Tuesday, October 6, 2026
3-minute read
3
AI UPDATES
2
RESOURCES
1
PRO TIP
👋 Good morning.
Claude started watermarking its text on August 17. Today OpenAI says ChatGPT and Codex text in the EU will carry an invisible mark of its own, textGrain, over the coming weeks, while API orgs anywhere can switch the same signal on today. If you ship AI writing into Europe, the provenance conversation just got a second seat at the table, and the detector is still for approved researchers first.
3 AI updates
textGrain, OpenAI's EU text approach. Source: OpenAI.
OpenAI is rolling out textGrain, an invisible statistical watermark in model word choices, so a detector can ask whether a passage carries an OpenAI mark. Over the coming weeks it will land on eligible ChatGPT and Codex text in the European Union. API customers globally can opt in today for select models; it stays off by default in the API. Detector access starts with approved researchers and expert organizations, per OpenAI and TechCrunch. Short, edited, math, and translated text are harder to catch, Replacing about 10% of words with synonyms dropped detection from roughly 92% to 66% in one OpenAI test.
If your EU drafts leave ChatGPT or Codex for customers or regulators, assume the text may carry a machine-readable signal soon, and treat a missing watermark as inconclusive, not proof of a human.
Claude got there first. OpenAI just caught the train.
Instinct group chats. Source: TechCrunch / Getty.
Noah Shinn says early-access users can add Instinct to group chats starting today. A new Instinct joins and works for the whole group, so friends can explore options, agree on a plan, and finish it in one thread, and they do not need Instinct themselves. Examples he lists: weekend trips, roommate logistics, tickets that go on sale, fantasy leagues, carpools, and Thanksgiving lists. Permissions are deliberate: your personal Instinct asks before connecting, the group's Instinct has no direct access to your personal accounts, and a new member freezes pending replies until you approve the larger circle. Early access gets it now; everyone else is "soon," and you can ask your Instinct to join the list.
If your team still plans in a text thread that dies at "next weekend?", this is the agent that sits in the chat instead of living in a separate app tab.
The group chat finally got a roommate who takes notes.
Beam, Reflection's open-weight MoE. Source: Reflection.
Reflection introduced Beam: a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, trained end-to-end from scratch for coding, reasoning, and agentic work. The company says it advances the Western open-weight frontier on coding and agentic tasks, with full weights, a technical report, and developer artifacts due later this month. Pretraining used 23.8 trillion curated tokens. A high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks. Early access signup is open while red-teaming finishes.
If you self-host coding agents and watch Western open weights, put Beam on the shortlist for when the weights land, and weigh the active-parameter efficiency claim against the closed models you already rent.
501 billion parameters, 23 billion on duty. The rest are on call.
2 Resources
Evals cover. Source: Sergii Makarevych / X.
Sergii Makarevych wrote an X Article for everyone who ships, buys, or signs off on LLM software (engineers, product managers, and the CEO) in plain language from the big picture down. An eval is a repeatable test that returns a number you can trust about whether the system does what you want, and whether a change helped. His four-step loop: collect inputs with gold answers, run the system, grade with code or a judge model or a human, then turn grades into a number with an error bar. The grader itself must be checked, or the team fools itself with a confident fake score. Worth it if "it looked fine in the demo" is still your release gate.
Hermes Agent user stories. Source: Nous Research.
The Hermes Agent user-stories page collects what people are actually building with the agent: scraped from X, GitHub, Reddit, Hacker News, YouTube, blogs, podcasts, and Discord. The page currently lists 326 stories across 15 categories and 11 sources, from solo job-site apps to French SMB installs to overnight "dreaming" cron jobs. Browse it for concrete workflows, not a pitch deck, each tile links out to a real post or gist.
1 Pro Tip
Two bare-minimum Dots basics. Source: YouTube / digitalSamaritan.
A YouTube short from digitalSamaritan lays out two bare-minimum setups so ChatGPT Dots do not feel like ChatGPT work on voice mode. First: list the tools you use for work. If they exist as plugins inside ChatGPT (not Codex), connect those plugins. For everything else, take control of the Dot's own computer and log into the apps and sites you want the agent to use, skip banking and anything super sensitive, so your machine can stay off while you talk to Dots. Second: give the agent its own email, either a new Gmail or a tool like AgentMail, and start looping that inbox into the threads where you want the agent to act.
If voice mode still feels like you are babysitting a chat window, these two are the floor before you chase fancier automations.
🎤 Join Kushank Live on Friday!
How to edit videos with Claude
Build in real time and bring your questions.
📅 Friday, October 9, at 10AM Pacific / 1PM Eastern
Reply to us if you wanna be a part of this!
Hope to see you there!
That's your 1% for today. Kushank "Fingerprint Duty", with AI teammates and the human team.
PracticalyAI
