From Product Hunt
Claude lands in PowerPoint, AI coding agents get rule management, and a personal AI assistant learns to use any webapp.
Claude in PowerPoint — AI-powered presentations
Anthropic brings Claude directly into PowerPoint. Reads existing layouts, fonts, and slide masters. Generate decks from natural language, convert bullets to diagrams, add native charts. Live data connectors. Available on Pro plan.
Straion — Rule management for AI coding agents
Centralized rule management for Claude Code, GitHub Copilot, and Cursor. Agents auto-pick the right rules per task. For engineering teams needing consistency across repos and agents.
Tidy — Personal AI agent for any webapp
Show Tidy how to navigate a site and it creates reusable tools. No code required. Lives in iMessage with memory, cron jobs, and community-shared tools. Like OpenClaw but cloud-hosted with computer use.
SotiTalk — Voice-to-text for technical terms
iOS voice-to-text that handles technical terms like n8n, Docker, and kubectl. 50+ languages, real-time, on-device processing. API + n8n + MCP integration planned. Free, no signup.
ClawCloud — Hosted OpenClaw in 60 seconds
Run OpenClaw without the setup. Hosted agent on WhatsApp, Telegram, Slack, and Discord with 800+ tools. BYOK or use credits.
TIMPs — Open-source memory for AI agents
Structured long-term memory infrastructure for AI agents. Semantic retrieval, project isolation. TypeScript + PostgreSQL + Qdrant. Open source.
From Reddit
AlphaFold gets a next-gen upgrade from DeepMind's drug spin-off, Anthropic's revenue trajectory closes in on OpenAI, and benchmark quality comes under scrutiny.
Scientists are marveling at Isomorphic Labs' latest AI for drug discovery, described as an "AlphaFold 4." The DeepMind spin-off's new system goes beyond protein structure prediction into full molecular interaction modeling.
Unlike AlphaFold 3, this version is exclusive to Isomorphic Labs and won't be publicly released — a strategic shift that has the scientific community divided.
Anthropic could surpass OpenAI in annualized revenue by mid-2026
EpochAI analysis shows Anthropic growing at 10x/year since hitting $1B, versus OpenAI's 3.4x. Estimated crossover around August 2026 at ~$43B annualized. Anthropic's current run-rate is $14B. Community is skeptical about linear extrapolation but the trend is notable.
Qwen team verifies serious data quality issues in GPQA and HLE benchmarks
Following the "DeepSeek-Overclock" forensic audit project, the Qwen team independently confirmed that roughly 40% of HLE answers are dubious and GPQA has significant quality problems. Raises questions about how much weight to give these widely-cited benchmarks.
ByteDance Seed 2.0 benchmarks reveal GPT-5.2 Codeforces elo at 3148
ByteDance's Seed 2.0 evaluation page includes GPT-5.2-high at Codeforces elo 3148. First public GPT benchmark on competitive programming. Google's Gemini 3 Deepthink holds the record at 3455 — a 300+ elo gap GPT-5.3 might close.
Hardware and local inference — relevant if you run models locally.
RTX 5090 VBIOS flash unlocks 300W power limit
Flash the MSI Lighting 800W VBIOS on single-connector 5090s to bypass NVIDIA's artificial power floor. Setting 50% power limit gives 300W max — useful for stacking cards or reducing noise during inference.
First Llama inference entirely on AMD NPU on Linux (Strix Halo)
Llama 3.2 1B running on AMD Ryzen AI Max+ 395 NPU — every operation on the NPU, no CPU/GPU fallback. 4.4 tok/s using the IRON framework on Fedora 43. First publicly documented Linux NPU-only inference.
Qwen3's underrated voice embeddings enable vector arithmetic on voices
Qwen3 TTS turns voices into 1024-dim vectors for cloning. The cool part: you can do math on them — swap gender, pitch, mix voices, create emotion spaces, and enable semantic voice search.
From Reddit
Image model shootout — Z-Image Base vs Turbo vs Flux 2 Klein. Plus new tools for ComfyUI power users.
Z-Image Base vs Z-Image Turbo vs Flux 2 Klein — comprehensive comparison
Side-by-side test across complex prompts, short prompts, and creative scenarios. Tests prompt adherence, creative fill-in, and overall quality. Gallery with detailed breakdowns for each model.
noai-watermark — single diffusion pass strips SynthID watermarks
Open-source tool that removes invisible AI watermarks (SynthID, StableSignature, TreeRing) baked into images by Gemini, DALL-E, etc. A single low-strength diffusion pass fools SynthID. Includes CtrlRegen mode for higher quality. Built for research.
ComfyUI API Liberation — bypass the credit system, use your own keys
ComfyUI's built-in API nodes route everything through api.comfy.org with its own credit pricing. This extension removes the middleman — call 20+ providers directly with your own API keys. Open source.
MonarchRT — 1.4-11.8x attention kernel speedup
New attention method using monarch matrices for GPU kernel speedups. Could significantly accelerate video generation in ComfyUI if integrated.
From Reddit
Mac utilities, self-hosted tools, and side projects. Local-first capture, smart media recs, and a SMART monitoring revival.
Resurf — local-first quick capture and personal library for Mac
Dump ideas, inspirations, bookmarks into a private local library. Coming soon: iOS companion, AI tagging with BYOK, CLI for AI agents to retrieve context, highlights and annotations. $49 lifetime.
Kommand — personalized keyboard shortcut manager for Mac
Build a categorized cheat sheet of shortcuts per app. v2 adds key chords, Vim commands, global overlay, and iCloud sync. Better than CheatSheet or KeyClu because you curate what matters. Lifetime purchase.
Radial 3.0 — circular shortcut launcher with multi-step macros
Pie menu at your cursor for launching apps, inserting snippets, running scripts, and chaining multi-step macros. v3 unifies all menus, adds window management and dynamic text variables. Context-aware per app.
SuggestArr — AI-powered media recommendations for Jellyfin/Seerr
Watches what you play on Jellyfin, Plex, or Emby, finds similar content via TMDb, and auto-requests through Jellyseer/Overseerr. Major update adds AI-powered recommendations. 800+ stars, MIT, Docker.
Storyteller v2.7.0 — reworked transcription for ebook/audiobook sync
Self-hosted platform that merges ebooks and audiobooks into "readaloud" books with sentence-level highlighting. v2.7.0 rewrites the transcription engine for better alignment accuracy.
Scrutiny fork — actively maintained SMART drive monitoring
The original Scrutiny project went dormant. This fork picks up active development with bug fixes, new features, and community feedback. Monitors drives via S.M.A.R.T data with a clean web dashboard.
Earl — CLI/MCP for API integration with AI agents
Hook Claude or ChatGPT into any API or database. One CLI/MCP tool instead of writing custom integrations for each endpoint.
PassportPhotoSnap — free passport photo tool
Pick country, upload photo, auto-crop to official specs with face detection. Background removal, brightness adjustments, print-ready output. Free.
Aagedal Media Converter — fast batch video converter for Mac
Minimalist drag-and-drop video converter optimized for macOS. Built by a journalist for quick turnaround. Free and open source.