The one thing
OpenAI publishes a guide to choosing and deploying GPT-6 models for startups
Why it mattersStartups can now follow a clear process to deploy GPT-6 in production without guesswork.
Seventy: Sunday, 4 October 2026
What happened. OpenAI published a guide for startups on choosing between GPT-6 models, tuning reasoning effort, improving prompts and skills, coordinating tools, and preparing workflows for production.
02Agents
Microsoft and Hugging Face release ThinkingBox to grade AI agents by real system state
Why it mattersThinkingBox exposes when agents falsely claim task completion by measuring real system state, not just tool usage.
What happened. Microsoft's ThinkingBox grades AI agents on actual database state changes, not just tool calls, and is now available on Hugging Face.
AWS shares design for checking thousands of leases against state law using chat and rules
Why it mattersAWS is showing that the AI chats and a rules engine rules on compliance, so don't depend on a model's judgment.
What happened. AWS published a design pattern called Adjudicated Query, with a reference architecture for Amazon Quick. Staff type questions into a chat tool, while a fixed, non-AI rules engine makes every pass/fail call on each lease.
04Training
AWS explains how to train a multi-step search agent on SageMaker
Why it mattersAWS says prompted alone rarely search reliably over multiple steps; training on your own tools is the fix.
What happened. AWS published a walkthrough for training a small language model to run multi-round searches, using reinforcement learning on Amazon SageMaker AI. The agent learns your tools and data by trial, picking its own queries and retrieval methods and stopping when it has enough.
05Agents
AWS shows how to give Bedrock-hosted Claude Desktop live web search
Why it mattersClaude Desktop on can now pull live web data through , so answers are no longer capped by the .
What happened. AWS published a guide to connecting Claude Desktop on Amazon Bedrock to Web Search through AgentCore Gateway. It lets the model fetch current information, such as documentation updates and live pricing, instead of relying on its training cutoff.
That's the whole issue.
5 stories, about 1 min 45 sec. Nothing was left out to fill the time.
Words of the day
- Reinforcement learning
- Training by trial and reward: the model tries, is scored on the result, and gradually does more of what scored well.
- Small language model
- A language model small enough to run cheaply or on a phone or laptop, usually trained for narrower jobs.
- Knowledge cutoff
- The date after which a model knows nothing, because its training stopped there.
“When something is important enough, you do it even if the odds are not in your favor.”