Happy Sunday, !
Welcome back to your weekly AI news roundup.
In case you missed it, here’s this week’s Thursday post:
Some Prompts Should Just Be Tools
TL;DR If you keep asking AI to do the same mechanical task, you should instead turn it into a standalone tool that needs no chatbot or agent at all.
Note: If you’re consistently missing out on my emails, remember to check your “Promotions” tab and mark whytryai@substack.com as a “Safe Sender.”
Here’s what happened in AI last week:
👩💻 AI releases
Amazon now lets you text Alexa for Shopping a description or photo of what you need, and Alexa will find the item and buy it for you.
Anthropic launched Personal Finance in Claude, letting you connect your bank and card accounts and ask about your spending in plain language.
Cognition released SWE-2 Flash, a fast, low-cost coding model that holds its own against much bigger open models.
Cursor can now build functioning apps with their own database that you can share by simply sending a link.
Google news:
Gemini Omni 1.2 Flash can make 15-second videos that you can refine over several rounds by just talking to it.
Google Vids now lets anyone with a free Google account make up to three 15-second AI videos a month.
Meta news:
Voice chats in the Muse app let you talk to expressive avatars that react to your tone and let you interrupt them mid-sentence.
Spaces in Muse lets users build custom mini-apps like a gift tracker or a daily word game just by describing them.
Netflix rolled out a new AI Discover tab that suggests shows by theme and remembers when you reject a recommendation so it doesn’t show up again.
OpenAI rolled out Images 2.6 in ChatGPT, which can turn your photos and prompts into animated GIFs that you can edit frame by frame.
Pangram launched Pangram Image 1.0, an AI image and video detector that can highlight exactly which parts of a picture were generated. (Try it for free.)
Source: Pangram Runway launched Dev plugins for Claude Code, Codex, and Cursor, so your coding agents can build apps on top of Runway’s video models.
SpaceXSI gave Grok Bot the ability to make phone calls on your behalf. (Also, can you guys settle on a single company name already?! My head hurts!)
🔬 AI research
Apple will add Live Rewind and Siri Recap to Apple Watch to show a transcript of the last 15 seconds or summarize conversations, now in developer beta.
OpenAI is going to merge the Chat and Work tabs in ChatGPT, so you can move from chatting to building stuff in one place, without having to switch.
Selendis is opening early access to Archon, a video model that can cut raw footage into a finished edit in minutes while learning each creator’s style.
Tencent announced Hy4, its first open model that can understand images and operate your computer.
📖 AI resources
“Claude 2.0” [VIDEO]: A great round-up of new Claude features by Futurepedia:
“HackerTube” [TOOL]: Andrej Karpathy’s project that collects YouTube videos recommended by Hacker News users in one place (built with Claude in 2 hours).
“How Much Does Reasoning Effort Affect Score and Cost?” [STUDY]: a deep dive into the quality vs. cost trade-offs by Vals AI.
“LifeVal” [BENCHMARK]: OpenAI’s benchmark that tests AI agents on everyday errands like budgeting, finding childcare, handling government paperwork, etc.
“StateBench” [BENCHMARK]: Allen Institute for AI’s benchmark for how well AI models keep track of details that change over the course of a long document.
“We may release models we can’t control” [VIDEO] another solid look at the state of things by AI Explained:
🤦♂️ AI fail of the week
I asked for a hype bro chatting to his AI agents. Surprisingly accurate. AGI is here!
🔐 All my paid goodies in one place
Take a peek at what’s behind the paywall:
