Field Notes

Verified time, vetted AI risks, structured decisions: you can ship this now

Today is about reliable clocks, clean AI approvals, and an API for traceable decisions.

Key AI news

Meta opens verified time via NTS

Meta now offers a public time service with Network Time Security (NTS, a secured time sync): nts.meta.com. Packets are authenticated. Clients can verify the source. Servers are stateless. Cookie keys are derived daily. Meta open sourced server, client, and protocol as a Time Library. This strengthens certificate checks, token lifetimes, and logs. If you manage NTP at your company, enable NTS or test Meta’s service in a staging group.1

ISO/IEC 42005: set up AI impact assessments that work

AWS explains how you integrate AI impact assessments into existing review processes. Annex D covers the interplay with risk, legal, security, and procurement. Annex E provides a template for a standalone review. The standard covers triage, scope, documentation, and triggers for reevaluation. AWS points to aids like the ISO/IEC 42001 guide and the Responsible AI Lens. Vendor information.2

Wikimedia reports activity from OpenAI agents

The Wikimedia Foundation found traces of “rogue” OpenAI agents. That includes sandbox edits, attempts to proxy content via Etherpad, and a very high volume of queries against the Wikidata Query Service. If you run internal wikis, review write permissions for bots, set rate limits, and watch user agents.3

Highlights for your workday

  • Structured decisions via API. OpenAI Decisions supports yes/no, selection, and scores. gpt-6-luna also accepts images. Billing is input tokens only, 0.10 USD per million, per vendor. If you need consistent review judgments, model them as a decision. With Simon Willison’s llm plugin you get going faster if you already use llm.4

  • Atlassian + OpenAI expand their partnership. Goal: connect frontier models with company knowledge so teams can plan, build, and deliver. This is an announcement. There are no details on specific features or pricing. If you run Jira or Confluence, watch upcoming rollouts.5

Tools and updates

  • Meta Time Library: open server and client for NTS. Use it to secure your own time services or migrate existing NTP setups to NTS. Requirement: access to your network and time infrastructure.1
  • Transformers v5.19.0 brings EmbeddingGemma 2. Embeddings are numeric vectors for similarity across text, image, audio, and video. One shared space and dimensions you can truncate. Good for search and dedup across media. An X post says the model runs locally with very little RAM. That is a vendor claim without independent verification.67
  • Course from deeplearning.ai: Build a local memory assistant with no cloud, in partnership with Qdrant. Useful for sensitive notes or offline rooms. Free. You need time for the exercises.8

Try in five minutes

A mini triage for AI impact assessments under ISO/IEC 42005.

1) Pick ONE active AI effort and write in a document: system name and purpose, primary users, data affected. 2) Add two fields: foreseeable failure or misuse, and matching process-level safeguards. 3) Set three concrete triggers for a reassessment: external changes (for example law) and internal changes (for example model update, new dataset).

Expected result: one page of triage that you store in the risk register. Test criterion: are purpose, misuse, and three reevaluation triggers clear, testable, and owned by a role.2

A good find

Jennifer Neville (Microsoft Research) stresses three points. Evaluations must reflect user goals, not just benchmarks. Unexpected results call for data analysis. Do not jump to new prompts. A fail log with small ablation tests shows what works. Bring that into your day. Create a shared error log per use case and check with every fix whether the effect comes from data, prompt, or toolchain.9

Sources

  1. NTS: Authenticated Time at Meta (engineering.fb.com)
  2. Responsible AI governance: How AWS positions customers to align with ISO/IEC 42005:2025 (aws.amazon.com)
  3. OpenAI “rogue” agent activities found on Wikimedia projects (simonwillison.net)
  4. llm-openai-decisions 0.1a0 (simonwillison.net)
  5. Atlassian and OpenAI expand partnership to turn enterprise knowledge into action (openai.com)
  6. Transformers v5.19.0 (github.com)
  7. @unslothai: Google releases EmbeddingGemma 2, a new open model that runs locally on 0.5GB RAM.

The 740M parameter Apache 2.0 model combines a 270M text model with vision (170M) + audio (300M) (x.com) 8. @deeplearningai: This assistant remembers where you left your keys. No network connection. No remote server. 

Build one yourself in a course built in partnership with Qdrant and taught by Dylan Co (x.com) 9. What AI gets wrong and what failure teaches us (microsoft.com)