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,
We missed our own 3rd birthday! 10th October 2025 marks three years since I started writing this newsletter. So grateful for all of you along for the ride and all the AI tools keeping us busy for so long đ .
The latest entry in AI browsers is Atlas - A new browser from OpenAI. Atlas would feel similar to Dia or Comet if youâve used them. It has an âAsk ChatGPTâ sidebar that has the context of your page, and choose âAgentâ to work on that tab. Right now, Agent is limited to a single tab, and it is way too slow to delegate anything for real to it. Click accuracy for Agent is alright on normal web pages, but it will definitely trip up if you ask it to use something like Google Sheets.
One ambient feature that I think many people will like is âselect to rewriteâ - You can select any text in Atlas, hover/click on the blue dot in the top right corner to rewrite it using AI.
You can also share your browsing history and access it in ChatGPT as a form of memory. That seems to be the bet to make Atlas sticky, but I havenât been a fan of Memory in general. Searching in Atlas defaults to ChatGPT but has a few panes you can switch to, including one for your 10 blue links, one for images and a couple more. Overall, it feels like a beta version with too many gaps yet to be filled.
Google AI Studio has had a vibe-coding mode since May, but they have upgraded the experience a lot since then. It is one of the fastest ways to build with Gemini models and have Gemini-based features in your apps. The app files are saved to your drive, can be pushed to GitHub, and you can preview and deploy them as well. This is one of the unique features of the developer console of a major AI lab. Related: Web traffic for AI Studio has also been growing rapidly since last year.
NEW: Our âBenâs Bites AI expert matching serviceâ has launched! h/t to my brother, Adam for building it with Shanice. If youâre an AI dev or no-code/automation expert, list yourself and get hired. If youâre a company needing AI projects built, post jobs and connect with the right people. Weâll be trialling a personalised matching service soon too.
đ What Iâm consuming
Use the saw, fear the saw - Donât stop making powerful (and potentially dangerous) tools; rather, we should empower people to use them safely.
Neural audio codecs - how to get audio into LLMs.
Zero Framework Cognition: A way to build resilient AI applications.
AI is making us work more - How the AI boom revived a 996 work culture.
There is no God Tier video model.
As you know, Iâm learning to be more technical so the Solve It With Code course by Jeremy Howard (Inventor of the LLM) & Eric Ries (The Lean Startup) is one course Iâm taking - Iâve learnt a ton from both of them over the years. Their method teaches how to think with AI, not just prompt. Break challenges into small pieces, build iteratively. Starts Nov 3 with 5 weeks of live problem-solving. Bonus: Access to their Solveit platform. Join here.
Also Andrej Karpathy set a challenge: âCan you take my 2h13m tokenizer video and translate [into] a book chapterâ. and Solve It did - read it here.
âď¸ Tools and demos
Speechmatics gives your Voice AI Agents the best ears in AI. Capture who said what in noisy, heavily accented, multi-speaker conversations. Get $200 free đ*
Workflow Debugger by LlamaIndex - Visualise a multi-agent workflow, with human-in-the-loop support and runtime comparisons.
Lovable now integrates with Shopify, allowing you to create a storefront and make it live.
Director - Free app to automate web tasks with a single prompt.
Spiral by Every - Write content youâre proud of with AI. (blog)
ElevenLabs Voice Isolator now supports video - Clean up the audio in your videos to match studio-grade quality.
Claude Desktop app now supports speech input, quick access and more.
Semantic Art - Instead of generating art, discover what humans have already created.
𼣠Dev dish
X finally has a pay-per-use API for developers to build upon. But itâs still too expensiveâwe pull 30-40k tweets per month to look for AI-relevant tweets from our feed, and thatâs easily $200/mo.
GPT-5 Image Mini is now available on OpenRouter.
ACP - A protocol for connecting any editor to any agent.
Chandra - A highly accurate OCR model that converts images and PDFs into structured HTML/Markdown/JSON while preserving layout information.
Scrapybara is sunsetting its product for spinning up remote desktop instances for computer use agents. As a result, they have open-sourced a part of their codebase.
đ° Who got that bag?
Langchain raised $125M at a $1.25B valuation.
Sesame (the viral voice demo makers) has raised $250M Series B led by Sequoia and opened an iOS beta.
Tempo Labs has raised $5M to let design engineers collaborate on production codebases.
đŚ Afters
The context engineering hackathon - Make agents and compete against pre-defined evals, not demo-ware. Nov 15th in SF, sign up here.
Anthropic and Google are in talks for a deal worth tens of billions for compute.
Samsung, Google and Qualcomm released an Android headset, GalaxyXR, with a bunch of Gemini features (at half the price of Apple Vision Pro).
Thatâs it for today. Feel free to comment and share your thoughts. đ
Read about me and benâs bites
đˇ thumbnail creds: @keshavatearth,
Hey Ben! I have been wondering about the privacy implications of AI features in browsers. Say your bank statement or email inbox is open in one of the tabs, and you run a query on something else â like a topic you are researching which are open on other tabs â wouldnât the AI tool have access to all this data which you didnât really mean to reveal?