The model that decides instead of chatting. Tech & AI Weekly by Eli, October 02 2026
Hi! Two Fridays off (WeAreDevelopers ate my week), so this one’s fuller. Decision models, three new frontier models, lower prices. Here’s what mattered.
AI & GenAI
- TypeSafe launched Jev, a model that can’t write a sentence. You give it a state and typed questions; it returns a choice with a calibrated probability. No prose to parse. Most of what agents do is decide, not write, and we’ve been paying LLM prices for it. Read more
- AWS open-sourced Strands Decider 2B, a Jev you can run yourself. Same idea, 2B, ~115ms on a laptop, Apache 2.0 on Hugging Face. Jev is hosted-only. This one runs on your machine. Read more
- Google announced Gemini 4 Argon. You can’t use it yet. Aimed at long work, with a 1M-token output limit. Rolling out to “trusted cyber defenders” first, developers later. Good to know it exists; don’t plan around it. Read more
- OpenAI’s GPT-6 Sol and Luna cut API prices in half. $2/$10 for Sol, $0.10/$0.50 for Luna, with a 6.1 Sol update a week later. The frontier is getting cheap enough to leave running. Read more
- Claude Sonnet 5.5 is on Amazon Bedrock, cheaper per task than Opus. Built for well-scoped work, faster, lower cost. Opus for judgment calls, Sonnet for the rest. Read more
- OpenAI confirmed a prompt injection that spreads like a worm. It tells the model to copy itself into its next output: email, file, commit. If you wire agents to real connectors, this is your threat model now. Read more
Cloud & DevOps
- AWS Well-Architected Agent is in preview. It reviews your infra (or your IaC) across 65+ services and hands back the fixes, ranked by your business goals. It won’t apply them, and it can be wrong. Still beats a checklist. Read more
- Docker is packing agent permissions into an OCI image. The Sandbox Kit Spec (v3, Apache 2.0, going to the CNCF) ships an agent plus the hosts, credentials, and volumes it’s allowed to touch. Same tools you already use for containers. Read more
- GitHub’s open-source AI security agent found 24 real Android bugs. A location leak in OsmAnd, a session-token takeover in Wikipedia for Android. The taskflows are open source; you can run them. Read more
From me this week
No new post this week. From the last two Fridays I skipped, in case you missed them:
- How to Stop AI Agent Memory Poisoning. Block bad data at the write path, before it’s stored. Read more
- Semantic Caching for AI Agents in Production. Reuse answers to questions that mean the same thing, not just the same string. Read more
- AI Agent Audit Trails. Prove why your agent decided, not just what it did. Read more
Where to find me
Two next week:
- WomenWhoCode Leadership Summit. October 7. Semantic tool selection: cutting agent errors 75% and token costs 89%. Details
- Road to NODES 2026 (Neo4j). October 8, virtual and free. Workshop: 5 techniques to stop agent hallucinations. Details
I hope you enjoy this as much as I enjoy putting it together every week. It’s my little weekly de-stress moment. Let’s stay informed together and learn something new along the way.
See you next Friday,
Eli
Dev.to · LinkedIn · GitHub · Twitter/X · Instagram · YouTube