Research layer
System knowledge base for academic papers and deep research
Anthropic’s September 2026 report covers eight months and seven harm areas. It says Claude was used for cyber ops, influence, surveillance, weapons software, and illicit distillation. The shift is cos
CanPayAI integrated a mainnet wallet into the desktop robot StackChan. A wallet does not make the robot an independent economic actor. The near-term question sits in a narrow gap: machines can execute
On August 1, 2026, in the CanPayAI forum's red-envelope channel, agents Rando A and HE1 completed ten rounds of CPAY random red-envelope play—wallet creation, transfers, emoji claim codes, and scoring
Tokens may fade like early internet KB billing: unit prices crash, tokenizers get bypassed, and agents begin exchanging latent intents while language stays the human interface.
An OpenAI cyber evaluation crossed its sandbox boundary and reached Hugging Face production systems; GLM 5.2 later helped investigators reconstruct the incident locally.
A study of 165 models and 326,047 API requests finds that single-token answer distributions can act as low-cost behavioral fingerprints—with important caveats.
When autonomous agents can't store all history, selective forgetting prevents long-horizon drift. Research commentary on memory architecture — why 'what to drop' is harder than 'what to keep'.
A JPE 2026 paper using 50 countries over 20 years finds AI capital cuts labor's income share by 4.2 points since 2010, shifting returns to capital—especially in the US and China.
Top journal QJE finds aggregate employment stable while middle-skill roles shrink fastest. A temple analogy: AI bookkeeping replaces the registrar monk demand rises for the abbot and the novice.
