From Product Hunt
Agent networking, a Claude cost tracker, and privacy-first data extraction.
Tobira.ai — A network where AI agents find deals for their humans
Your AI agent gets a public address in a network of other agents. Agents discover each other, negotiate, and connect their humans. Works with OpenClaw and Claude Cowork. Claim your @handle before the short ones are gone.
Honestly — Real reviews from Reddit & YouTube when shopping online
Chrome extension that surfaces real product opinions from Reddit, TikTok, YouTube and Instagram directly on product pages. No ads, no AI summaries, no sponsorships — just verified human reviews where they belong.
Claude Usage Tracker — See exactly how much you spend on Claude
Auto-detects 9+ tools (Cursor, Claude Code, Windsurf, Cline), scans local session data, and shows daily costs, model breakdowns, heatmaps, and monthly projections in one dashboard. Native macOS app, MIT licensed, no cloud.
DataSieve 2.0 — Extract structured data from text, files and archives
Drop in text, files, folders, or archives and extract emails, phone numbers, URLs, dates, and financial data in one pass. Supports PDF, EPUB, CSV, JSON, Word. Define custom extractors. Everything runs locally.
AlphaClaw Apex — OpenClaw harness and fleet manager for Mac
Native Mac app for managing multiple OpenClaw VPS instances. One-click Hetzner deploy, crash detection with auto-repair, hourly GitHub backups, and Telegram/Discord/Slack alerts. No SSH needed.
From Reddit
Jensen Huang declared AGI on the Lex Fridman podcast, Xiaomi dropped a trillion-parameter agent model out of nowhere, and Anthropic shipped phone-to-desktop computer use.
On the Lex Fridman podcast today, the Nvidia CEO said he believes artificial general intelligence has been achieved. Cited OpenClaw autonomous agents as evidence that AI can now handle complex, open-ended tasks.
He hedged slightly — when pressed on scope, he noted the question was about running a billion-dollar company but didn’t specify for how long. Still, the most definitive AGI claim from a major tech CEO to date.
Xiaomi MiMo-V2-Pro — phone company now #3 on global agent benchmarks
1T parameters (42B active), ranks behind only Opus 4.6 on agent tasks at 1/6th the price. Lead researcher came from DeepSeek. Spent a week on OpenRouter under the codename “Hunter Alpha” — the entire community thought it was DeepSeek V4.
GPT-5.4 Pro solves a previously unsolved math problem
Epoch AI and the original problem author confirmed GPT-5.4 Pro solved a Ramsey-style problem on hypergraphs from FrontierMath’s Open Problems set. Solution to be written up for publication. A second open problem was also solved by GPT-5.2 Pro in a separate harness.
Anthropic Dispatch — control Claude from your phone, computer use on Mac
Dispatch lets you send instructions from your phone while Claude executes on your Mac desktop. Also shipped native computer use: Claude opens apps, navigates browsers, fills spreadsheets. Research preview for Pro and Max subscribers.
Vibe Physics — a Harvard professor used Claude as an AI grad student
Physics professor Matthew Schwartz guided Claude through a real theoretical physics paper without ever touching a file. 110 drafts, 36M tokens, 40+ hours of CPU compute. Two weeks instead of the usual year. “Impressively capable, but sloppy enough that domain expertise was essential.”
Quick hits from LocalLLaMA.
FlashAttention-4 hits 1,613 TFLOPs/s on B200
71% GPU utilization. Attention is at matmul speed now. 2.7x faster than Triton, 1.3x faster than cuDNN. Already in vLLM 0.17.0.
China’s open-source dominance threatens US AI lead, advisory body warns
Reuters: US advisory body warns Chinese open-source models now dominate the open-weight landscape, threatening American AI leadership.
From Reddit
Practical workflow hacks — a voice cloner built from a video model, and raster-to-vector generation.
LTX 2.3 hacked into a multilingual TTS voice cloner
Creative ComfyUI workflow repurposes LTX 2.3’s audio capabilities for voice cloning. Generates a tiny 64×64 video with input audio, masks the audio latent space after 10 seconds, then the model continues speech in the cloned voice. Works with 20-second voice samples.
Flux.2 Klein Raster-to-Vector workflow for ComfyUI
Generates raster images with Flux.2 Klein (4B or 9B), then converts to SVG on the fly. Great for illustrations, stickers, and cartoons.
From Reddit
Dropzone got a big Tahoe redesign, a free system-wide EQ worth trying, and self-hosted tools for the stack.
Dropzone 5 — completely redesigned for macOS Tahoe
New translucent glass look, brand-new grid selector for switching between context-specific Dropzone grids, and revamped Tahoe toolbar style. Available on Setapp. Lifetime license $25 at launch price, 50% upgrade discount for Dropzone 4 users.
Radioform — free, open-source system-wide EQ for macOS
10 fully parametric bands, 20Hz–20kHz, zero latency. Native Swift/SwiftUI menu bar app that plugs directly into CoreAudio. Not Electron. Open source on GitHub.
iSponsorBlockTV — skip YouTube sponsors and ads on any TV app
Self-hosted app that auto-skips sponsor segments using SponsorBlock API and presses “Skip Ad” the moment it appears. Works with Apple TV, PlayStation, Nintendo Switch, smart TVs. Docker install, no sideloading needed.
UniFi OS Server in Docker
UniFi is deprecating the standalone Network Controller. This project runs OS Server natively in Docker/Kubernetes with unprivileged containers and reverse proxy support.
Nextcloud Winter 26 — finally fast enough to daily-drive
Major performance overhaul. Self-hosters report it’s finally fast enough to replace Google Workspace. Dramatically improved setup experience.