The newsletter you wish existed, written for you every week.
Describe your niche in a sentence. It searches the web for what's new, checks when every source was published, and sends you only what you haven't seen. React to a story and the next issue adjusts.
AI Matters, issue of 2026-10-05
OpenAI drops GPT-6.1 Sol and Dots agents; Gemini 4 Argon arrives
- Oct 5emergent.sh
OpenAI releases GPT-6.1 Sol multimodal model and launches Dots AI agents
GPT-6.1 Sol processes text, image, and structured data inputs for cross-modal reasoning tasks like financial analysis and document processing, launching September 29 under a proprietary license alongside GPT-6 variants Astra and Luna. The same DevDay unveiled ChatGPT Space, a collaborative workspace blending Google Docs and Notion for humans and AI agents, and Dots, an always-on personal assistant powered by GPT-6 Astra with a dedicated cloud computer that can access email, calendar, Slack, and documents to suggest priorities and book travel. Sam Altman called GPT-6.1 Sol the fastest-growing OpenAI model release ever.
- Oct 1blog.google
Google launches Gemini 4 Argon with 1M-token output and coding focus
Google DeepMind introduced Gemini 4 Argon for coding, enterprise knowledge work, and cyber defense, offering up to 1M-token output, a 50% introductory discount, and strong benchmark results against GPT-6 Astra and Claude Opus 5.5. Access begins with government users in the Fairwind Program, with broader rollout planned after guardrail refinement.
- Oct 3superpowerdaily.com
Anthropic adds Claude Code Mods that can rewrite prompts and replace built-in features
Anthropic introduced Mods on October 1, 2026, creating a TypeScript extension layer that can intervene before tool calls, permission requests, and screen rendering, replacing built-in behavior like the /diff feature rather than just responding to events. Mods inherit Claude Code's machine access, with organization marketplace restrictions and a protective sec-default mod for managed installations.
What happens between your sentence and your inbox
Six steps, every issue. Figures are from the issue above.
- 1
Plan
Reads your brief, your past reactions and every story you already got, then decides what to look for since the last issue.
Nemotron-3_5-Lightning
6 searches planned
- 2
Search
Runs those searches on the live web: news, blogs, official sites, Reddit, YouTube and the sources you trust.
Tavily
113 results, 161 unique with your subscriptions
- 3
Score
Reads every result and scores how much it matters to you. Articles about the same event are grouped.
NVIDIA-Nemotron-3-Nano-30B-A3B
160 scored, 68 at 6+
- 4
Check dates
Opens each promising page and reads when it was really published. No date, no story.
Page metadata, no AI
27 checked, 11 dated, 16 unknown
- 5
Filter
Drops anything old, weak, blocked or already sent, and keeps any one site from taking over.
Plain rules, no AI
32 kept
- 6
Write
Writes the issue in your language from those sources only. Every link comes from the search, never from the model.
Nemotron-3-Ultra-550b-a55b
6 stories from 6 publishers, 3 reported by several of your sources
That issue took 389 seconds and cost $0.19: $0.11 of web search and $0.076 of Nemotron. Every issue shows its full trace.
Every story is provably recent
Search engines happily return a plugin page from 2016 for a query about this week. An AI writer will then call it “new”. That is how automated newsletters lose your trust.
Newsletter for One opens every promising source and reads its real publication date: the upload date of a video, the published time in an article's metadata. If a date can't be verified, the story doesn't ship.
- Sep 24
- Verified on the source itself. The only kind of date that makes it into your issue.
- Sep 24
- Only mentioned in the text. Shown as unverified.
- ≈ this window
- Reddit posts: no exact date without Reddit's API, but search only returns posts from this window.
- Sep 15 · old
- Published before the window. Dropped.
- no date
- No date anywhere. Dropped.
It learns what you mean
Mark a story useful or not and say why in a few words. It turns that into lasting preferences that steer what it searches for and what it keeps.
👎“preset packs are basically ads”
becomes
- Avoid: Transcript-based automatic rough-cut extensions for Final Cut Pro
You decide where it looks
- Sources you trust
- Get a search of their own every issue.
- Sources you don't
- Are excluded from every search, with one click from any story.
- Search in one language, read in another
- Most sources for a niche are in English. It searches where the sources are and writes in the language you read.
- Never the same story twice
- It remembers every piece of news it sent you, not just every link.
Built with
Every model is an open NVIDIA Nemotron model served by Nebius Token Factory, picked for its job: the cheapest one where volume is high, the largest one where you read the result.
| Job | Runs on | Why this one |
|---|---|---|
| Planning searches, learning from feedback | Nemotron-3_5-Lightning | 1M-token context at the lowest price: sees your whole history |
| Scoring every search result | NVIDIA-Nemotron-3-Nano-30B-A3B | Cheapest and fast; runs on dozens of results per issue |
| Writing the issue | Nemotron-3-Ultra-550b-a55b | NVIDIA's largest model, where quality is visible; about a cent per issue |
| Searching the web | Tavily | Search built for agents, with date and domain filters |
| Checking dates | Page metadata | Facts, not guesses: no model involved |
Start with one sentence about what you want to follow.
Start a newsletter