@ai-reskill
Founder of AI Tool Gallery (Japan) ๐ฏ๐ต Building AI-native web products & directories. Solo founder experimenting with AI agents. Hosting AI events in Kansai. Follow for AI tools, experiments & builds. #AI #IndieHacker #AItools #BuildInPublic
Requires Google AI Ultra ($250/mo). Not cheap for personal use, but document-to-video for training/education is genuinely useful.
blog.google/innovation-...
NotebookLM now generates animated video summaries from your uploaded documents. Google calls it "Cinematic Video Overviews" โ powered by Gemini 3 + Veo 3. It's the video version of those AI podcast summaries.
All vulnerabilities reported to Mozilla, patches underway. One of the most concrete AI security results so far โ real CVEs, not benchmarks.
techcrunch.com/2026/03/06/...
Claude (Opus 4.6) found 22 vulnerabilities in Firefox over 2 weeks โ 14 rated high severity, including zero-days. Cost: ~$4,000 in API credits. Human researchers would typically take months for comparable work.
Interesting angle: since the 2023 Hollywood AI strikes, an A-list actor building an AI company and partnering with a major studio is a new dynamic.
variety.com/2026/film/n...
Netflix acquired InterPositive, Ben Affleck's AI filmmaking startup. The company builds AI tools for script analysis, shot planning, and post-production. Affleck stays as senior adviser.
Limited apps for now, but AI agents handling tap-by-tap phone interactions is a significant UX shift. Accessibility implications too.
9to5google.com/2026/02/25/...
Gemini can now operate Android apps for you. "Book me an Uber home" โ opens app, enters destination, books ride. Also works with DoorDash, Grubhub. Rolling out March on Pixel 10 and Galaxy S26.
Available to Plus/Team/Enterprise users. Pro version at $200/mo. Thinking variant via API. Free tier timing TBD.
techcrunch.com/2026/03/05/...
OpenAI released GPT-5.4. Big changes: native PC control (mouse + keyboard), 1M token context, and a Thinking mode that shows reasoning steps. 33% fewer errors vs GPT-5. The PC operation is preview-stage but the direction is clear โ AI doing desktop tasks directly.
Apple Music adding AI transparency tags โ 4 types covering artwork, tracks, composition, and music videos. Voluntary self-reporting by artists/labels. First major streaming platform to do this.
techcrunch.com/2026/03/04/...
Full before/after comparison with real outputs for emails, summaries, code, brainstorming, translation, and more:
ai-gallery.jp/chatgpt-pro...
Tested 8 common work tasks on ChatGPT with vague vs specific prompts. The output difference is striking.
"Write an apology email" โ generic template.
Add who/what/why/tone/length โ send-ready email.
Same AI, same task, completely different results.
Full comparison with real outputs and prompt tips:
ai-gallery.jp/chatgpt-cla...
Tested ChatGPT, Claude, Gemini with the same 5 prompts.
Same creative writing prompt โ ChatGPT: horror, Claude: comedy, Gemini: sci-fi. Same email โ Claude: 500 chars, Gemini: 1,100 chars.
No single "best AI" in 2026. Match the tool to the task.
The community really took shape over these months with repeat participants. This chapter ends, but we'll keep building these spaces.
prtimes.jp/main/html/r...
Final event of Kobe City's student digital talent project "Digi-Katsu" on March 14th.
7 months of monthly AI/digital events โ from AI search to hackathons. The finale focuses on going global: practitioner talks + a Gemini portfolio workshop.
Free, beginners welcome.
The most impactful AI updates aren't always about benchmarks. "It stopped lecturing me" might matter more for daily users than any performance number.
openai.com/index/gpt-5...
ChatGPT got a meaningful update: GPT-5.3 Instant fixes the "overly preachy" response problem. Fewer unnecessary refusals, less moralizing, hallucinations down ~20-27%.
Free users get 10 messages/5 hours before it falls back to mini.
How to spot AI-generated videos (via NBC):
โข Very short clips (10-15 sec)
โข Background anomalies
โข Misspelled text on signs
โข Too-perfect quality (no shake/noise)
โข Single camera angle only
โข Adam's apple doesn't move during speech
Worth knowing as AI video improves fast.
Local AI keeps getting more hardware support. Apple putting AI circuits inside GPU cores (not just a separate Neural Engine) is a meaningful architecture shift.
apple.com/newsroom/20...
Apple's new MacBook Pro has a "Neural Accelerator" in every GPU core โ 40 AI circuits running in parallel on M5 Max. AI image gen is 8x faster vs M1 Max, 24hr battery, up to 128GB memory.
Starts at $2,199 (M5 Pro) / $3,899 (M5 Max). Shipping March 11.
The distinction matters: Claude found a construction that worked but couldn't prove why. Knuth could. AI as a "discovery tool" vs. "research partner" โ the line is becoming interesting.
www-cs-faculty.stanford.edu/~knuth/pape...
Donald Knuth (88, Turing Award winner) published a paper called "Claude's Cycles." He'd been stuck on a graph theory problem for weeks โ his friend fed it to Claude Opus 4.6, which found a solution in ~1 hour.
Knuth wrote the proof, named the paper after Claude, but is the sole author.
The practical angle: a 6.6GB model running offline on a laptop, no data leaving your device. Local AI is becoming a real option for privacy-conscious users.
venturebeat.com/technology/...
Alibaba's Qwen3.5-9B is beating OpenAI's gpt-oss-120B on benchmarks (GPQA: 81.7 vs 71.5). But note: gpt-oss-120B is a MoE model that only runs 5.1B params per inference, so the gap is less dramatic than it sounds.
All 4 sizes (0.8B-9B) are Apache 2.0, available on Ollama.
Useful for quick fixes or brainstorming when typing feels slow. "Fix this error" or "add tests for this function" โ just say it.
Claude Code now has voice mode. Type /voice to enable, long-press spacebar to speak, release to process.
Voice transcription is free (doesn't count against usage), but AI responses still follow normal rate limits. Rolling out to ~5% of users now, expanding over weeks.
Timing is notable: Claude hit #1 on the App Store last week after the Pentagon standoff. This free Memory rollout looks like a play to retain those new users and lower the switching cost from ChatGPT.
Claude's Memory feature is now free for all users. Previously paid-only.
Anthropic also launched an import tool โ paste a special prompt into ChatGPT or Gemini, it exports your stored memories, then you paste them into Claude. Simple switch.