Field Notes

LLMs 10–20x faster in production, API credit at OpenAI, and AI tests before merge

Today is about speed, cost, and reliability. Plus two small practice levers you can use right now.

Important AI news

10–20x faster: Booking.com combines Knowledge Distillation and Medusa

Amazon Science reports on a pipeline with Booking.com. A small model reaches near the quality of a 70B teacher, yet runs 10–20x faster and at less than a third of the cost. Knowledge Distillation explained. A large teacher model creates training targets for a smaller student model. Medusa explained. Extra heads predict multiple tokens in parallel, which speeds up output. They measured 6x speed-up from the smaller model plus 2–3.6x from Medusa, with minimal quality loss according to the provider. The team used TinyLlama 1.1B as the student and Llama-3.1 70B as the teacher, with domain-specific finetuning. Numbers come from the provider and have not been independently verified.1

Meta deploys JiTTests in production

Meta brought Just-in-Time tests into the development infrastructure. The AI generates targeted missing tests for risky diffs through two paths: intent-aware and “dodgy diff”. An LLM-as-Judge ensemble scores findings, a rule-based filter (RubFake) cuts false alarms, and humans stay in the review loop. Goal: show real bug differences before merge instead of only commenting on code. Meta describes scaling, a true-positive focus, and lower review load. Results and methods are provider claims.2

OpenAI: API credit for $100/$200/$500 subscriptions

OpenAI gives paying chat subscriptions monthly API credits in the same face value. The credits expire and can be assigned to an API organization. Anthropic had introduced a similar model earlier. If you already pay a Pro subscription at that level, you cut test and prototype costs.3

Meta + CNIB: AI glasses for blind and low-vision people

Meta supports the Canadian CNIB with money, training material, and device donations. Four focus areas: community feedback, peer training, input for product development, and accessible learning resources. In addition, AI glasses go into the CNIB program “Come to Work”. Claimed benefits by Meta: reading aloud, object recognition, description of surroundings.4

Highlights for your workday

  • Sharpen your resume with Gemini. You turn raw notes into crisp, ATS-friendly bullets. This raises your chance to pass automated filter systems. Start with an old CV and revise just one section.5

  • Intent-based UI tests as a service. Epsilon builds tests in plain language and runs them with Amazon Nova Act and the Bedrock AgentCore Browser. You get reliable steps without fragile selectors and without your own browser infra. Requirement: AWS access and Bedrock integration.6

  • Agents can now click in the browser themselves. LangChains Managed Deep Agents control websites without an API, test UIs end to end, and can hand the window to you briefly for login. This helps with bookings, forms, and urgent QA checks.7

Tools and updates

  • Simantic: firmware simulation for AI agents. Test MCU firmware without hardware, with virtual devboards and CI integration. Useful for embedded teams in automotive, robotics, and space. Requirement: supported chips like STM32F-/H7, ESP32-S3/C3, RP2040.8
  • Mistral Large 4 “Le Chonk” on Hugging Face. Multimodal large model as a practical entry via the HF platform. Suited for prototypes that combine text and media.9
  • AnySpeechRecognizer and AnyDecisionModel. Open models for speech and decision logic on Apple platforms. For local dictation, transcripts, or simple automations.10

Try it in five minutes

Revise resume bullets with Gemini

1) Open Gemini and switch to the full-screen editor. Paste these raw notes: “Introduced CRM, 3 months, team 5, sales happier. Did newsletter, more clicks. Sorted support emails, replied faster.”11 2) Prompt: “Write 5 ATS-friendly bullet points in Task-Action-Result format. Use strong verbs, concrete numbers or time spans. Add 3 relevant keywords from marketing/CRM.”5 3) Read the bullets aloud and cut filler words. Add real metrics if known.

Expected result: 5 short bullets with clear action and measurable outcome. Check criterion: each bullet has a strong verb, an effect, and at least one domain keyword.511

A good find

“If AI does not help you today, that is a finding, not a failure.” Amazon director Terk Terkowitz advises: document the reasons, share them in the team, and look for solutions. This builds judgment instead of automating blindly. Concrete next step: write a one-pager with three tasks where AI hinders or helps, and attach evidence.12

Sources

  1. Speeding up language models for travel applications (amazon.science)
  2. Catching Bugs Before They Land: JiTTesting’s Human-on-the-Loop Approach at Full Scale (engineering.fb.com)
  3. API credits benefit for eligible ChatGPT subscriptions (simonwillison.net)
  4. The Future Is for Everyone: Meta and CNIB Partner to Support Canadians Who Are Blind, Deafblind or Have Low Vision (about.fb.com)
  5. How to use Gemini to update your resume (blog.google)
  6. Building intent-driven UI testing with Amazon Nova Act (aws.amazon.com)
  7. @langchain: 🌐 Managed Deep Agents can now use a browser

Your agent can: - Take actions on sites with no API, like booking a reservation - Test an application end to end through its UI - Hand (x.com) 8. Launch HN: Simantic (YC F26) – Firmware Simulation for AI Agents (news.ycombinator.com) 9. Welcome Le Chonk: Mistral Large 4 on Hugging Face (huggingface.co) 10. Introducing AnySpeechRecognizer and AnyDecisionModel: Open Models for Speech and Decisions on Apple Platforms (huggingface.co) 11. You can now generate, preview, and edit documents, PDFs, presentations, spreadsheets, and more directly in Gemini. (blog.google) 12. Quoting Terk Terkowitz (simonwillison.net)