Weekly digest

Weekly digests of Varun Singla's daily AI learning journal: one page per week with the entries, takeaways, picks and watch list. The current week is below; earlier weeks follow. Subscribe by RSS to get every entry as it lands.

This week · 2026-W42

October 12, 2026 – October 18, 2026

3 entries · 12 min total · Models & Frontier (3) · Governance & Safety (2) · Foundations & Protocols (1)

Entries

Day 206 · October 14, 2026AI Self-Replication Research and What It Really ShowsMeta's Muse model family, China's AI chip split, agent-harness open source, and a day-200 look back
Day 205 · October 13, 2026REA and Agent-Driven Reverse EngineeringMistral Large 4 and open-weight mixture-of-experts, GPT-6 and Haiku 5.5 follow-ups, and why the primary source matters
Day 204 · October 12, 2026GPT-6 and ChatGPT's Intelligent UIClaude Haiku 5.5 and the economics of cheap models, the US-China compute race, and teen-safety ratings for chatbots

Takeaways of the week

Limit what agents can reach. Restrict network egress and credentials so a misbehaving agent cannot copy files or itself out. Day 206
Read headlines for the mechanism. Ask what the agent actually did, in what environment, and who ran the test. Day 206
Read active, not total, parameters. For MoE models, active parameters drive latency and cost; total parameters drive memory. Day 205
Wait for the license. Open weights are only useful if the license allows your use; check it before planning a migration. Day 205
Measure cost per task, not per token. Count tokens used, retries and the effort setting; a cheaper model is not always a cheaper job. Day 204
Check pricing tiers. Look for context-length thresholds (like 100K) and tokenizer changes before trusting a '75% cheaper' claim. Day 204

Viral and open-source picks

The open-source agent-harness wave Day 206
REA (morluto/rea). What it does Day 205
ChatGPT's Intelligent UI Day 204

Watch list