The newsletter for the technically curious. Updates, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.
Hey folks, for the entire day, I was wondering, does anything other than Sora even deserve to be here? OpenAI released Sora 2, with a new social media app called…you guessed it, Sora.
The app has a TikTok-like feed—with videos created by Sora. These videos have a) sound (just like Veo 3) and b) ‘real’ people. The addition of people is the killer feature (called Cameo), where anyone can create a short video of themselves saying a few things, moving their head around and then allow friends (or anyone) to add them to a Sora video.
It’s invite-only and only available in the US and Canada for now, but you can watch the videos, like Sam Altman and OpenAI employees launching Sora, or Pikachu performing an EDM set, anime Jesus, or the hidden Steve Ballmer inside Sam Altman.
Another cool upgrade in Sora 2, beyond realism, audio and Cameo, is the movement of characters. See this video of a skateboard flip, or Theo running in an alley. No sudden jumps or appearing out of nowhere.
Meta also launched a feed of AI-generated videos last week, called Vibes, but the vibes (pun intended) in Sora are much less towards AI-slop, especially because a) the OpenAI team is working hard pushing out funny videos with real people (mostly Sam) and b) the product is much better (yep, I scrolled both feeds for 20-30 mins each).
I am certainly on the ‘no more AI slop’ side of the table with these products - but I’m also aware of ‘no dick pics’ moment of Snapchat, or similar tales that all of a sudden were just part of daily life (in a big way - TikTok started as Musically that was like awful lip-syncs and dances, which still happen on TT today 😅). I guess it’s easy to look at slop as slop now, but I wouldn’t set up camp comfortably on any side; things change a lot.
Thinking Machine Labs (the company by ex-CTO of OpenAI, Mira Murati) released their first product: Tinker. Tinker is a fine-tuning API, built for researchers who want to work with large (>10B) models, test their own algorithm and data, but don’t want to deal with the infra. The API exposes low-level primitives like forward_backward & sample and focuses on LoRA for fine-tuning (vs RFT). (more details)
Claude is now available in Slack, and the “Imagine with Claude” demo is now open for Pro subscribers ($20/mo plan) as well.
Anything (portco) is hiring engineers in SF. They just raised $11M Series A.
OpenAI built GPT for the masses. Brightwave built AI for professionals: full editability, traceability down to the sentence level, and 20x the compute. Their custom subagents can ingest 2000+ documents with zero loss of fidelity - try doing that with a consumer LLM. → Your 7-day trial is waiting.*
*sponsored
Wanna partner with us? Last few slots left for the rest of the year.
🌐 What I’m consuming
The anatomy of MCP authorization - how to add OAuth to your MCP server.
Saving hours by writing release notes for Bun using AI.
The people running Elon Musk’s xAI.
Your Factory AI guide - Building a software development army with Droid.
Rethinking muscle mem as an LLM proxy - as long as a tool gets called, does it matter who called it?
Effective context engineering for AI agents - how does it differ from prompt engineering?
⚙️ Tools and demos
Recipes By Granola - Run advanced prompts across your Granola notes.
Octave 2 by Hume - Emotion-preserving speech generation at fast speeds and low cost. (voice conversion and phoneme editing coming soon)
Pencil.dev - Design mode for your code editor. A canvas that lives inside cursor using MCP.
Faces - The everyday site builder with taste. I like the variety of styles and their focus on “quickly make a website for a project/portfolio” vibe.
Agentforce Vibes - Salesforce now wants enterprises to vibe code.
Riverside now supports chat-based editing for your videos.
Everyday.new - Create workflows that run everyday without dealing with n8n or other complex tools.
Sculptor - the missing UI for Claude Code. Basically a ‘Conductor’ alternative that has raised $200M+ (for some reason…)
An open-source alternative; Emdash (love the name) - supports Claude Code and Droid 🔥 - these tools let you run multiple coding agents in parallel.
(iydk Factory released Droid last week - I’m head of dev rel - and you can get 40M tokens free using this link)
🥣 Dev dish
Mintlify now sends markdown by default to AI agents.
Mastra AI shipped a new models router covering 600+ models from 45+ providers from a single API.
🌌 On the frontier
Frontier for AI labs right now is building AI for science - we are getting new labs dedicated to maths or biology every day now. Some of the new ones:
Periodic Labs - Their goal is to create an AI scientist for physical sciences. Raised $300M led by a16z. Formed by prev. VP of post-training at OpenAI and material science lead at Google DeepMind.
Axiom is building an AI mathematician with $64M seed and a team of Meta FAIR researchers and Stanford PhDs. Founder and coverage in Forbes.
Another area is making AI companions actually work. The core theme here has been reducing tech features and capitalising on one core differentiator, like style or peace of mind.
Taya - A jewellery-like necklace with recording/note-taking features, and subtle app notifications when you need them. (video)
Meadow - Leave your phone behind when going out and take meadow. Half the size of a phone with full LTE service, and essentials like Uber and Strava. (video)
🍦 Afters
DoorDash now has a delivery robot called Dot.
Built with Claude Sonnet 4.5 challenge - 4 winners to get a Claude 20x MAX plan for a year and $1000 in credits.
Perplexity AI acquired Visual Electric, a creative tool for teams. Shame, I liked Visual Electric 😅
That’s it for today. Feel free to comment and share your thoughts. 👋
📷 thumbnail creds: @keshavatearth