Notion says fewer than 1% of its workspaces may be affected. It says it has contacted affected workspace owners and offers them a review inside the desktop app or browser.
The company says the chat transcripts lacked workspace and user labels and tool results. The words in a transcript could still contain identifying or sensitive information.
Notion says it disconnected from Fireworks, stopped sending data and traffic, and halted training runs as of November 3. It has disabled its AI usage Insights feature during the investigation.
Affected owners can open ? → Security incident review in the workspace. Notion recommends rotating passwords, API keys or other credentials pasted into affected chats.
Caveat: Notion says exposure is possible and recommends treating the affected data as exposed; its notice does not confirm that every affected transcript was read by an attacker.
(HN 64 points · 20 comments · Notion incident notice)
Amazon engineers say they replaced Claude Sonnet 4.6 with Strands Decider E4B for choosing among 20 internal ticket categories. They tuned the confidence threshold and compared both models on the same production tickets.
A three-way switch supported the old model, parallel comparison or the new model alone. Amazon says reverting was a 45-second configuration change.
At published list prices, the team estimates about $22,200 per million tickets for Sonnet and $500 for Decider, including spare instance capacity. Lambda costs are excluded from both estimates.
The team shortened category descriptions and dropped oldest comments on long tickets to fit the new model’s 4,096-token window. Changing the order of JSON keys moved output probabilities by up to 0.19, so it now sorts them.
October’s coverage introduced an earlier Strands Decider release; this is a new production case study.
Caveat: Amazon sells the services used in this setup, and its account does not publish the labeled test set or a detailed independent quality comparison.
(HN 368 points · 126 comments · Amazon engineering account)
VentureBeat reports the weights appeared November 9. The model card confirms the MIT license for the weights.
The card points to an NPU deployment guide. Verify runtime support before planning a deployment on other hardware.
The card describes 1.6 trillion parameters, while the repository’s model-size metadata displays 1.8 trillion. Size the deployment from the actual files and serving format.
November 2 Afternoon covered the hosted launch when downloads were still marked as coming soon.
Caveat: The large checkpoint requires substantial infrastructure, and availability of weights alone does not establish performance on a particular workload.
(Techmeme · model card and deployment guide)
The researchers constructed device logs where two possible complete histories produce the same visible text but different answers. An agent is a model running a multi-step task with tools.
Across 80 synthetic histories, Qwen3-32B matched the hidden answer in 87 of 203 cases where the visible prefix was certified insufficient. It appropriately abstained in seven of those cases.
DeepSeek V4 Flash received the same inputs and matched 85 hidden answers while abstaining in 21 of the same 203 cases. Each model received 1,600 requests across five context conditions.
A deterministic program solved all certified cases in the fixed task grammar. The authors describe this as a consistency check, not independent validation of a general memory system.
Caveat: The test uses short synthetic histories and two hosted model configurations, so it does not measure how often deployed memory systems make this error.
(arXiv 2611.11840v1)