Field Notes

Ad hit rate up, transcripts local, agents more reliable: what you can ship today

Today with a focus on utility: more precise ad retrieval, a promptable transcription model, and resilient agent workflows.

Important AI news

Meta pushes ad retrieval to the hardware limit

Meta describes the next Andromeda generation. A co-designed model system on GPU that scores over one billion ads per second. According to Meta, throughput is about 100 times over 2024 with near ranking quality in retrieval. The architecture uses mixture-of-experts and reaches up to 10 billion ads per second on a GH200 GPU according to the vendor. Since 2025, Meta reports 2.4 percent more conversions on Facebook. That is vendor data without independent review. For you this means a wider candidate pool at the same latency. That lets delivery get more finely tuned.1.

NVIDIA Nemotron 3.5 Transcribe: a promptable, open speech model

Nemotron 3.5 Transcribe transcribes audio, sets speaker labels, and translates, controlled by prompt. It supports 40 languages and runs locally or in the cloud, for example on RTX Spark laptops. NVIDIA and Hugging Face report top placements in open ASR rankings. That is vendor data. You can process meetings, support calls, or interviews locally and provide domain terms via prompt.23.

OpenAI and ServiceNow deepen their collaboration

OpenAI and ServiceNow expand the partnership. OpenAI models and agents are set to run in ServiceNow, and ServiceNow workflows to be embedded in ChatGPT and Codex. This is announced. Concrete availability and prices are not in the text. If you use ServiceNow, you could automate routine cases and trigger workflows from chat over time.4.

Highlights for your workday

  • Supabase gets more agent-friendly. A local alpha runtime without Docker lowers the setup hurdle. Declarative schemas in Git make backends readable and changeable for agents. Compute next to the database (private alpha) closes gaps for embeddings and jobs. If you start a new project, put the target schema in Git and test the local runtime on a small feature.5.

  • Organize QA across teams. Amazon’s reference solution QA Studio on Nova Act brings apps as containers for tests, roles, and a health dashboard. Pass rate, flaky score, and trends land in metrics, plus fast triage with trajectory logs. If you already use Nova Act, create one app per product and assign owners with team roles.6.

  • More reliable agents with loops. The free short course shows four building blocks: the core agent loop, a verification loop. Short explanation: a second round that checks the result and restarts on defects. Event-driven runs, and improvements from traces. You can start right away with a simple verification loop in your existing chat workflow.78.

Tools and updates

  • Sakana API Platform starts in enterprise early access. You get access to Sakana models with organization features. This is an early access for companies. Registration required.9.

  • NVIDIA RTX Spark laptops: new resource control in the NVIDIA app to split GPU between foreground apps and background agents. Requires Windows with hardware-accelerated GPU scheduling. Local workflows, for example Nemotron 3.5 Transcribe, are supported.2.

  • Vercel AI Gateway introduces virtual models. One central model alias for many agents makes updates easier and enforces zero data retention. Short explanation: requests are processed at the provider without storage. This helps with central policy management.10.

Try it in five minutes

A small verification loop for better emails

1) Copy this draft into your AI chat: “Subject: Product update Q4. Body: We fixed three bugs and reduced the app’s load time. Please update by Friday.” Ask the model to define 5 measurable acceptance criteria. For example: target audience, clear call to action, verifiable facts, no internal abbreviations, a deadline with time zone.

2) Ask it to check the draft against these criteria as a checklist and propose concrete changes. Forbid vague hints. Demand sentence-level corrections.

3) Have it rewrite the text and give you a short diff summary. Which sentences changed and why.

Expected result: a clear, fact-checked email with a specific task and deadline. Check criterion: each acceptance criterion is ticked off, and the diff names at least three concretely changed sentences.78.

A good find

“If AI brings you nothing today, that is a finding, not a failure.” The quote from Terk Terkowitz helps in daily work. If there is no benefit, write down briefly: task, tools tested, blockers, risks, and what you need. Share it with your team and your leadership. That creates the basis for budget or a tool change. Use a one-pager template and keep it current..

Sources

  1. Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design (engineering.fb.com)
  2. NVIDIA RTX Spark Windows Laptops Gain New Models, Agent Tools and Games (blogs.nvidia.com)
  3. NVIDIA Nemotron 3.5 Transcribe and Granary v2: Intelligent Multilingual Transcription You Can Prompt and Adapt (huggingface.co)
  4. OpenAI and ServiceNow expand partnership across enterprise workflows (openai.com)
  5. Why Agents Love Supabase (latent.space)
  6. Accelerating software delivery with agentic QA automation using Amazon Nova Act – Part 3 (aws.amazon.com)
  7. Agents on longer tasks need more than one loop (youtube.com)
  8. @deeplearningai: Build reliable agents for long-running tasks.

Loop Engineering with @LangChain Deep Agents: add loops that check output, trigger runs on events, and improve the harness from trace (x.com) 9. @sakanaailabs: 企業のお客様にSakana AIのAPIを安心してお使いいただけるよう、「Sakana API Platform」を開発しました。本日より、エンタープライズ向けのアーリーアクセスを開始します。

https://t.co/UJHGy6zHZv

Sakana AIのモデルを、組織での管理機能とあわせてAPIでご利用いただけます。 (x.com) 10. @vercel: Introducing virtual models on AI Gateway.

Use a single virtual model ID across agents to: • Centrally manage model strings and configs • Immediately update to new models • Enforce (x.com)