🤖 AI Agents Can Now Use Your Computer, Here's What Changes Today.

Hey, it's Oliver, here's your AI update for this week!

In today's issue:

• OpenAI's Operator agent can now book your flights, order groceries, and fill out forms for you
• Google's Gemini 3 just launched with agentic features that handle multi-step tasks across your apps
• Why Kosmos AI completing 6 months of research work in 12 hours changes everything about scientific discovery
• And more...

🛠️ Tool of the Week: Kosmos AI Scientist (Not Sponsored)

What it is:

An autonomous AI researcher from Edison Scientific that reads 1,500 scientific papers, executes 42,000 lines of analysis code, and generates fully cited research reports—all in a single 12-hour run.

Why it matters now:

Launched this month and already praised by Sam Altman as "one of the most important impacts of AI." Unlike ChatGPT, Kosmos actually conducts complete research cycles autonomously—planning tasks, analyzing data, searching literature, and generating hypotheses.

Three ways to use it:

• Give it a research objective – Provide a dataset and goal like "find how neurons protect themselves during hypothermia" and it works autonomously for 12 hours, running parallel analyses and literature searches.
• Verify traceable conclusions – Every statement links directly to code output or original papers, so you can check its work. About 80% of conclusions are accurate, with data analysis being most reliable.
• Accelerate research cycles – Scientists report it completes 6 months of work in one run, finding novel insights across metabolomics, materials science, and neuroscience.

The catch: 1 in 5 conclusions needs human verification. It's strongest at data analysis and literature review, weaker at synthesis.

Pricing: Free for academics. Commercial pricing for enterprise.

Bottom line: This is AI doing the exhausting grunt work—reading thousands of papers and running endless analyses—so researchers can focus on creative discovery.

🤖 1. OpenAI's Operator Agent Actually Uses Websites For You (Not Just Searches Them)

Why This Matters

• Operator doesn't just find information—it takes action. It clicks buttons, fills forms, types text, and navigates websites like a human would.
• It's powered by a new "Computer-Using Agent" model built on GPT-4o that combines vision with advanced reasoning.
• Unlike traditional APIs, it works with any website's front-end, meaning it doesn't need special integrations to use services.

The Reality Check

Currently available only to ChatGPT Pro subscribers ($200/month) in the US as a research preview. OpenAI plans to expand to Plus, Team, and Enterprise users once safety testing is complete. The agent can get stuck on complex interfaces, CAPTCHAs, or password fields and will ask you to take over when that happens.

The Practical Impact

• Book concert tickets by searching venues and completing checkout
• Fill online grocery orders from handwritten lists uploaded as photos
• Make restaurant reservations for specific dates and party sizes
• Compare products across sites and complete purchases

OpenAI partnered with Uber, DoorDash, Instacart, OpenTable, Etsy, and StubHub to ensure Operator respects their terms of service.

Bottom line: This is the shift from AI as information retrieval to AI as action-taking. When agents can navigate websites like humans, the definition of "using the internet" fundamentally changes. Expect booking travel, comparison shopping, and form-filling to become background tasks you delegate entirely.

🕊️ 2. Kosmos AI Does 6 Months of Scientific Research in One Day

Why This Matters

• Kosmos reads approximately 1,500 research papers and executes 42,000 lines of code per run
• It doesn’t just summarize existing research—it generates original hypotheses, tests them, and makes new discoveries
• Seven genuine scientific discoveries so far, including three that independently reproduced unpublished findings and four that contributed novel insights

The Reality Check

Sam Altman called it "exciting" and predicted it would be "one of the most important impacts of AI." But 79.4% accuracy means about 1 in 5 conclusions still requires expert verification. It sometimes pursues statistically significant but scientifically irrelevant patterns.

Built by FutureHouse (backed by ex-Google CEO Eric Schmidt) and now commercialized through spinout Edison Scientific.

The Practical Impact

• Metabolomics research identifying protective mechanisms in cells
• Materials science discovering how humidity affects solar cell efficiency
• Neuroscience finding protein decline patterns in Alzheimer's patients
• Statistical genetics uncovering diabetes protection mechanisms

Collaborating scientists report Kosmos findings scale linearly—the more cycles you run, the more valuable discoveries you get.

Bottom line: AI is moving from helpful assistant to actual research collaborator. When a system can read thousands of papers and run tens of thousands of analyses overnight, the bottleneck shifts from "doing the work" to "asking the right questions." Scientists who learn to direct AI research agents will outpace those who don't.

