Anthropic ships Claude Opus 4.8 — Anthropic's new flagship stopped guessing. Opus 4.8 is roughly 4x less likely to let an error slip by unflagged, and when it isn't sure, it tells you instead of inventing a clean-looking answer. Feed it a rent roll that doesn't tie or an OM with a junk assumption and it raises its hand. Fast mode also got 3x cheaper.
Google launches Gemini 3.5 Flash at I/O 2026 — Google's new Flash model delivers near-Pro reasoning at half the cost and 4x the speed, and it's already the default behind Google Search's AI Mode so you're probably using it without realizing. To put it in CRE terms: the 1M-token context window means you can drop a full OM, the rent roll, and a stack of leases in at once and ask it to find the discrepancies. It reads scanned PDFs, photos, and video natively, too. No more chopping a deal into pieces to fit.
Microsoft launches Agent 365 — Microsoft shipped the control tower for AI agents. Every agent your org runs gets a managed identity, lands in one registry, and answers to your compliance policies all from a single dashboard. The second your team points agents at leasing inquiries, AP, or underwriting prep, this is what keeps them out of the tenant and financial data they've got no business touching.
Anthropic overtakes OpenAI as the world's most valuable AI company — A $65B raise nearly tripled Anthropic's February number and crowned Claude the most valuable AI lab on earth at a $965B valuation. Claude is getting wired straight into PE real estate portfolios (see below). The model your auditors, lenders, and capital partners are increasingly running is Claude. Standardizing on it stopped being a fringe bet.
Blackstone and Brookfield bet on custom AI — and proptech could pay the price — The biggest owners are buying engineers. Anthropic's $1.5B JV (Blackstone, Goldman, H&F) and OpenAI's $10B Deployment Co (Brookfield, TPG, SoftBank) both drop teams inside CRE portfolio companies to rebuild underwriting and asset management around the model. That's Palantir's playbook, not a SaaS license. If you're not operating at that scale, your edge isn't out-spending them. It's knowing exactly which off-the-shelf tools close the same gap for pennies on the dollar.
Here's the problem almost every CRE professional has and very few have solved:
Your best intelligence is scattered. Deal notes in one app, broker conversations in your inbox, market reads in your head, submarket comps in a folder you haven't opened since March.
Every time you start a new analysis, you're rebuilding context from scratch and your AI is doing the same thing, because it has little to no memory of anything you've told it before (if you’re using Claude Cowork, there is no built in memory - link here for lesson on that in the Studio: Week 28: Making Claude Cowork More Powerful: Building Claude.md and Memory.md files).
I've been building a fix for this, and it's good enough that we're dropping a full lesson on it in the Studio this week.
The concept comes from Andrej Karpathy's (co-founder of OpenAI, now at Anthropic) "LLM knowledge base" pattern, which is basically a living wiki that your AI maintains for you and, critically, improves on its own over time. The more you use it, the smarter it gets.
No code, no Obsidian, no new subscription. Just Claude and a folder on your computer.
Here's the architecture and the five-step build so you can start before the lesson drops.
The architecture: one brain file, three folders. At the center is a single CLAUDE.md file, which you can think as your knowledge base's operating manual. It tells Claude who you are, what you care about, how your business is structured, and where everything lives.
Around it sit three folders:
Inbox — where you dump raw, unstructured stuff: meeting notes, voice memos, broker emails, half-formed thoughts.
Knowledge — the organized wiki: clean, structured notes on markets, sponsors, lenders, asset classes, playbooks.
Projects — your live work: active deals, the disposition you're underwriting, the LP update you're drafting.
That's it. The whole system is just text files Claude can read and write.
Step 1 — Set up the structure. Create the three folders and an empty CLAUDE.md. In that file, write a few plain-English paragraphs: your role, your portfolio, your investment criteria, your hurdle rates, the way you like work delivered. This is the context Claude reads every single time, so it never starts cold again.
Step 2 — Do the information dump. Don't organize anything yet. Just unload. Drop your last ten deal memos, your market notes, your broker contacts, your standard underwriting assumptions into the Inbox. Messy is fine.
Step 3 — Let Claude build the wiki. Now you point Claude at the Inbox and ask it to organize everything into structured notes in the Knowledge folder like a page per market, per sponsor, per recurring workflow. Claude decides what belongs where and cross-links it. What was a pile of fragments becomes a second brain in minutes.
Step 4 — Create the compounding loop. This is where it stops being a static file dump and starts self-improving. You add an instruction to your CLAUDE.md telling Claude that after every meaningful task, it should append what it learned ( what worked, what didn't, the new comp it found, the lender preference it picked up) back into the right note. Every conversation now makes the system permanently smarter. That's the whole magic.
Step 5 — Schedule a monthly health check. Set up a recurring task (Claude can easily run on a schedule) that audits the knowledge base once a month: flag stale notes, merge duplicates, surface gaps, and clean up the index. Your second brain maintains itself instead of rotting like every other folder you've ever made.
The payoff is a one-day-vs-hundred-day thing.
On day one it's a tidy set of notes. By day one hundred it's an asset that knows your markets, your deals, and your judgment cold and gets sharper every time you touch it.
This week's lesson goes deep. In the Studio I'm walking through the entire build live, start to finish, including the exact CLAUDE.md template, the prompt that turns your Inbox into a wiki, and the health-check task that runs it all on autopilot, all adapted specifically for CRE workflows.