Tokens & Signals for 10/2/2026. We scanned ~1,200 Twitter accounts (1361 tweets), 13 subreddits (0 posts), Hacker News (6 stories), 3 newsletter posts, 5 podcast episodes, 50 Discord messages, and leaderboard data for you. Estimated reading time saved: ~9 hours.
* OpenAI launched Dots, persistent personal agents that sit in the background, learn your workflow, and handle things like inbox management. $20/month on Pro. x.com/sama/status/2106085986606403684
* Claude Code "Mods" is live — developers can now inject custom plugins. Community builds already include Doom clones and dev-workflow helpers. x.com/ClaudeDevs/status/2106118517447876618
* Perplexity open-sourced pplx-decider-27b, a 27B MoE decision model built for intent routing at $0.04/M input tokens. x.com/AravSrinivas/status/2105774153903268288
* @giffmana on where research is heading: "We're past the point of 'AI isn't that good.' In structured research, not using AI might be malpractice." Models are hitting near-100% accuracy on accounting tasks. Humans? 37%. x.com/giffmana/status/2105910335391596928
* llama.cpp now runs decision models locally via a new /v1/systemone endpoint, so your agent logic stays completely offline and private. x.com/ggerganov/status/2106029758350032937
* NYC is pushing strict legislation that requires third-party audits for any AI model used in municipal decisions — $50k fines per violation. x.com/GaryMarcus/status/2106021408518353226
* @karpathy on the new agentic reality: "The shift from 'AI as a tool' to 'AI as a coworker' is happening faster than anyone predicted."
* McDonald's has rolled out AI pricing across 14,000 U.S. stores, optimizing menu prices in real-time based on local demand. x.com/MTSlive/status/2106071415078449425
* Anthropic is taking heat over its internal culture, with critics calling it "cult-like" due to an intense focus on alignment frameworks. x.com/giffmana/status/2106114331481952539
* @theo on T3 Code's momentum: "400k users. Doubled in 43 days. The appetite for agentic dev environments is insatiable." x.com/theo/status/2105921113603952853
Best to Build With Today
* Coding: claude-opus-5.5-max (WebDev/Fullstack leader)
* Reasoning: gemini-4-argon-high (Math/Hard Prompts leader)
* Chat: gemini-4-argon-high (Overall #1)
* Image Generation: gpt-image-2.5-sunburst (Visual quality/Editing leader)
* Video Generation: gemini-omni-1.1-flash (Text-to-Video leader)
* Agent Performance: Claude Fable 5.1 (Max)
* Open-Source: kimi-k3-max (Text/Coding performance)
* Value Pick: pplx-decider-27b ($0.04/M input for reliable routing)
Deeper Dives
💼 Industry & Business
NYC Proposes Strict AI Model Validation
NYC Speaker Menin introduced legislation requiring third-party audits for any AI model influencing municipal decisions. Companies will need to submit bias mitigation and explainability reports — or face $50,000 fines per violation. It's a clear signal that compliance-first regulation is coming to major cities.
� Twitter� Hacker News
Anthropic Faces Culture Scrutiny
Anthropic is under fire over reports of "cult-like" internal dynamics. Critics are arguing that the company's "Constitutional AI" approach has drifted from being a technical alignment method into something more like an insular philosophical identity.
� Twitter
McDonald's Dynamic AI Pricing
McDonald's is now running an AI-driven pricing engine across all 14,000 U.S. stores. The system chews through millions of transactions and sets prices location-by-location based on local willingness-to-pay. Your Big Mac price may vary.
� Twitter
Meta Parts Ways With Virtue AI
Meta has ended its partnership with Virtue AI, the firm it brought in for ethical AI research. It's the latest in a string of big tech companies quietly walking back independent ethics oversight.
� Twitter
🧠 Models & Research
AI Agents Outperform Humans in Research
New research has AI models hitting 100% accuracy on accounting tasks versus 37% for humans. On scientific questions, agents were 1.4x faster with 92% accuracy. At this point, using AI for structured research isn't a nice-to-have — it's becoming the baseline expectation.
� Twitter� Reddit
Perplexity's New Open-Source Decider
pplx-decider-27b uses a Mixture-of-Experts architecture for routing and intent classification. Apache 2.0 licensed, cost-efficient, and 15% better than previous open-source alternatives. Solid option if you're building anything that needs smart query routing.
� Twitter
Google's TEE Federated Learning
Google Research dropped a system that uses Trusted Execution Environments to enable private, verifiable Federated Learning. It threads the needle between server-side speed and strict on-device data privacy — worth a read if you're building anything privacy-sensitive.
� Twitter
🚀 Products & Launches
OpenAI 'Dots' Personal Agents
Dots is OpenAI's answer to the persistent agent problem. It runs quietly in the background using a low-latency GPT-4o variant, learns your context and personal style over time, and does things without you having to ask. Less chatbot, more autonomous assistant.
� Twitter
Claude Code 'Mods'
The new MIT-licensed Mods framework lets developers plug custom tools directly into Claude Code. Early results show a 22% reduction in hallucination rates, and the community is already building things like repository management tools on top of it.
� Twitter
DeepSeek Harness Desktop
DeepSeek shipped native desktop apps for Windows and macOS. You get unified management of local inference and API connectivity — no command-line wrestling required.
� Twitter
llama.cpp Adds Decision Model Support
The latest llama.cpp update lets you run decision-making models locally via a /v1/systemone endpoint. Supports 1B–3B quantized models, which means your agent's "brain" can stay completely offline. Big deal for anyone who cares about privacy.
� Twitter
Launches
* OpenAI Dots: Always-on personal agent with persistent context.
* pplx-decider-27b: 27B parameter MoE model for intent routing.
* Claude Code Mods: Extensible plugin framework for coding agents.
* DeepSeek Harness: Native desktop model management for Windows/macOS.
Closing thought: The AI agent era isn't coming — it's here. From persistent Dots running in the background to decision models you can run entirely on your own machine, the tech has quietly crossed from "cool chatbot" to "reliable coworker." The only real question left is how you're going to manage your new team.