New AI Features to Make Your Life Easier
PLUS: OpenAI's plan to replace LinkedIn and why AI confidently gives wrong answers
Happy Friday.
Claude's new file creation feature reminds me why I'm glad we're past the days of AI hallucinating documents it created.
You know that frustrating experience where AI would confidently tell you it made a Figma file or Notion doc, you'd get excited, then realize there was nothing actually there?
Things are starting to feel different.
The gap between AI theory and actual usability is shrinking. Tools are finally starting to work the way non-technical people expect them to, not just how technical people can make them work.
1/ New AI Features to Make Your Life Easier
What happened: Major AI platforms rolled out features that help you integrate with your existing workflows instead of forcing you to work around their limitations.
Claude Can Now Create Files For You
Claude rolled out document creation this week. You can now ask it to build Excel spreadsheets, Word docs, PowerPoint decks, and PDFs directly in the chat.
The feature works for Max, Team, and Enterprise users now. Pro users get it soon. Just enable "Upgraded file creation and analysis" under Settings > Features > Experimental.
Fair warning: there are some data privacy concerns since Claude needs internet access to build files.
Replit Launched Agent 3 for No-Code Automation
Replit's new no-code Agent 3 can code for over 3 hours without you touching anything. It builds apps, tests them, fixes bugs, and even checks for security issues.
And for the first time, Replit can build agents and automations for you. Exciting for non-technical workers looking to build.
The company went from $3M to $150M revenue in a year. They're valued at $3B now. Insane growth.
What this means: Building custom business tools just became more accessible to everyone. If you do routine tasks, want to create custom workflows, or need to connect different tools, this is a step towards more seamless AI integration.
ChatGPT Projects Goes Free & They Introduce Branching
OpenAI made its Projects feature available to free users. This is one of my most-used AI features. Highly recommend you try it out if you haven't already!
Free users can upload 5 files per project, organize chats by topic, and give each project custom instructions that help AI remember context between sessions.
Plus, they added a chat branching feature to be stoked about. Start a conversation about one thing, branch off a response to explore a side question, then return to your original thread without mixing up too many details. Pretty clever.
2/ Why Does AI Get Things Wrong?
What happened: Researchers at OpenAI and Google DeepMind published studies explaining flaws in how AI systems work and why they confidently give wrong answers.
The Hallucination Problem Has an Easy Fix
OpenAI published research explaining why AI models confidently give wrong answers instead of saying "I don't know."
The issue: current training rewards lucky guesses over admitting uncertainty. Models get full points for correct answers, zero for saying they're unsure.
This creates a system that always guesses, even when it's completely lost.
The fix: AI orgs should redesign training to penalize confident wrong answers more than honest uncertainty. Seems obvious in hindsight.
Google Found a Fundamental Flaw in RAG Systems
Google's DeepMind discovered a major problem with RAG (retrieval-augmented generation) - the tech that lets AI search through database documents.
RAG works by having AI search your data files first, then use that info to respond. It's what powers AI tools that answer questions about a company's docs.
The problem: RAG hits a mathematical wall when databases get too big. The system can only handle so many question-document combinations before performance collapses.
Think of it like this: imagine a filing cabinet that can only hold so many folders. Once you exceed that limit, the whole system breaks down.
This explains why some AI search tools work great with small datasets but completely fail with large ones.
3/ OpenAI Plans to Disrupt the Job Market
What happened: OpenAI is planning to enter the job search market.
Will OpenAI Compete with LinkedIn & Indeed?
OpenAI is launching an AI-powered job platform mid-2026. It's goal is to match employers with AI-skilled candidates and include certification programs.
Their goal: certify 10 million Americans in AI skills on the platform by 2030.
Coinbase: 40% of Our Code is AI-Written
In other news, Coinbase CEO Brian Armstrong confirmed that 40% of their daily code comes from AI tools like GitHub Copilot.
They're targeting 50% by October.
Here's the part that should wake you up: they've terminated programmers who refused to use AI tools.
4/ Other Signals Worth Knowing
- Google Admits Open Web is in Decline. The company acknowledged in legal filings that the open web is "already in rapid decline," contradicting their public claims that search is thriving. Interesting timing.
- Microsoft Diversifying Beyond OpenAI? They're reportedly adding Anthropic's Claude to Office 365 apps. Claude reportedly performs better than ChatGPT models for spreadsheet and presentation tasks.
- Oracle Landed $455B in AI Infrastructure Deals. $300B from OpenAI alone. Larry Ellison briefly became the world's richest person after the announcement.
