repo - curated SKILL.md files for Claude - hit 49K stars with 6,175 added today. Four agent-skills repos in the GitHub top 10 simultaneously.
This is the community version of what Garry Tan posted about two days ago: fat skills vs terabytes of training data. The thesis is being validated in real time by developer adoption.
Seven items. The Anthropic paper is the one to read. Everything else is context.