From Product Hunt
Computer use agents and Mac notch utilities lead today's launches.
Ara — Computer Use Agent in Your Notch
BYOLLM computer use agent that lives in your Mac's notch. Hit a hotkey, give it a task, watch it control your cursor, click apps, and fill forms. 1,000+ integrations with Gmail, Slack, Notion, GitHub. Think Wispr Flow but with actual screen control.
Monid 2.0 — OpenRouter for Agent Tools
Plug in once and your agent can discover, compare, and pay for any of 200+ tools on demand. Social scrapers, search APIs, ecommerce data, lead gen. Unified API for agent tool access.
SuperIsland — Dynamic Island for macOS
Free, open-source Dynamic Island with a Raycast-style extensions SDK. Music player, calendar, weather, file tray, plus WhatsApp, Linear, AI agent status, and teleprompter extensions. Growing extension library.
From Reddit
A Fields Medalist crosses a threshold with GPT-5.5, DeepMind sets a new math benchmark record, and Mythos gets its first independent evaluation.
Fields Medal winner Timothy Gowers published a blog post about using GPT-5.5 Pro to solve open mathematical problems. He described crossing an "interesting threshold" where the model's mathematical reasoning has become genuinely useful for research-level work.
Gowers believes mathematical research will face a "crisis" very soon at the current rate of AI progress. A sobering assessment from one of the most accomplished living mathematicians.
DeepMind AI Co-Mathematician Scores 48% on FrontierMath Tier 4 — New SOTA
Google DeepMind's agentic AI co-mathematician set a new high score on FrontierMath Tier 4, beating GPT-5.5 Pro (39.6%) and GPT-5.4 Pro (37.5%). Multi-agent system with coordinated parallel workstreams, hypothesis tracking, and LaTeX write-ups. The base Gemini 3.1 Pro scores just 19%.
GPT-5.5-Cyber — Fewer Guardrails for Vetted Security Teams
OpenAI launched GPT-5.5-Cyber in limited preview for government, critical infrastructure, and security vendors through the Trusted Access for Cyber (TAC) program. Defenders get a less restricted version for bug hunting, malware analysis, and reverse engineering. Advanced Account Security required from June 1.
METR Evaluates Early Claude Mythos — 50% Time Horizon 16+ Hours
Update: METR evaluated an early version of Claude Mythos Preview during a limited window in March 2026, estimating a 50% time horizon of at least 16 hours (95% CI 8.5-55 hrs). Only 5 of 228 tasks in their suite exceed 16 hours — they're hitting the ceiling of their evaluation framework.
Claude for Excel, PowerPoint & Word — Now Generally Available
Claude add-ins for Microsoft 365 are now GA (Outlook in public beta). Key innovation: Claude carries full conversation context as you move between apps. Triage email in Outlook, draft in Word, analyze in Excel, present in PowerPoint — all in one continuous session.
DS4 by antirez — DeepSeek V4 Flash Inference for 128GB Macs
Salvatore Sanfilippo (creator of Redis) released DS4, a DeepSeek 4 Flash-specific inference engine optimized for 128GB MacBooks. Purpose-built for the V4 Flash architecture rather than being a general engine.
Quick hits from LLMs and AI.
AI identifies 10,000+ new exoplanet candidates, potentially tripling known alien worlds
Scientists used AI to find candidates that traditional methods missed entirely.
DeepSeek now targeting $7.35B funding round, V4.1 next month
Update: Significantly higher than the $3-4B initially reported. Founder plans to contribute maximum allowable amount.
AI2 EMO — 14B MoE with document-level expert routing
Novel architecture: experts cluster around domains (health, news, etc.) instead of surface patterns. 1B active params.
Ring 2.6 1T — InclusionAI's latest trillion-parameter model
Ant Group's successor to Ring-1T-2.5. Available for free on OpenRouter. Open weights expected.
Helix 02 robot tidies a bedroom autonomously
Picks up items, folds, and organizes without intervention. One of the most practical humanoid demos yet.
From Reddit
Real-time video processing hits 30 FPS, and aggressive quantization makes local video gen more accessible.
Open-source pipeline that processes webcam video in real-time at 30 FPS with ~0.2s latency on a single RTX 5090. Uses a spatial-aware KV cache that only recomputes tokens where motion occurs, plus frame interpolation for smoothness.
Free and runs locally. A milestone for real-time AI video processing on consumer hardware.
LTX-2.3 PolarQuant Q5 — 88% Size Reduction, Near Lossless
New quantization method reduces LTX-2.3 22B to 88% of original size while maintaining 0.9986 cosine similarity. Makes local video generation much more accessible on limited VRAM.
Distilled Qwen3-1.7B text encoder saves 21% VRAM for Z-Image
Finetuned 1.7B adapter imitates Qwen3-4B hidden states, dropping peak VRAM from 21.35GB to 16.76GB with nearly identical output. Generation time also slightly faster.
From Reddit
macOS utilities and self-hosted updates. Good day for small, focused tools.
FileFluss — Open Source Multi-Cloud File Manager for Mac
Free, Mac-native file manager for searching, moving, and syncing files across Google Drive, Dropbox, OneDrive, and more from a single interface. From the creator of NetFluss (9,000+ downloads).
NaiveChat — Raycast AI Chat Alternative for Power Users
Hotkey-driven AI chat built for extreme efficiency and intensive prompt engineering. Designed by a Raycast power user who hit the AI chat's limits.
Technitium DNS Server v15 — now with OIDC single sign-on
Long-requested SSO support. Works with Authentik, Keycloak, and other OIDC providers.
Kavita v0.9.0 — CVE security fix, update now
Two medium CVEs patched. If you run this reading server, update immediately.
AutoFocusMonitor — One less click per monitor switch
Free macOS utility that auto-focuses the last window when you move between monitors.