The newsletter for ai builders of all levels. Mini-tutorials, tool reviews, and lay of the land from an exited founder turned investor and forever tinkerer.
Hey folks,
Let’s talk about voice.
It feels like there is a bit of a barrier to sitting at your computer and talking to it vs sending voice notes to friends. Voice is quicker. But clarity is better when it’s written down. So being able to see your own speech would be helpful for clarifying your thoughts.
But the problem with voice at the moment is it's an extra cost, extra latency, and can you get the accuracy? Whenever I use ChatGPT Voice Mode, it transcribes what I say into Welsh because I am Welsh and I've got a faint Welsh accent, but I can't speak a word of it. Not particularly helpful.
AssemblyAI have been partners of ours for a while now, and their latest Streaming Speech-To-Text API170 is here. So think of streaming like when you ask ChatGPT or Claude a question, the output starts appearing a few lines at a time. That’s a much nicer experience instead of waiting for its full answer. For voice tools, streaming is hard because what your next few words are changes the interpretation of what you’ve already said.
AssemblyAI's API170 now balances that with insane speed, cost, and accuracy. So you can build an experience that combines what your users are saying and the transcription that they see, without it feeling broken. I tested it, and did a little review, and I think you should check it out if you're building anything with voice or you want to add voice capability to your applications or workflows. I'm going to be adding it to a bunch of mine where I actually do like to see what I'm saying, which my current voice tools don’t do!
Or, watch on YouTube110.
ChatGPT now has a study mode233. Instead of just doing the work, it guides students towards learning and finding answers on their own. Currently, it’s just a special prompt268 to nudge model behaviour, but OpenAI plans to bake that behaviour into its models over time.
NotebookLM can now make Video Overviews378. It pulls in images and diagrams from your documents, creates new infographics and visuals as slides. I’ve already seen a few startups on the beat of “make a 3blue1brown-type animated video on a topic”, so expect more of such tools in the coming weeks.
And AI mode (in Google Search)86 is getting some new features like file uploads, search with live video and images, and a writing canvas. Google’s promoting both these features from a studying pov, you can guess why.
Attio172 is the AI-native CRM for the next generation of teams. Sync your email and calendar, and Attio instantly builds your CRM—enriching every company, contact, and interaction with actionable insights in seconds. Join fast growing teams like Granola, Flatfile, Modal, and more. Start for free today.172*
Mark Zuckerberg released a new memo called “Personal Superintelligence319”. I loved reading Om’s breakdown of it638. Also, a scoop from Wired claims Zuck made offers worth a billion dollars103 (spread over a few years) to many people at Thinking Machines Lab (Mira Murati’s company), and everyone declined.
Behold the first AI-native investment bank408.
Cursor’s lead designer built an operating system325 with Cursor.
Logan Kilpatrick’s latest podcast with Matan Grinberg, CEO of Factory AI153 (I’m an investor).
YouTube: youtube.com/watch?v=o3rE6cLVOH4153
Airia103 - The enterprise AI platform with built-in governance and security. Deploy agents across teams while maintaining enterprise compliance.*
Snaptrude248 - Instantly turn messy docs into customizable requirements with an end-to-end platform for concept design.
Gumboard452 - A free, real-time sticky note board to keep track of your tasks. (it’s open-source37)
Magic Patterns255 - The design tool to create and prototype new features for your product.
Ollama209 has a desktop app to use local models with a ChatGPT-like interface.
Eigent212 - A team of AI agents collaborating to complete complex tasks in parallel.
Chilled Sites317 - Super simple text-to-app builders with a TON of features (great for non-technical folks) - Paul’s shipping like crazy.
Parse.bot307† - Turn any website into an API.
Terragon277 - Background agents for Claude Code. I’m using it a lot for firing off lots of agents, especially on mobile.

Want customised help from AI devs for your company’s project? We’re trialling a matching service connecting companies with AI experts. Register your interest here21.
Exa Fast83 - A search API from Exa with <500ms latency (40% faster than Google Search).
Claude Code119 can now work in multiple directories in a single session.
Crush by Charm108 - A glamorous coding agent in your terminal (works with all major model providers)
AutoRL156 - Write a single sentence to train a task-specific model.
LangExtract159 - OS Python library from Google to reliably extract information from documents.
Applications for the latest batch of HF063 (residency for ex-founders) are now open.
First-ever Claude Code Office Hours146 - today, 3 pm PDT.
Anthropic is looking to raise at $170B valuation41, while Meta just increased its market cap by a similar number in just after-hours trading yesterday (puts all those offers from Zuck into perspective)
Also, OpenAI announced Stargate Norway57, its first AI data centre initiative in Europe.
That’s it for today. Feel free to comment and share your thoughts. 👋
📷 thumbnail creds: @keshavatearth17 - I remembered Asimov’s “The Fun They Had” when coming up with the idea for today’s thumbnail.
Changing how they'll learn
found my voice
3,104 readers clicked at least once. 43 links, 43 with clicks. Heat is relative to the most clicked link in this issue.
Substack · essay