SUBSCRIBE | ADVERTISE

October AI news sans hype

A monthly digest of the 1% of AI news and tools that mattered

Was this email forwarded to you? Sign up here.

Howdy wizards,

AI companies pushed shiny new browsers, agents and context systems—and achieved some insane valuations. Meanwhile, smart people are cautioning against hype and workslop.

Here’s what’s brewing in AI.

New tools & product features (part 1)

• OpenAI launched Atlas—a browser with ChatGPT at the core. It sees the pages you’re on and remembers what you’ve seen, and has an agent mode that lets it click around to do tasks. With this launch OpenAI has stepped fully into the agentic browser wars—along with other contestants like Dia, Perplexity’s Comet, Opera’s Neon and the open source BrowserOS—all trying to get a piece of Chrome’s sweet monopoly-situation (and the ad dollars that come with it). Atlas is currently Mac only, but will extend to the lesser alternative (🪟 hehe) and mobile soon.

• Claude now has skills. Skills are basically folders of information (files, code scripts, etc) that you create, and Claude loads them when it’s relevant to the task. Let’s say you’re making a presentation for work. In your arsenal you might have the skills “brand guidelines” and “make presentation”. Claude can now take your input, find the relevant skills by itself, and use them to give you an on-brand slide deck. Available on paid Claude plans.

• OpenAI shipped a bunch of goodies on their DevDay. With the introduction of Apps in ChatGPT, OpenAI aims for the lucrative position of the middleman between you and your apps. They also introduced AgentKit, a set of visual tools that makes it easier to automate workflows with AI, and give your workflow a chat interface (e.g. make your own support agent). See also: my full breakdown of DevDay.

• Google launched Gemini Enterprise, basically bundling its workplace AI into a single $20-30/user package. It comes with a no-code agent builder and admin controls—basically Google’s answer to OpenAI’s AgentKit. However, the interface is more conversational, while OpenAI’s product is more Zapier-like. Google’s product currently has a tighter integration with Gmail, Calendar, etc, but lacks the ability to use MCPs.

IN PARTNERSHIP WITH EMERGENT

Create powerful apps on Emergent

Building full-stack apps just got easier. Emergent lets you create complete applications—with authentication, database, and deployment—using nothing but plain ol' English. No coding required.

Emergent lets you build and ship your app using top LLMs (GPT-5, Claude Sonnet 4.5), while handling complex stuff like Auth and database for you.

Perfect for non-developers prototyping internal tools, or teams that want to ship MVPs fast without wrestling with code.

Start your project now

New tools & product features (part 2)

• ChatGPT is transitioning to a more human-like personality and will get less restrictive for things like erotica (for adult users) in December. Brace yourselves: NSFW AI companions are going mainstream. Grok did this first. The question is, with ChatGPTease on the horizon, will Claudia and GeminiXXX be far behind?

• Google launched an MCP for Chrome DevTools, which lets the AI see what happening in the browser on a technical level. If you’re doing AI-assisted coding with command line tools like Claude Code or Codex, MCPs are a god-send. The exact installation depends on which CLI tool you’re using (you’ll find what you need in the docs).

• NotebookLM from Google got beefed up: now has a 1 million token context window, 6x longer memory and it automatically explores content from multiple angles.

• Perplexity’s Comet browser is now open to everyone (ie you don’t have to beg for an invite anymore).

• Google’s viral image generator Nano Banana is now generally available, and you can now choose which aspect ratios it should use. Are you a nano banana convert yet? What’s proving useful to me is that it’s actually great at editing images, too, not just generating new ones.

Models

• Cursor, the grandaddy of vibe coding tools, got 3 major updates:

  • Composer—their own agent model that’s 4x faster than similarly intelligent models. It’s based on the realisation that devs want smart models that are interactive, ie they don’t take 30 minutes of thinking time to center a div (it completes mosts requests in <30 seconds).

  • They’ve also made it easier to run multiple agents in parallel inside Cursor, without them interfering with one another (powered by Git Worktrees). This includes having multiple models attempt the same problem, and letting you choose the best result.

  • They’ve added voice mode for native text to speech. Just tested this feature and it’s great—will definitely give some competition for tools like Paraspeech and Wispr Flow.

• Google wouldn’t just let Sora 2 slide, so they rolled out Veo 3.1 this month—an improved version of their video model with richer audio, more realism, and better editing controls. You can now add up to 3 reference images, choose start/end frames and extend scenes up to 1 minute.

