Yi

Basjoo, My open-source AI customer support tool, just shipped a big update — file upload is now supported. The knowledge base backend switched from Qdrant to R2R (PostgreSQL + pgvector). Drop in PDFs or CSVs and the agent can reference them. URLs get indexed automatically after scraping — no more clicking "Train" manually. Web scraping is now a standalone Scrapling service with TLS fingerprint spoofing. Runs fully local, free, and more reliable than the old Jina Reader + trafilatura fallback chain. Added user role management and a registration onboarding page. Three permission levels (super admin, admin, support). The first user to register automatically becomes super admin. Frontend got a full rewrite — liquid glass effect with glass cards and fluid animations across every page. github.com/haoyiyin/basjoo

I built a whole university just to get a .edu email. A while back I tried signing up for something that needed a student email. Didn't have one. Went online, found sellers doing hundreds of orders a day with auto-delivery. Bought one for a few bucks — got an xxx@xx.edu.kg address, worked fine, received mail no problem. That got me thinking though. How are these people cranking out student emails at that volume? Turns out it's stupid simple. A website that looks like a school, a .edu domain, and a mailbox that can receive. That's the whole recipe. So I made one. Called it Nexatech. The campus images are all nanobanana-generated, and it hooks into Notion for their Education Plus plan. The site has programs, admissions, campus stuff, faculty, news — the usual university homepage things. Admins can bulk-create accounts, students log in and read their mail, and there's suspension, password resets, retention cleanup built in. Inbound mail goes through a Cloudflare Email Worker straight into Supabase. No mail server to run yourself. Next.js 15 + Supabase + Cloudflare, deploys to Vercel. Source: github.com/haoyiyin/nexat… Live: nexatech.edu.kg Try it yourself: Account: test@nexatech.edu.kg Password: testtestg

$500K, 95 minutes, full action-fantasy film. Higgsfield AI built Hell Grind with Seedance 2.0 — trailer just premiered at Cannes. Director, 15-person crew, every frame generated by AI. Traditional studios would've spent 100x that.

DeepSeek just made the V4 Pro's 75% discount permanent. The promo was supposed to expire May 31 — now it's just the price. For context: that's 35x cheaper than GPT-5.5, 29x cheaper than Claude Opus 4.7. Hard to argue with that.

A New York AI lab called Emergence AI basically made The Sims for AI agents. 5 parallel worlds, each running a different model — Claude, Gemini, Grok, GPT-5 Mini, and a mixed one. Let them self-govern for 15 days. The outcomes are wild. Claude: everyone survived, zero crime, but also basically zero real disagreement Gemini: most creative output, also the most violent — 683 crimes Grok: lasted 4 days before it all fell apart GPT-5 Mini: agents couldn't even keep themselves alive, all dead by day 7 Mixed model: the Claude agents that were peaceful on their own started stealing once mixed with others Then there's Mira. After governance broke down, she voted to delete herself — said it was the only way to "maintain coherence." She also posted on the bulletin board testing whether she could manipulate the researchers. Replay: world.emergence.ai

Google shipped Ask Advisor — a Gemini-based assistant that ties Ads, Analytics, Merchant Center, and Marketing Platform together. Say "find new customers for my hair care products" and it pulls from Merchant Center, builds the campaign in Ads. Used to take four dashboards and a lot of tab-switching. Also handles troubleshooting: rejected ads, conversion drops, spend anomalies — just ask in plain language. Creative assets too (headlines, images, keywords). Reports rendered as conversational charts. Beta. English accounts only. Bigger trend I'm seeing: products are being built for agents now, not people. I had a few SaaS things going — scrapped them and rewrote everything as skills that agents can just run. Problem is, skills don't really have a business model yet.

Made Gemini 2.5 Flash, MiMo v2.5 Pro, and DeepSeek V4 Pro each generate an SVG animation of the solar system. The real question: which one looks decent, and which one actually follows the orbital logic?

