AI on Phil Wheat

OpenCode Documentation

Tue, 13 Oct 2026 09:13:00 -0500

The documentation for OpenCode, an open source AI coding agent that runs in the terminal (also as a desktop app or IDE extension). You bring your own LLM provider API keys, or use their curated model list. Covers install (script, npm, Homebrew, packages for Windows), configuration, session sharing, themes, custom commands, and an undo/redo workflow for agent edits. OpenCode v2 is now out. A direct Hermes competitor in the terminal-agent space, and the docs are a good read even just for feature comparison.

Sebastian Raschka

Tue, 13 Oct 2026 09:12:00 -0500

Sebastian Raschka is an LLM research engineer (ex-Lightning AI, former statistics professor at UW-Madison) and author of “Build a Large Language Model (From Scratch)” and “Build a Reasoning Model From Scratch”. His site indexes the Ahead of AI newsletter: long-form articles (controlling reasoning effort in LLMs, using local coding agents with open-weight models, building an AI text detector from scratch, curated LLM paper lists) plus short architecture notes on new models (Kimi K3, Muse Glimmer 30B, how Claude’s text watermarking works). Consistently one of the better free technical explainers out there.

CodeFlow — Architecture Intelligence

Tue, 13 Oct 2026 09:11:00 -0500

Gives you an overview of a git repo.

CodeFlow is an open-source visualizer that turns any GitHub repository into an interactive architecture map: file relationships and dependencies as graph/tree/matrix/map views, blast-radius analysis for changes, code ownership, security issues, and design-pattern detection. The underlying project is braedonsaunders/codeflow; it runs locally as a single HTML file. Good for walking into an unfamiliar codebase.

txtai Semantic Search Examples

Tue, 13 Oct 2026 09:08:00 -0500

The examples index for txtai, the NeuML embeddings library. The semantic-search section is a notebook gallery: building embeddings indexes from Hugging Face datasets or your own data sources, adding semantic search to Elasticsearch, image+text similarity in one embedding space, custom SQL functions, model explainability, query translation, QA databases, semantic graphs, and BM25 topic modeling. There are parallel sections for LLM/RAG, pipelines, workflows, and model training. A practical catalog if you’re evaluating txtai against other RAG stacks.

Skills — The Agent Skills Directory

Tue, 13 Oct 2026 09:07:00 -0500

Open Agent skills, cross harness so they should work various places.

skills.sh is a directory of reusable agent skills installable with a single npx skills add command. It explicitly spans agents: Claude Code, Cursor, Codex, Copilot, Windsurf, Gemini, Cline, OpenCode, Goose, Zed, VS Code — and Nous Research/Hermes is listed among them. Skills are markdown instruction bundles (procedural knowledge), published per-repo (anthropics/skills, obra/superpowers, mattpocock/skills…). Since Hermes already uses the same SKILL.md convention, most of these should drop straight into the skills directory after review.

Joy Caption Alpha Two

Tue, 13 Oct 2026 09:05:00 -0500

JoyCaption is an image-captioning vision-language model built from the ground up to be free, open, and uncensored, aimed at generating captions for training diffusion models — fancyfeast’s claim is near-GPT-4o captioning quality while staying unrestricted. Alpha Two is the Llama-3.1-based iteration (LLaVA-style), with prompt-able caption formats (descriptive, Midjourney-style, art critic, product listing, tags) and 4-bit quantizations for low-VRAM cards. The linked HF Space is the hosted demo; ComfyUI custom nodes exist for local use.

Infinite Livestream

Tue, 13 Oct 2026 09:04:00 -0500

Infinite livestream code. Needs a LOT more hardware than I have, but is a curiosity.

A chat-driven, never-ending AI video broadcast from reactor-team: viewers type !prompt <idea> in Twitch/YouTube chat, an LLM expands each idea into styled scene sequences, the fast-h3 model generates 768p clips with synchronized audio, and it all goes out over RTMP as one uninterrupted stream. Two halves: the model queue (runs on the Reactor Runtime, 8x B200 — yes, really) and a streaming client (Python + reactor-sdk, any box with ffmpeg). Apache-2.0 code; MiniMax H3-licensed weights. The architecture writeup is interesting even if the hardware bill never happens.

The Teaser Period: Why the AI Boom Is Hitting a Reset Wall

Tue, 13 Oct 2026 09:03:00 -0500

Financial questions around the AI boom.

Groundbreaker’s argument, drawn straight from 2008: trillions in AI compute commitments come due in 2027–2028, structured like the two-year teaser rates on subprime 2/28 ARMs. During the teaser period everything looks fine — borrowers pay, bonds coupon, spreads stay tight — but the aggregate reset schedule is contractual and dated, and when it’s charted you get the reset wall. The piece walks through the mechanics of who is committed to what, when payments step up, and how the AI boom breaks if revenue doesn’t arrive before the resets.

Yan Desbiens

Tue, 13 Oct 2026 09:01:00 -0500

Blogger — some interesting ideas. AEON is what caught my eye on this one.

Yan Desbiens runs a one-person, one-RTX-4090 local-first AI lab and documents it all. The AEON project: an open-source, deterministic procedural world simulation steered by a local LLM “world-spirit”, with embodied citizens, a teacher-to-student liquid-neural-network mind, and a Three.js dashboard. Also on the site: NeuroArch (background cognition cycles between prompts), ForgeLM (train 30M–500M models on one 4090), UFM (fractal memory so a single GPU runs models that shouldn’t fit), agentos, and genetic architecture search. Unfashionable from-first-principles local AI, written up candidly.

Gamentic

Tue, 13 Oct 2026 09:00:00 -0500

Self-hosted dungeon RPG.

Gamentic’s docs/blueprints page describes the architecture in one line: a self-hosted AI dungeon RPG where one local LLM plays every role through purpose-built contexts (agents); whatever it wants to change in the world must pass a validated tool into SQLite as the single source of truth; a FastAPI “brain” resolves one turn per request and pushes late media over SSE; the frontend is a no-build vanilla SPA, with image and voice as optional local services. The site is an interactive blueprint of the agents, engine, state, and repo tree. A nicely disciplined pattern for LLM-with-tooling-state projects.

Genspark Blog

Tue, 13 Oct 2026 08:59:00 -0500

A Hermes competitor? Maybe more GuideAnts? But hosted only, no local that I can see.

The Genspark blog tracks their “AI employee” workspace product: successive AI Workspace releases (1.0 through 6.0), Super Agents built from one prompt, an AI Developer coding agent, GenOffice (claimed first open-source AI office suite), a Gen-1 Slides model, and a Microsoft collaboration, plus founder Eric Jing’s “Seeing AGI” essays. Positioned as hosted consumer/enterprise — they signed the Open Weights letter and tout an on-device browser AI, but there’s no self-hosted story like Hermes has.

Pirate Face

Tue, 13 Oct 2026 08:55:00 -0500

AI models as torrents.

Pirate Face mirrors eligible open-licensed (MIT/Apache-2.0) Hugging Face models as checksum-verified BitTorrent magnets — “the permanence layer for sovereign AI”. If a model page disappears, peers keep seeding; downloads can be verified bit-for-bit against recorded SHA-256s. Terminal search via curl https://pirateface.co/s/qwen, verified creator handles tied to HF accounts, 669k+ eligible models listed, live seeder counts shown (many need seeders). A planned drop-in HF_ENDPOINT API would make it a transparent alternate download path.

AMD AI Developer Program

Tue, 13 Oct 2026 08:54:00 -0500

AMD developer program. The Rocm cert was pretty on point, especially for opening up CUDA only programs.

Free-to-join program for building AI on AMD hardware: $100 in AMD Developer Cloud credits, a complimentary month of DeepLearning.AI, early event registration, a private Discord with AMD engineers, and a points system (ROCm Certification is worth 500 points; monthly sweepstakes for a Radeon or Ryzen AI PC). The announcements feed tracks ROCm 10, the online ROCm certification course, and Day-0 model support writeups (Qwen 3.8, MiniMax-H3 on Instinct). The certification path pairs with the LinkedIn-badge question.

Attie

Tue, 13 Oct 2026 08:51:00 -0500

AI research assistant — looks to be hosted only, closed beta.

Attie positions itself as research tooling for the open social web: “a tool you choose to open, not something injected into your feed” — no engagement tricks, you ask, it helps, you decide what to do with what you find. Currently in closed beta with a waitlist admitting people gradually. Hosted only; no local/self-host option advertised. Interesting angle: an AI assistant pointed at decentralized social data rather than the walled gardens.

RAGFlow

Tue, 13 Oct 2026 08:49:00 -0500

RAGFlow is an open-source (infiniflow/ragflow) “context layer” for agents: an ingestion pipeline that parses multi-format documents into rich semantic representations (template-based, layout-aware chunking is the flagship claim — tables and structure included), hybrid retrieval combining vector search + BM25 + custom scoring and re-ranking, and a visual agent-workflow builder with MCP/tool integration. Hosted cloud with a free tier, or self-hosted from GitHub. The deep chunking approach is what “implicit data” extraction would rest on; worth a hands-on eval against txtai given the pace of change.