Industry moves

• OpenAI just finished their transition to a for-profit company, giving Microsoft a 27% stake. They’re now the world’s most valuable private company after a $6.6B employee tender priced the company at a $500B earlier this month.

• OpenAI is also like a stoned teenager on 7/11 on a Friday night, desperate for chips. They’re now going to design 10GWs worth of their own chips in collaboration with Broadcom. It will be a long project, starting in 2026 and aiming for a full rollout by the end of 2029.

Research

• Stanford surveyed ≈1,000 workers about workslop; a new term for purely AI generated outputs that are given to other colleagues to do work based off of. Shockingly, it ends up in wasted time due to people on the receiving end having to decode and redo a bunch of things. As the Buddha says: if you didn’t put effort into writing your workslop, don’t expect others to put time, energy and care into reading it.

• The AI tools that startups spend most money on are OpenAI, Anthropic and Perplexity. Investment firm a16z analysed transactional data from Mercury to show which AI companies are actually getting B2B revenue. Vibe coding platforms and creative tools were also high on the list. Agent-type tools for specific industries were also trending on the charts. In the sea of hype, it’s great to see some hard data like this. But—which of these companies are actually profitable? That’s a different topic entirely.

Latest reviews from whatplugin

Me and Jake put leading AI tools through real-world tests so you can easily compare them and decide what's worth your time and money.

• Emergent (app builder)*: I’ve reviewed Emergent, as well as created a small full-stack app with the platform to show off its capabilities. PS check out the digital swear jar I vibe coded; feel free to use it to track your unfortunate habits.

• Our writer and musician Jake reviewed all the top options for AI audio mastering (the final step of an audio production). LANDR was a step above the rest. Check out the full review.

*Sponsored

Talks & tutorials

• Andrej Karpathy went on the Dwarkesh Podcast. It’s an fascinating episode. Among other things, he tried to dampen some of the expectations around AI agents. This resulted in a bit of a backlash for AI-assisted coding where many established developers took to twitter with some variation of “I told you vibe coding was useless”. My take is that Andrej’s points about LLMs struggling with truly novel code are totally fair—but they don’t “debunk” the usefulness of AI coding tools because the vast majority of business applications involves creating variations of existing solutions, where LLMs do excel as they have abundant training data. Truly original work is a different matter entirely.

• Came across this great example on using Claude Code for sales, even if you’re non-technical. If you’re feeling intimidated by command-line coding tools, seeing someone stumble through using it for their workflow like this is surprisingly helpful.

What I’m actually using

• I’ve finally succumbed to the FOMO and been playing with Nano Banana this month. It’s really good, and I’ve used it to generate this newsletter’s featured image. The main negative I’ve found versus ChatGPT, is that keeps generating hands and feet in weird places, whereas Chat seems to be more accurate/better at adhering to my prompt.

• For the people coding with AI, I recommend Google’s MCP for Chrome DevTools. It has become the second MCP I’m using daily in my development workflow, and saves a lot of time troubleshooting and copy pasting information from DevTools into Codex. The other one is the Supabase MCP, which lets the AI read and take actions on the database I’m using at the backend of the apps I’m creating.

What’s on my radar

• Confession: I haven’t tried Atlas yet. After testing Dia, Comet and the crappy beta of Claude for Chrome, it hasn’t exactly made me lust for more. Someone please give me a compelling reason to try it!

• Claude Skills looks interesting, but I’ve cancelled my Claude subscription recently in favour of ChatGPT Pro and Codex. Might be back to Claude soon though, apparently Codex’ performance is on a decline (last month everyone was going from Claude Code to Codex, and now it’s the other way). Anyways—have you found Claude’s Skills useful so far? If so, for what? I’d love to know.

• The new voice mode in Cursor is actually great and it transcribes while you talk, which makes it really fast. I still prefer Paraspeech for dictation though, because it works across my apps, is fully offline and has a cheap lifetime license (my readers can still get it at 15% off with the code HOWDYWIZARDS at checkout).

THAT’S ALL FOR THIS WEEK!

Affiliate disclosure: To cover the cost of my email software and the time I spend writing this newsletter, I sometimes link to products and other newsletters. Please assume these are affiliate links. If you choose to subscribe to a newsletter or buy a product through any of my links then THANK YOU – it will make it possible for me to continue to do this.