🌐 3. Google Launches Gemini 3 With Multi-Step Task Agents

Why This Matters

• Gemini 3 integrates directly into Google Search with "AI Mode" that generates full answers, not just links
• New "Gemini Agent" features can organize your email, book travel, and coordinate tasks across Calendar, Gmail, Drive, Photos, Maps, and YouTube
• Uses "2.0 Flash Thinking" to break down complex tasks and show its reasoning as it works

The Reality Check

Rolling out across Gemini app, AI Studio, and Vertex AI now. Warren Buffett’s Berkshire Hathaway just took a stake in Alphabet, signaling institutional confidence in Google’s AI strategy. Stock up 55% in 2025.

Google’s also expanding globally—just opened a new DeepMind research lab in Singapore focused on real-world AI applications and Asian-language capabilities.

The Practical Impact

What you can do right now:

• "Analyze everything my team discussed about Q4 and create a summary"
• "Find a restaurant with good happy hour deals for next Wednesday and make a reservation"
• "Look at my calendar and brief me on upcoming meetings based on recent news"

The agent mode coordinates across apps without you switching between them.

Bottom line: The AI assistant vision is finally real. When your AI can actually read your emails, check your calendar, search the web, and book restaurants all in one request, productivity shifts from "doing tasks" to "directing outcomes." The companies with the best multi-app integration will win.

1. 💡 Use AI to Generate “Expert Panel” Perspectives

The problem: You're facing a big decision and need diverse viewpoints, but assembling advisors is slow and expensive.

The solution: Use AI to simulate a panel of experts with different priorities examining your decision from multiple angles.

2-minute setup:

I need to analyze [your decision] from three perspectives:

1. A risk-averse CFO focused on financial implications
2. A growth-focused CMO looking at market opportunities
3. A technical CTO concerned with implementation feasibility

For each perspective, provide:
- Their main concerns
- Critical questions they'd ask
- Their recommendation

Decision to analyze: [describe your situation]

What you get: Three distinct expert viewpoints that challenge your assumptions and reveal blind spots you hadn't considered.

Real example: Deciding whether to invest in new software? The CFO highlights ROI, the CMO sees market positioning, and the CTO flags integration issues.

The trick: Add more perspectives like legal counsel, HR, or customer success. The AI adapts to each role's priorities.

Tools:ChatGPT (best for distinct personas) | Claude (nuanced reasoning) | Gemini (contextual reasoning)

Click the underlined tools to check each one out.

2. 🔄 Turn Long Articles Into Instant Audio Briefings

The problem: You have 10 articles bookmarked but no time to read them, and want to absorb the insights during your commute.

The solution: Use NotebookLM to convert any article, PDF, or research paper into a natural-sounding audio conversation between two AI hosts.

2-minute setup:

1. Go to notebooklm.google.com
2. Create a new notebook and upload your articles (paste URLs or upload PDFs)
3. Click “Generate” under the Audio Overview section
4. Get a 5–10 minute podcast-style discussion in ~2 minutes

What you get: Two AI hosts debate key ideas, explain concepts conversationally—much more engaging than text-to-speech.

Real example: Upload 5 competitor reports → receive a 12-minute audio briefing highlighting key patterns and contradictions.

The trick: Upload up to 50 sources per notebook. NotebookLM finds cross-source insights automatically.

Advanced move: Upload a draft of your article or presentation and hear what’s missing.

Tools:NotebookLM (free with Google account)

Click the underlined tools to check each one out.

3. 💬 Have AI Write From Reference Examples (Not Generic Instructions)

The problem: You tell AI "write in a casual tone" but it still feels robotic or misses your style.

The solution: Instead of describing tone, give AI actual writing examples. It will analyze and match the style.

2-minute setup:

I'm going to show you 2-3 examples of writing I like. Analyze the style, tone, sentence structure, and approach. Then write [your content request] using that same style.

Example 1: [paste writing sample]
Example 2: [paste writing sample]
Example 3: [paste writing sample]

Now write: [your specific request]

What you get: AI output that truly matches your voice and style—no more endless edits.

Real example: Paste 3 great email subject lines → ask for 10 more in the same tone. Or mimic a blog post style.

The trick: AI learns from multiple examples—sentence length, humor, word choice, etc. Two show a pattern, three confirm it.

Advanced move: Build a style library of your best writing by category and reuse it often.

Tools:ChatGPT | Claude | Gemini

Click the underlined tools to check each one out.

A quick note before you go

Thanks for reading this week’s Brain Bytes — I hope something here helped you move faster or think better.

How’d this one land?

• A ✨ Super Helpful!
• B 🤔 Some Good Stuff!
• C 😕 Could be better.

See you next week, — Oliver