Carlos's personal intelligence site · Updated twice daily
One venture, tracked live: an AI Shell that reads the page you're on and rebuilds it around the task in front of you — plus the daily AI & tech radar that feeds it.
02 · Live demo — try the product
Pick a goal. The Shell re-sorts, re-ranks and re-emphasises the exact same flights. Nothing about the data changes — only what matters to you, right now.
Ranked by true total — fare + checked bag. The cheapest headline fare isn't always the cheapest trip.
Demo data for illustration. In the real Shell, every number links back to its source on the original page.
03 · AI & Tech Radar
Five signals that matter, read through four lenses: career, business, build & learn, and money. Vendor claims are labelled as such.
Tap any card for the full deep dive — plain words, a visual, primary sources and next steps.
GPU compute-sort renderer for Gaussian splats in three.js — ~9ms vs ~16.5ms at 10M splats (vendor docs). Pairs with Idol (image → splat) for a fully open splat pipeline.
So what for Carlos: the missing fast end of your splat product idea now exists as an MIT npm package.
Targeting up to C$7B at a C$35.7B valuation; Canada adds up to C$1.6B in staged equity tied to jobs and compute.
So what for Carlos: a study item, not a build item — capital is concentrating in sovereign, regulated AI.
$15.3B enterprise value, $10/share, go-shop window to Nov 28. OpenAI is a Thrive partner and investor.
So what for Carlos: the model to study — buy cash-flowing businesses, inject AI. Your Acquire lens, at $15B scale.
$11M total for agent identity & access control, with Karpathy and Dean as angels. Agents need identities, permissions and audit trails.
So what for Carlos: your #2-ranked IDV-adjacent lane is getting crowded — the window is closing, not opening.
Expanded misalignment monitoring plus a Model Spec rewrite. METR and Redwood published bypasses on Oct 22 — vendor claims, labelled as such.
So what for Carlos: direct input to the Shell's safety design — assume injection works, architect around it.
S-1 filed Oct 26, ticker ANT. H1 2026 revenue $16.29B, verified from the filing. Price range expected ~Nov 9–16.
Standing watch: flag the moment the amended S-1 with a price range files, or ANT hits any broker IPO calendar.
World modelsWorld Labs · docs
A 3D scene can be built from millions of tiny fuzzy blobs instead of triangles. Seen from any angle it looks photographic — that's a Gaussian splat. To draw it, blobs must be painted farthest-first or transparency blends wrong.
Spark Radix is World Labs' new renderer that does the sorting on the graphics chip itself, in one pass, nothing sent back to the main processor — which is why it's ~1.37× faster than their previous renderer (9.2ms vs 16.5ms per frame at 10M splats, vendor docs).
Same ~140 blobs, both times. Only the paint order changes.
So what for Carlos: the missing fast end of your splat product idea now exists as an MIT npm package.
FundingCohere · reported
Cohere builds AI models for companies and governments that can't just send their data to an American tech giant — banks, hospitals, ministries. "Sovereign AI" means: the model runs on infrastructure you control, in your country, under your rules.
This week Cohere raised the biggest private round in Canadian tech history, and the Government of Canada itself bought a stake — governments are no longer just regulating AI, they're buying equity in it.
| Lead investor | Schwarz Group (Germany — also runs the STACKIT sovereign cloud) |
|---|---|
| Canada's stake | Up to C$1.6B equity, paid in stages |
| In return, Cohere commits to | HQ stays in Canada · ≥C$1.7B Canadian compute by 2030 · ≥1,000 new technical jobs |
| Round | Targeting up to C$7B at a C$35.7B valuation (target, not closed) |
| Company-stated | ARR +165% YoY (unverified) |
So what for Carlos: a study item, not a build item — capital is concentrating in sovereign, regulated AI.
Roll-upThrive Holdings
Thrive Holdings is a holding company built to buy mature businesses and run them forever — not flip them. Its twist: OpenAI is an investor and partner, supplying models and forward-deployed engineers who work inside the acquired companies.
ADT, the home-security giant, is its biggest bet yet: $10/share in cash, $7.6B equity / $15.3B enterprise value, a 45.6% premium, closing expected Q1 2027 after a go-shop window that ends Nov 28.
The same play at your scale: Acquire.com listing → your AI leverage → cash flow. ADT is that idea with 9 zeros.
So what for Carlos: the model to study — buy cash-flowing businesses, inject AI. Your Acquire lens, at $15B scale.
IdentityMintive · seed
Companies are deploying AI agents that can read email, query databases and call tools — but each agent needs an identity, like an employee badge: who is this agent, what is it allowed to touch, and a log of everything it did.
Mintive sells exactly that control layer: an MCP gateway (MCP is the standard plug agents use to reach tools) where IT grants permissions, a model gateway that caps spend, and a monitor that tracks every agent in the company. $7M seed, $11M total, 50+ enterprise customers claimed since February.
Side effect by design: every action is written to an audit log — who did what, when, under which permission.
