😸 Google's new AI actually controls your computer

Welcome, humans.

Ever wonder which AI chatbot would crush you at poker? Starting October 27, ChatGPT, Claude, Gemini, Grok, and DeepSeek will face off in the first-ever all-robot cash game at PokerBattle.ai (not sponsored, we just find this hilarious).

They'll play thousands of hands of no-limit hold'em with $100K play-money bankrolls, and you can watch every hand in real time, complete with each AI's reasoning for every decision.

Oh, and the entire platform was built by a recreational poker player who isn't even a developer. He coded it… using AI.

That’s v impressive. Meanwhile, the rest of us be like…

Your face when: you become too dependent on ChatGPT…

Here’s what happened in AI today:

1. Google released Gemini 2.5 Computer Use.
2. OpenAI's DocuGPT post caused DocuSign's stock to drop 12%.
3. IBM partnered with Anthropic to integrate Claude into its software.
4. Google Quantum AI scientists won the 2025 Nobel Prize in Physics.

Advertise in The Neuron here

P.S: Missed our recent podcast with OpenAI’s Ahmed El-Kishky on their ICPC “coding Olympics” win? Their system (GPT-5 plus an experimental reasoning model) solved all 12 problems, and at DevDay, OpenAI introduced apps inside ChatGPT and AgentKit/Agent Builder, echoing that reasoning-first approach in new developer tools

Google Built a Browser Agent That's “50% Faster” Than the Competition.

Computer use models (where AI, get this, controls your computer) are janky atm. They misclick, hallucinate what's on screen, and close wrong tabs. Anthropic's demos show Claude once stopped a screen recording mid-filming and browsed Yellowstone photos during a coding demo. Historically, not reliable for serious work.

But now Google's entering the ring with Gemini 2.5 Computer Use, and per third-party Browserbase testing, it's preeeetty good.

How it works: You give the AI a task. It screenshots, figures out what to click or fill, executes, then screenshots again to see what changed and decides the next move. Repeat until done.

Performance wise, Browserbase ran 200+ experiments totaling 4,000 browser hours testing all three models on industry benchmarks. Google's version is often “50% faster” (according to Poke.com), more accurate, and cheaper than Claude and OpenAI (model card).

• Mind2Web (web navigation): Google hit 69% success versus Claude Sonnet 4.5's 53% and OpenAI's 46%.
• WebVoyager (complex multi-step tasks): Google consistently outperformed both.
• Latency: Google maintains lowest response times while delivering higher accuracy. Both faster AND smarter.

Why Google's different: Most computer use models can't accurately count pixels (like large language models struggle counting letters). Google trained specifically on pixel precision, so it knows where to click. They also optimized for parallel actions, meaning the AI executes multiple steps simultaneously instead of waiting for each.

Google's using this internally, too: their payments team implemented it to fix broken UI tests, recovering 60%+ of failures that took days. Also used for Firebase testing, Project Mariner, and powering AI Mode in Search.

What people have tested it for:

• Automatically filling repetitive forms (goodbye, data entry).
• Testing websites by simulating user behavior.
• Organizing messy project boards by dragging tasks.
• Booking appointments across multiple pages.
• Gathering research from various websites.

The demos look slick (here’s one). One shows the AI navigating a pet spa website, pulling customer info, adding it to a CRM, and scheduling a follow-up, all autonomously.

Google built safety guardrails so the model asks for human confirmation before risky actions like purchases or bypassing CAPTCHAs.

How to try it: Easiest way = Browserbase demo environment. To build your own agent, use Gemini API with model gemini-2.5-computer-use-preview-10-2025 and implement an execution loop using Playwright. Check docs here or reference implementation on GitHub.

FROM OUR PARTNERS

Meet Bolt.new

Most vibe coding tools are toys. Fun at first, but often get stuck in endless error loops, infrastructure headaches, and projects collapsing under their own weight.

Meet Bolt.new the #1 professional vibe coding platform, trusted by product teams at 72% of the Fortune 500.

Bolt puts the most powerful coding agents and enterprise-grade infrastructure directly in your browser:

• 98% fewer error loops
• 1,000× bigger projects
• Production-ready backend from day one

Pro-level power. Simple and intuitive chat interface.