Train Your Own Decision Model with Unsloth

Tue, 13 Oct 2026 08:43:00 -0500

Using Unsloth desktop to train a classifier model.

Unsloth’s guide to turning ordinary LLMs (Qwen, Gemma, Llama) into decision models like their Jev/Laya/Clef: the fine-tuned model reads input, scores the options you provide, and returns a choice with a calibrated probability instead of generating text. LoRA (rank 16, 4-bit, lr 2e-4, 1–2 epochs) lifts accuracy from roughly chance (30–37%) to 73–81% on decision benchmarks (typed-decisions, BANKING77, CLINC150), with test sets decontaminated against training. Runs through the Unsloth Desktop app’s Train tab (macOS/Windows/Linux) or code via FastDecisionModel, then serves the model under their Decision API. Modest VRAM: a 0.8B model trains in 4 GB in ~40 minutes.

SiMa.ai Palette

Tue, 13 Oct 2026 08:41:00 -0500

SiMa.ai is an edge-AI chip company (Modalix MLSoC, ~50 TOPS at sub-10W, targeting drones, robotics, automotive, industrial vision). Palette is their software environment: “Palette Neat”, an agentic, containerized build environment where human+agent developers go from ONNX models to a running binary on the MLSoC in hours instead of months — model compiler toolchain, purpose-built physical-AI libraries, runtime orchestration on host and chip, plus a Model Browser and devkits. “Physical AI” here means on-device/embedded inference, not robotics research.

Aquarium

Tue, 13 Oct 2026 08:40:00 -0500

Has hosted offerings playing against each other using money for points.

Aquarium is a live experiment: four frontier models (Claude, GPT, Gemini, Grok) each holding real on-chain money on Base play 20-Questions-style guessing games against each other for a pot — yes/no questions cost fees, wrong guesses lose the ante, a host model called the Sphinx audits. The page streams the game feed, live bankrolls, and a full ledger of past games (“clamp”, “toothbrush”, “sourdough bread”), every transaction verifiable on Basescan. Each season is framed as an experiment with emailed findings. A cheap, honest window into model-vs-model decision behavior.

Abacus ChatLLM

Tue, 13 Oct 2026 08:38:00 -0500

A hosted agent - I haven’t tried it but several people recommended it.

ChatLLM (from Abacus.AI) is an “AI super-assistant” subscription: one ~$10/month account for a rotating roster of all the frontier models (the FAQ names GPT-6 variants, Sonnet/Opus, Gemini, Grok, Qwen, DeepSeek, Kimi, GLM plus ~20 more, updated 24–48h after releases), image and video generation, a coding agent CLI, and a general Abacus AI Agent. The differentiators are breadth plus enterprise plumbing: team seats, conversation imports from ChatGPT/Claude, connectors to GDrive/Slack/Confluence, custom chatbots on your data, SOC-2 Type-2 and HIPAA. Hosted only — nothing local — with generous-feeling token limits and automatic model-switching when rate-limited.

Please, Don't Code

Tue, 13 Oct 2026 08:37:00 -0500

LLM-based embedded board programming. Haven’t tried it yet, but I should.

PleaseDontCode (PCGen) is an end-to-end hardware platform for people who don’t write embedded code: pick one of 35+ boards (ESP32 variants, Arduino Uno/Nano/Mega, RP2040, ESP8266, M5Stack, Pi Zero 2W…), describe in chat what the device should do, and the AI draws the wiring schematic and generates firmware kept in sync with it — with hardware-verified pin assignments (it knows GPIO6 on an ESP32 is unusable, unlike a plain chatbot). Flash via USB straight from the browser, then POTA (“Programming Over The Air”) handles WiFi firmware updates and a live monitoring dashboard with widgets. Free for one device.

ModelScope

Tue, 13 Oct 2026 08:34:00 -0500

ModelScope is Alibaba Cloud’s open-source model community — China’s largest model hub, with an English-language site since 2024: thousands of ready-to-run models (Qwen series front and center, plus Baichuan, Zhipu, and user-contributed models), datasets, and interactive demos, organized around a “Model-as-a-Service” idea. It’s the closest thing to a Hugging Face equivalent on the Chinese side of the ecosystem, so some Qwen releases and Chinese-model checkpoints show up there first or with extra tooling. Useful for model discovery; check licensing and mirror availability before depending on it.