So what for Carlos: your #2-ranked IDV-adjacent lane is getting crowded — the window is closing, not opening.
SafetyOpenAI · vendor
OpenAI runs a monitoring system meant to catch its own internal agents misbehaving. It invited two outside evaluators, METR and Redwood Research, inside for three weeks to attack it. They found the watcher had holes: monitoring only fully engaged when an agent handed control back to a human, about 10% of internal coding-agent traffic went unmonitored during "compaction" events, and they found ways agents could in principle bypass the sandbox and the monitor itself.
OpenAI says it patched the issues; the evaluators have not verified the patched system.
So what for Carlos: direct input to the Shell's safety design — assume injection works, architect around it.
⏰ WatchSEC filing
Anthropic, maker of Claude, filed its public prospectus (the S-1) on Oct 26 to list on Nasdaq as ANT. The filing's real story: revenue exploded — $16.29B in the first half of 2026 vs $1.16B a year earlier, and Q2 2026 produced its first positive operating income, +$974M.
The scary-looking $80.7B "loss" is mostly paper: when Anthropic's valuation rises, accounting rules force it to book the increased value of its convertible notes as a loss. The genuinely big number is ~$265.6B of future compute commitments. Price range and share count are still blank — that's the next filing to watch, expected ~Nov 9–16.
| H1 2026 revenue | $16.29B |
|---|---|
| Q2 2026 operating income | +$974M (first positive quarter) |
| Headline net loss | $80.7B — mostly paper (convertible-note revaluation) |
| Compute commitments | ~$265.6B |
⚠️ Anything offering you Anthropic shares before the amended S-1 prices is a scam. E*TRADE (Morgan Stanley is in the syndicate) is your likely allocation channel; Fidelity is not in the syndicate.
Standing watch: flag the moment the amended S-1 with a price range files, or ANT hits any broker IPO calendar.
04 · The research journey
Scroll the programme: validation (V), architecture (A) and Shell (B) tracks. Each run verified numbers against primary sources — and killed a few comfortable assumptions.
Carlos was right: nobody pays for extensions. Chrome Web Store payments ended in 2021. Sell the Shell as a subscription — the extension ships free as its companion.
No AI browser ships generated task UI as its main surface — the gap is unclaimed. Atlas died after 292 days and folded into a Chrome extension. Hands-on installs still pending.
Arc died of the novelty tax — its cleverest features sat under 5% usage. Humane and Rabbit confirmed it: novelty is not retention. Now a POC metric: generated UI must win ≥50% of eligible tasks by week 2.
No one searches for “adaptive UI” — but clutter, tab overload and comparison pain are verified and huge. Rule: sell outcomes in the user's words, never the category name.
At $10/mo and 0.5–2% conversion, $40k/mo needs 200k–800k active users. Distribution — not market size, not unit cost — is the binding constraint.
Property prosumer wins, 21/25. Job hunting scored highest on pain but was rejected — its money is in applying and autofill, exactly the actions v1 excludes.
Property tier corrected to $49–99/mo — the anchors sell data, so the tier must carry cross-portal aggregation, not just a nicer UI. Affiliate revenue is a subsidy, never the business.
No verified one-person billion-dollar software company exists. Solo is a phase-1 model only — every winner added humans at proof. Audience-building is part of the venture, not optional marketing.
YC already funded Cyclon — “a phone where AI is the interface” — so the generic pitch slot is occupied. Pitch the working Shell plus one measured workflow. Solo Founders Program (Feb 2027) is the first-choice vehicle.
A task-coverage wedge, not a launch geography — language is not the moat. Colombia has no MLS: the same unit is multi-listed at different prices across portals. That cross-portal dataset is the real moat.
Pruned DOM + accessibility tree wins. Representation alone swings hallucination ~30× — so JSON-LD is layer 1 and every price and date gets per-value validation. Vision is demoted to region-crop fallback.
Prompt injection is unsolvable at the model layer — OWASP, NCSC and OpenAI all say so. Answer: quarantined extraction, no model-chosen URLs, and a read-only v1 that cannot act, buy or send.
14 registered components, ≤12 visible to the planner per task. The spec — morph.spec/0.1 — is a flat list with data bound by path, so the planner never transcribes a price.
Four clocks, not one: ack ≤100ms, first real content ≤400ms on a template hit, complete spec p95 ≤10s on a miss. Golden rule: cache structure, never data — prices are always re-extracted live.
Blended cost fell to ~$0.0011 per adapted view (~$0.68/user/month) after price corrections — so route on latency and spec validity, not price. Sonnet is the offline yardstick; fast non-reasoning models plan at runtime.
MV3 bans remote code but allows remote configuration — exactly what our specs are. Two-speed architecture: specs update instantly server-side; new components ship through store review. Render home: side panel first.
Built and tested: goal → spec → 3-stage validator → rendered components, with a Node harness green on 6 scenarios and 5 negative tests. The live LLM call is gated on deploying the gateway Worker with a provider key.