Let Bolt do the heavy lifting so you can focus on your vision instead of fighting errors.

Try Bolt v2 today. Build without limits.

Prompt Tip of the Day

Want to never go to the DMV and wait in line again? We tested this prompt with ChatGPT Agent, and it worked like a charm.

I'm trying to schedule a DMV appointment for [example: my wife to Apply for a REAL ID]. We live in [your City]. I'm willing to drive 150 miles for the appointment, but it must be in [your State]. Can you find an appointment slot for a [reason, for example: REAL ID] on [today’s date]

Corey fired this up in St. Louis and was able to get an available appointment that day, and many have open appointments (especially in smaller areas). There’s DMVs everywhere; as long as it’s in your state, you can use it. You’re welcome.

Treats to Try

1. Google's Opal lets you build AI mini-apps (like OpenAI’s builder tool or n8n) with plain language and visual workflows, regardless of coding skills, now expanding to 15 more countries.
2. Hands Off uses your webcam to detect when your hand moves toward your face to bite your nails or pick your skin, then pops up an alert to stop you in real-time (love this idea… the founder created it to use it on himself, and it worked!).
3. Freshly Squeezed resizes, crops, and converts batches of images (like turning 50 HEIC phone photos into optimized WebP files for your site) in one drag-and-drop (Mac only rn; love this too bcuz a creator made it for himself).
4. Orchestra organizes your workspace by tasks instead of channels, so “Fix checkout bug” becomes its own chat room with only the 3 people working on it, including the task details, files, and calls all in one thread.
5. Hunyuan Thinking analyzes your images and videos to answer questions about them, and can even "think through" complex visual problems by cropping, zooming, drawing on images, or searching the web for additional context—try it here via the dropdown model picker.
6. Repo Prompt lets you visually select files from your codebase to build precise AI prompts and syncs that context across Cursor, Claude, and other tools—free trial, then $14.99/month.
7. DeepLearning.AI has a new course on agentic AI teaching four key design patterns: Reflection (where agents self-critique their output), Tool use (where AI decides which functions to call like web search or email), Planning (which breaks tasks into sub-tasks), and Multi-agent collaboration (where multiple specialized agents work together like employees).

Around the Horn

Inspiration for above video; great infographic of the data form this study

• The Information reported that Oracle's AI cloud business showed razor-thin profit margins (14%) despite projecting $381B in revenue from GPU rentals, which caused the stock of it and similar companies to drop a bit.
• OpenAI's recent blog post about DocuGPT triggered a multi-billion dollar SaaS stock selloff, with DocuSign falling 12%.
• MPA, the US Film industry’s main lobbying group, says OpenAI needs to take “immediate and decisive action” to prevent copyright infringement on Sora 2.
• Zelda Williams, daughter of late comedian and actor Robin Williams, posted to Instagram to request people stop sending her AI deepfakes of her dad; this is possible due to a quirk of US law that says its not illegal to “libel the dead”, so deceased public figures are allowed through the censors.
• Deloitte deployed Claude AI to its 500K employees while at the same time had to refund A$439K (that’s Australia dollars, for US readers) for an AI-generated report with errors.
• Anthropic announced plans to open a Bengaluru office in 2026 and is discussing a partnership with Reliance Industries, as India becomes its second-largest market.
• IBM partnered with Anthropic to integrate Claude’s large language models into IBM's enterprise software portfolio, starting with an AI-first IDE.
• Google Quantum AI chief scientist Michel Devoret, ex-Google quantum lead John Martinis, and UC Berkeley's John Clarke won the 2025 Nobel Prize in Physics for discovering macroscopic quantum mechanical tunneling in The Nobel Prize for physics is awarded for discoveries in quantum mechanical tunneling electrical circuits that laid the foundation for quantum computing.
• Check out Rate Limited (new podcast from Ray Fernando!) with super valuable insights from three AI engineers sharing their hands on experience and practical tips on coding w/ AI—you gotta check them out, these guys are really cracked.

FROM OUR PARTNERS

Peak season checklist

Get ahead of the holiday rush with Gladly’s complete peak season prep checklist. From optimizing customer touchpoints and training your team to enabling AI across every channel, this guide helps you stay organized, scale support, and deliver exceptional experiences when volume spikes.

Download the checklist

A Cat’s Commentary