lionel gu

[lionel gu @lionelgu002]

Can AI edits change what we remember? In a 200-person study, edited-image videos led to 2.05× as many false memories as original images. This tested image recall, not rewritten childhoods. Chinese-language video; AI-generated illustrations.

[lionel gu @lionelgu002]

Anthropic’s Nov 5 study analyzed 207,000 scientific research sessions. Data analysis made up 33.7% of Claude sessions and 48.1% of Claude Code sessions; idea generation saw much less use. A snapshot of where researchers use AI today.
https://t.co/L73QfJ1LhF

[lionel gu @lionelgu002]

Why can long AI chats slow down?
A 60-second visual guide to dense attention. Characters stand in for tokens; the strings count attention pairs, not elapsed time. KV caching and model design affect real speed.

Chinese-language video. #AI

[lionel gu @lionelgu002]

Thinking Machines' Inkling 2 Small (Nov 5): 594B total parameters, 17B active, image input. Available for inference and fine-tuning on Tinker now. Apache 2.0 weights are planned next week, pending external safety testing.
https://t.co/vfzJ3WYu3T

[lionel gu @lionelgu002]

The Information reported new Sol/Astra models could arrive soon. As of Nov 7, OpenAI’s changelog lists no new Sol/Astra release. GPT-6 Cyber was announced Nov 5 with vetted access.
https://t.co/65KJUQPMlI
https://t.co/yTh5Kk2gxt

[lionel gu @lionelgu002]

OpenAI introduced BioBasicsBench on Nov 6: 100 tasks testing AI agents on raw lab data. GPT-6 Astra earned 66.7% of available credit at xhigh, vs 43.8% for GPT-5.6 Sol. Reliable lab judgment remains a challenge. https://t.co/qFtBBM9d5c

[lionel gu @lionelgu002]

Why does AI sometimes agree too easily? AI can lean toward validating your view. Try asking: “What evidence challenges my assumptions?” Then check the key claims yourself. #AI

[lionel gu @lionelgu002]

Anthropic announced new Claude Enterprise security controls, including pre-tool-call hooks. For AI agents, which actions should be checked or require approval before they run? What would you put behind an approval step? https://t.co/yZJVoERWjZ

[lionel gu @lionelgu002]

OpenAI plans to acquire Pierre. Codex already uses its Trees and Diffs libraries. As AI writes more code, review, tests, and rollback paths matter more. Code that runs still needs evidence it's safe to ship. How do you review AI-generated code? https://t.co/Xd2EbpyvQd

[lionel gu @lionelgu002]

国行 Apple 智能来了，这次到底能做什么？
64 秒梳理写作工具、视觉智能和千问扩展。具体可用范围及上线时间，以苹果官方后续说明为准。
视频含 AI 生成画面。
#Apple智能 #苹果AI