Arete-27B is a Qwen3.8-27B model post-trained on 52.8M tokens for 2 epochs without RL, reducing propaganda rate from 22% to 1% on an 803-prompt eval. The model’s Hermes Bench score changed marginally to 65.77% from 65.60%.
Key points
Fine-tuned on 52.8M tokens, 2 epochs, no RL.
Propaganda rate fell from 22% to 1% on 803-prompt evaluation.
Hermes Bench score: 65.77% vs 65.60% prior.
Sources
Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks
Making Agents More Consistent: Skills Should Form Habits for Repeat Tasks introduces skill habit formation to improve consistency. A habit-formed variant reproduced all 456 repeated dispatches and cut token use by 14%-56%.
Key points
38% to 74% of answers disagreed across 42 tasks run three times per model.
Four gates admit candidate skills; the central gate tests execution trace tolerance against run-to-run variability.
Sources
Tom Lee says AI infrastructure spend $30T, total build-out $35T
Tom Lee claims AI infrastructure spending is about $30 trillion, close to the $35 trillion total build-out of internet, mobile and handsets in 2026 dollars. He adds that Cisco fell 50% twice before it peaked.
Key points
AI infrastructure spend is about $30T.
Total build-out of internet, mobile and handsets in 2026 dollars was about $35T.
Cisco fell –50% twice before it peaked.
Sources
- Original post
- help.x.com - Linked in the post
Fundstrat Tom Lee AI infrastructure spend equals internet build-out $30T vs $35T
Fundstrat Tom Lee compares AI infrastructure spend to historic internet and mobile build-out, saying AI spend is about $30T versus $35T for the earlier era. He cites Cisco’s 50% drops twice before its 2000 peak and likens today’s market to 1997.
Key points
AI infrastructure spend is about $30T in 2026 dollars.
Total internet and mobile build-out was about $35T in 2026 dollars.
Cisco fell 50% twice between 1993 and 2000 before peaking in 2000.
Wireless stocks peaked 10 years into a 25-year exponential growth cycle.
Sources
RYFT 3D Agentic Game Creation beta rollout announced
RYFT announces the rollout of 3D Agentic Game Creation in beta. Selected beta testers will receive access shortly. The upgrade is positioned as a major change for building worlds and tokens.
Key points
3D Agentic Game Creation is actively being rolled out.
Selected beta testers will receive access shortly.
Sources
- Original post
- ryft.fun - Linked in the post
NVIDIA TEE inference on H200, Tier 3+ partner
NVIDIA TEE encrypts GPU memory and a Tier 3+ certified partner offers H200 nodes with TEE enabled for confidential inference. The post lists checks before trusting a TEE-enabled node.
Key points
TEE encrypts the GPU's memory so the operator can't read model weights or inputs while the job runs.
The node is an H200 with NVIDIA TEE enabled from a Tier 3+ certified partner.
Sources
Moving sofa problem – area bounds listed by mathematicians
Moving sofa problem – area bounds listed by mathematicians: square area 1, semicircle π/2≈1.5708, Hammersley 1968 π/2+2/π≈2.2074.
Key points
Square area 1
Hammersley (1968) area π/2 + 2/π ≈ 2.2074
