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,
Iāll be jetting off to Sardinia with my wife Sunday ššāļø (no kids! š«¢) - any recs, hit me up š. On to AI landā¦
ChatGPT has a developer mode now [ link ], and you can use it too. Enable developer mode in Settings ā Connectors ā Advanced ā Developer mode. This allows you to bring any custom MCP server with write actions into ChatGPT. For example, you can use Stripeās MCP to get your accountās data or even create an invoice. Or use Vercelās MCP server [ link ] to take care of deployment.
Letās talk about two new features: one each from Claude and Gemini, and the different ways both companies marketed their launch.
Claude can now create and edit files [ link ] from spreadsheets and documents to PDFs and slide decks. Itās essentially a code interpreter, i.e. Claude can run code in the background to do these file creation/editing tasks. But normal people donāt care about the tech; they care about what the tool can do for you.
Now, in complete contrast, Gemini now supports adding audio files [ link ] to your chats in the app. NO OTHER MAJOR AI APP HAS THIS, and they announced it as āpapercut fixedā š¤¦āļø . They are doing well onboarding people with nano-banana hype, but the excitement for these (not so) little features also adds up. Whoās gonna tell them?
Replit released v3 of their Agent. [ link ] Key upgrades: 1) It can go up to 200 minutes of working autonomously. 2) Agent tests your apps in the browser periodically, like clicking a button, trying to log in, etc. 3) In beta, but Replit Agent can build other agents and automations (powered by Mastra), not just web apps (like build an agent to ping me in Slack 20 minutes before every meeting with research info). Replit also has a design-only mode that builds the frontend and mockups for features (which is much faster) if youāre just prototyping. And it raised $250M [ link ] at a $3B valuation.
Why is running an agent for 200 minutes a thing to celebrate? Isnāt that slow? TLDR; long agent runtimes mean deeper, more complex tasks get done. Itās about capability and autonomy.
Brightwave's state-of-the-art multi-agent research system is the most powerful synthesis engine on the planet. 10k+ documents, long-running asynchronous background agents, fine-grained context control and a flexible, intuitive UI that thinks like you do. API access available. Try Brightwave today. [ link ]*
*sponsored
š What Iām consuming
What not to do when monetising a newsletter [ link ] from our (yes, Benās Bitesā) firsthand experience.
Inside the Man vs Machine hackathon [ link ] - 100+ participants, 6 final projects for a $12,500 top prize. Can you guess which ones used AI to build and which ones didnāt? (non-paywalled article [ link ])
How Factory builds agents [ link ] that help across the entire software development life-cycle.
Shawn Wang (aka swyx)ās thesis for joining Cognition [ link ] (which just raised at a $10B valuation)
20-minute crash course for AI SDK v5. [ link ]
Defeating nondeterminism in LLM inference [ link ] - blog by Thinking Machines (ex-OpenAI CTO, Mira Muratiās new company)
āļø Tools to tinker with
Oboe [ link ] - Use AI to become smarter, not stupid. Course with long reads, audio lectures, quizzes and more.
Voice Remixing by ElevenLabs [ link ] - Change any aspect of a voice (real or generated) like gender, age or accent.
Scheduled Runs in Julius [ link ] - Schedule any analysis to be run with just a single click, and have the results delivered straight to your Slack or email.
Cofounder [ link ] - AI agent that runs your business with you, remembers things, and knows everything about your business.
Google AI Edge Gallery [ link ] - Official app from Google (on Play Store) to run a local model (gemma 3n) on your mobile. (repo [ link ])
Design systems in v0 [ link ] - Define colour schemes and preview light/dark modes for your apps.
Napkin AI [ link ] - Create diagrams/mindmaps that you can actually use from just prompts.
*sponsored
š„£ Dev dish
Fartscroll [ link ] - Makes a fart noise as you open/close your macbook š
Modal Notebooks [ link ] - Cloud-hosted GPU notebook with collaborative editing and GPU swaps.
Veo 3 and Veo 3 Fast [ link ] are now generally available in the Gemini API. The models are roughly 50% cheaper now and support vertical videos and 1080p.
vt - the CLI for Val Town [ link ]. Deploy software instantly as you develop it.
Chroma package search [ link ] - Enable your AI agents to search the source code of your package dependencies.
Web fetch tool [ link ] in Anthropic API - Pre-built tool to get data from any webpage with no extra infra or cost.
I havenāt tried this, but this demo [ link ] looks really cool.
š Charts I saw this week
Simon Willison used ChatGPT to recreate the chart below using US census data.
Parallel (the new AI search company by ex-Twitter CEO Parag Agarwal) shared some Deep Research benchmark results, claiming their search is the best and cheapest.
š¦ Afters
Daniel from BB community is hosting Kieran (from Every) in Toronto to demo his workflow for shipping like a team of 5 but solo [ link ] with Claude Code.
Oracleās stock price jumped by ~35% yesterday, making Larry Ellison the richest man in the world. They revealed new revenue majority of which is coming from OpenAI, according to this scoop from WSJ. [ link ]
Enjoy this newsletter? Forward it to a friend.
Thatās it for today. Feel free to comment and share your thoughts. š
Find me on X [ link ], Linkedin [ link ], or Instagram [ link ]
Read about me [ link ] and benās bites
š· thumbnail creds: @keshavatearth [ link ]
Unsubscribe link