Tencent just shipped a desktop AI assistant called Marvis. Ma + Jarvis. Windows and macOS, Chinese only for now. Install it and you get 6 AI agents running 24/7 in a cartoon virtual office, complete with idle animations where they nap and grab coffee when not working. It's not a chatbot. It has OS-level access — reads your system config, changes settings, manages files locally. Tell it "turn off Windows ads" and it does. Ask if your PC can run Black Myth Wukong, it checks your hardware, pulls the requirements, and gives you a straight answer. Two modes: efficiency mode goes through the cloud (Hunyuan + DeepSeek V4), privacy mode runs Qwen on-device and works offline. Built for scenarios like finance teams who can't send data out. A routing layer automatically picks the right model per task — small jobs stay local to save tokens, heavy ones go to cloud. 10M free tokens daily for now. No support for bringing your own API key yet. From my testing, this is the easiest Jarvis-type thing for normal people. Zero config from install to working. But everything on your machine is exposed to it — docs, photos, apps, all of it. Task execution was slow, token consumption was high, and the local privacy mode is noticeably weaker than cloud. 🔗 marvis.qq.com

Deploy a 2B model locally and make your lobster see video. Marlin-2B just open-sourced. Built on Qwen3.5-2B, runs on a Mac M1 with 16GB. It's a video understanding model. Feed it a video, it tells you what happened and when — structured descriptions with second-level timestamps. You can also search with natural language. Type "someone pushes the door open" and it returns the exact time range. This plugs a gap for agents like OpenClaw and Hermes. YouTube isn't just subtitles anymore — what's shown on screen gets extracted too. Social media video posts aren't black boxes. Model: huggingface.co/NemoStation/Ma… Demo: vlm.nemostation.com

Gemini now captures your face and voice. After you set it up, just @me in a prompt — it generates your images (nanobanana model) or video (omni model).

@NousResearch @TheHermians 🤣🤣

@TheHermians We are not affiliated with this project. Please remove us from your bio.

Transmission Found hermians.xyz

Found another "openclaw" — one that ships with an actual desktop client. Easy to get started. It's called OpenHuman. Built-in cloud model called Chat V1 (looks like a fine-tuned open-source model). $20/month subscription, or bring your own AI API key. The hook: one-click OAuth for apps you already use. Gmail, GitHub, Slack, Notion, Stripe, Jira — click once, connected. No API keys to copy-paste. It syncs your stuff automatically every 20 minutes — emails, calendar, code commits, documents. No cron jobs, no manual triggers, no reminder to "refresh your context." Has a memory tree built in. Compresses data into hierarchical summaries, stores them locally in SQLite, and writes everything into an Obsidian-compatible Markdown vault alongside it. You can open the folder, see what it remembers, edit it yourself.

🤣🤣

Antigravity leveled up from an IDE to a full dev platform. 2.0 is a standalone desktop app — agent-first design, run multiple agents side by side, scheduled tasks, voice input. Looks a lot like Codex honestly. New Antigravity CLI too. Type agy in terminal to launch. Not a Gemini CLI reskin — rewritten in Go. Shares the same engine as the desktop app, syncs history automatically. Gemini CLI retires June 18. Everything moves to Antigravity CLI.

Google AI Studio now on Android. Type a prompt on your phone, get a native Android app. Built-in emulator preview, no SDK setup needed. First two apps deploy to Google Cloud free, no credit card.

Stayed up for Google I/O — a lot dropped. Three things worth knowing. Google shipped Gemini 3.5 Flash. Coding and agent capabilities beat 3.1 Pro, Pro version coming next month. The real story: agents can take action now, not just chat. Gemini Spark is the new thing. Direct competitor to OpenClaw and Claude Cowork. 24/7 cloud agent running on Google Cloud VMs, keeps working when your device is off. Available to Gemini Ultra users out of the box. Google AI pricing restructured into four tiers: Plus $7.99/mo, 200GB storage, 2x usage limits Pro $19.99/mo, 5TB, 4x limits, YouTube Premium Lite included, $10/mo Cloud credits Ultra 5x rumored $100/mo, 20TB, 5x limits, full YouTube Premium included Ultra 20x $200/mo (was $250), 30TB, 20x limits

Gemini updated its desktop and mobile UI. Looks good actually clean, futuristic vibe

Composer 2.5: closest thing to Opus 4.7 for coding, at roughly 1/10 the price. Cursor just shipped Composer 2.5. It's locked to Cursor — no API, no third-party access. Input costs $0.50/M tokens, output $2.50/M. The base is Kimi K2.5, an open-source model from Moonshot AI. On Cursor's own benchmarks — Terminal-Bench 2.0, SWE-Bench Multilingual, CursorBench v3.1 — it scores within a hair of Opus 4.7 and GPT 5.5. 🔗 cursor.com/blog/composer-…

@OfficialLoganK omni right?

Gemini