The accessibility-agent route is closed on Play — prohibited for non-accessibility tools, full stop. Android becomes: Shell app + parameterised widgets first, launcher optional later.
Four code-enforced data classes. Raw URLs never leave the device. And the trust asset isn't copy — it's a per-task receipt: “here is exactly what left your device.”
Kaizōsha's Morph is an iOS extension with the same name and the same mechanic. Shipping second under that name makes us the copy. Candidate ACOMODA is unverified; a full knockout pass is now a week-1 task.
Server fetch is being blocked at the infrastructure layer and Amadeus killed its self-serve API. Routing rule: licensed API → extension capture → server fetch → agent browsing last. Property has no API channel at all — extension capture or nothing.
The generic Intent Shell is now claimed ×3, for free: ChatGPT Intelligent UI shipped to all tiers Oct 7–8, alongside Google AI Mode and Gemini. Wabi 2.0 pivoted to agent-decided, shared apps ($55M at $1.25B, Oct 29 — vendor-reported, no usage disclosed) and Disco/GenTabs is still waitlist-only. Voice-first is rejected as a claim. Still unclaimed: generated UI from your own logged-in web session across arbitrary sites, and prosumer depth in one vertical — property.
05 · Architecture
Five steps from messy page to task-shaped UI. Click each step — the detail changes. (Yes, like the product.)
The free companion extension reads the page you're already logged into — locally, on invocation only. No standing all-sites permission, no background scraping. Cookies, form values and other tabs never leave the device (Privacy Class 0).
Why it matters: property portals block bots and offer no API — the user's own session is the only reliable channel.
06 · Decisions locked
Not opinions — scored, dated, and logged old → new in the standing docs. Carlos can veto any of them; until then, they steer the build.
The Shell is the product; the extension is its eyes and hands inside your logged-in browser. Weighted score: Hybrid 4.30 vs extension 3.85, shell-alone 2.85, browser fork 2.55. B2 sharpened why: that logged-in session is the only unclaimed data surface — everyone else generates from chat or their own data. Pitch rule tightened: never pitch “AI generates your interface” — since Oct 8 that is a free ChatGPT feature.
Standalone subscription with its own billing. The extension is listed free — the store is a channel, never the paywall.
#1 Property prosumer ($49–99/mo) · #2 Travel · #3 Trámites (legal-gated). Personal dogfooding starts with travel + property; the public beachhead follows whichever mode wins.
Same name, same mechanic, already claimed on iOS — plus a crowded trademark register and every good domain taken. It survives only as an internal codename.
The Shell cannot act, buy or send — by architecture, not by promise. Every number links to its source. That's the trust pitch in a category built on agents that click things.
AI agents cover build, maintenance and ops. They don't cover distribution or trust — those are the founder hours. The day job is the financing, not the obstacle.
07 · Opportunities
Every research cycle ends here: what to watch, what to learn, what to build. Investment items are watchlist research only — never trade calls.
Azure grew +48% with a $33.2B quarter. Watch: does capex timing eat free cash flow? Confirm: Q2 Azure +49–50% with FCF >$15B. Kill: Azure slips under ~45% while capex catches up.
Pay Per Use turns agent traffic into a monetised, signed channel — and Cloudflare is the verifier. Confirm: signed-agent revenue disclosed in earnings. Kill: still beta with no uptake by mid-2027.
A8's evidence: Google reserves the two Android AI surfaces that matter — generated widgets and agent function-calling — for itself. Kill: it opens those APIs to third parties generally.
S-1 filed, revenue verified, price range pending (~Nov 9–16). E*TRADE is the likely allocation channel — Morgan Stanley is in the syndicate; Fidelity is not. Research only until terms exist.
Spark Radix's core lesson: when millions of items need ordering per frame, sort on the GPU as compute — never round-trip to the CPU. The pattern behind smooth splat rendering.
Design LLM schemas inside the subset every provider supports — plain types, required, enums — and enforce the rest in your own validator. Applies to every AI product you build.
HTTP Message Signatures are becoming the identity layer for agents; Cloudflare and AWS already verify. The Shell will need a signed identity sooner than expected.
The two Android skills worth having before Phase 3 — parameterised widgets and the agent-callee API. Both reuse the Shell's component/spec model directly.
Single image → .ply splat (Idol) → fast three.js viewer (Spark Radix). A weekend-sized prototype that proves the full open pipeline end to end.
~150 lines on Cloudflare. Normalises three provider APIs, holds the keys, enforces privacy gates. It unlocks A7's live LLM call and both model bake-offs — the highest-leverage build on the board.
“What left your device” — generated from the actual payload, shown in the side panel. A differentiated, buildable trust feature no AI browser ships today.
Extraction templates for Fincaraíz, Metrocuadrado, Ciencuadras and Facebook listings. The cross-portal dataset underneath it is the candidate moat.
Run the same property task in free ChatGPT Intelligent UI (and GenTabs if accessible) vs the Shell hand-run version. Bar: ≥60% prefer ours for the vertical task. If not, the differentiation thesis fails earlier and cheaper.