Field Notes

Steer your Facebook feed, tune Grok agents, local RTX PCs: you can use this now

Three things for your week: fine-tune recommendations, make agents reliable, roll out local AI cleanly.

Key AI news

Facebook lets you steer the feed with text

Meta is rolling out a tool where you tell Feed and Reels in free text what you want to see. Meta says your inputs take effect in the same session. You reach it via the three-dots menu under “Your algorithm”. The rollout is gradual. If you do not see it yet, it is not enabled for you. Useful for you: you can sharpen research and inspiration in your own feed. The effectiveness and details come from the vendor.1

Grok 4.7 on Bedrock: more steps, steerable thinking

Grok 4.7 is now on Amazon Bedrock. New is a slider for “reasoning effort”. That controls how much thinking happens before the answer. AWS shows three agent tasks with 67 correct out of 72 runs. At maximum effort 18 out of 18. At low effort 15 out of 18 with far fewer tokens. These are measurements from the blog post. SpaceXAI also cites a 500K context length and better long-task performance. Those are vendor claims. For you that means: you can trade off cost and accuracy per request and align monitoring to it.2

RTX Spark: launch local agents on your PC this week

NVIDIA and partners are rolling out software bundles for RTX Spark laptops ahead of market launch on Friday, October 16. Included: agent apps with secure sandboxing via OpenShell and Microsoft Execution Containers, a local 74B MoE model Nemotron 3.5 Bolt on October 15, and hybrid routing between local and cloud. For creatives there are accelerated workflows in Photoshop and Premiere. NVIDIA says up to 2x. “Coming soon”. This is an announcement, partly with dates in the next few days. Practical: you can introduce local agents with clear security boundaries and reduce cloud use on purpose.3

DeepMind and Singapore: fraud protection on the device

DeepMind and GovTech Singapore show an on-device detection system against chat scams. Lightweight Gemma models step into private chats without cloud exfiltration and warn about patterns from conversation histories. Status: announced and described as a case study. If you plan customer protection in apps, note the pattern. On-device detection. Clear data boundaries. Early user warnings.4

Highlights for your workday

  • Security queries in plain language: Orca Security built an agent on Amazon Bedrock that turns text questions into Sonar filters. A multi-stage setup with pattern checks, guardrails, embeddings, and RAG. RAG is a technique that pulls knowledge from documents in a targeted way. Orca says it cut costs by 90 percent, input tokens by 93 percent, and latency by 75 percent. Use the pattern for internal search or compliance queries with strict schema checks. Numbers are vendor claims.5
  • Distribute private Bedrock prices centrally: AWS shows how you subscribe negotiated Marketplace prices for models once and pass them as “managed entitlements” to all accounts. Either as a one-off script or as a CDK automation with events and an allowlist. That reduces manual activations and secures the negotiated rate.6
  • Rate Gemini 4 Argon now: Argon is available via the Gemini API and, according to Simon Willison, currently at 2$/M input tokens and 10$/M output tokens. A doubling is announced from December 1. That creates a short cost window for benchmarks. The prices and efficiency comparisons are not from independent tests.78

Tools and updates

  • Gemini 4 Argon in llm-gemini 0.35: the plugin supports Argon directly. Willison says the current cost profile looks competitive, but it changes on December 1. Requirement: Gemini API access, sometimes a subscription.7
  • Higgsfield: longer AI videos: the integration uses GPT-6 Astra for runtimes over ten minutes, plus sharper images and a voice track. Useful for marketing tests with series formats. Access and prices are in the vendor text. Details are sparse.9
  • LangSmith Engine v2: red-teaming, more issue types, automatic fix tests for agents. Helps make agents more robust in staging. Requires the LangSmith stack.10
  • Lift Spark UDFs onto GPU: NVIDIA shows how agents can help turn slow UDFs into GPU counterparts. Relevant when big data jobs stall and you already have a GPU environment.11

Try it in five minutes

1) Open Facebook on your phone. For three to five posts tap the three-dots menu and pick “Your algorithm”. Add two to three topics you need today, like “B2B SaaS” or “Reels statistics”.1 2) In ten posts mark five as “Interested” and five as “Not interested”. Add a short free-text note about what exactly fits or bothers you.1 3) Scroll for five more minutes. Expected result: more hits for your topics, fewer distractors. Checkpoint: count the hit rate in the next 20 posts before and after the adjustment.1

A good find

Sources

  1. Your Algorithm, Your Words: New Ways to Shape What You See on Facebook (about.fb.com)
  2. Build long-horizon agents with Grok 4.7 on Amazon Bedrock (aws.amazon.com)
  3. All Systems Go: Global PC App Ecosystem Ready for NVIDIA RTX Spark (blogs.nvidia.com)
  4. Google DeepMind and Singapore Government announce new AI tool to protect citizens against scams (deepmind.google)
  5. Natural language security queries at Orca Security using Amazon Bedrock (aws.amazon.com)
  6. Manage private pricing offers for Amazon Bedrock with managed entitlements (aws.amazon.com)
  7. llm-gemini 0.35 (simonwillison.net)
  8. Gemini 4 Argon: Built to solve your hardest problems (deepmind.google)
  9. Higgsfield takes AI video from 1 minute to over 10 (openai.com)
  10. @langchain: ✅ Red team your agents

✅ Detect more issue types

✅ Automatically test proposed fixes

@bentannyhill on everything new in LangSmith Engine v2 and how we approach agent improvement (x.com) 11. How to Rewrite Apache Spark UDFs for GPU Acceleration with AI Agents (developer.nvidia.com)