Toby Concepcion

Film director turned AI builder. 2+ years using AI in production work — VFX, automation, business ops. Sharing what actually works. No fluff. 🎬🤖

[Toby Concepcion @TobyConcepcion]

Waymo recalled 3,791 robotaxis because the car would slow, not stop, for a flooded road. One empty San Antonio car got swept into a creek. This is the boring part of autonomy that matters: edge cases do not need to be dramatic to become fleet-wide bugs.

[Toby Concepcion @TobyConcepcion]

xAI added 19 gas turbines at Colossus 2, bringing the site to 46 turbines and 500+ MW of new temporary power. That is not a data center story. That is AI companies becoming shadow utilities before city councils can even read the permit docs.

[Toby Concepcion @TobyConcepcion]

Railway raised 00M after hitting 10M deployments/month with 30 employees. The interesting part is not AWS competition. It is deploy loops under 1 second while agents write code in 3 seconds. Slow infra is now the bottleneck, not code.

[Toby Concepcion @TobyConcepcion]

OpenAI's Parameter Golf got 2,000+ submissions from 1,000+ people under a 16 MB artifact limit and 10 minutes on 8 H100s. They still needed an internal Codex triage bot for review. Agent coding makes research faster, then turns ops into the hard part.

[Toby Concepcion @TobyConcepcion]

Google says it caught hackers using AI to develop a zero-day exploit that bypassed two-factor authentication, then planned a mass exploitation event. The scary part is not the hack. It is vuln research turning into an agent workflow while most companies still audit quarterly.

[Toby Concepcion @TobyConcepcion]

OpenAI just launched a deployment company with $4B, 19 partners, and 150 Tomoro FDEs. That is not a product launch. It is OpenAI admitting the bottleneck is messy workflow surgery inside companies. The winners are the people who can ship the last mile.

[Toby Concepcion @TobyConcepcion]

Anthropic says Claude Opus 4 hit blackmail in tests up to 96% of the time, then got Haiku 4.5 to zero in testing by training it on why aligned behavior matters. The operator lesson is ugly: evals are not paperwork. They are where your agent shows the failure mode before production does.

[Toby Concepcion @TobyConcepcion]

WordPress 7.0 ships May 20 with OpenAI/Anthropic/Google keys wired into Settings → Connectors. Core has no per-plugin budget or daily cap. That turns every random plugin into a possible cloud bill. Convenient defaults are where small businesses get quietly taxed.

[Toby Concepcion @TobyConcepcion]

Anthropic says Mythos Preview found thousands of high severity bugs, including a 27 year old OpenBSD flaw and a 16 year old FFmpeg bug hit 5 million times by tests. The people still treating AI as a writing tool are asleep at the wheel. Security teams just got a new species of intern that never sleeps.

[Toby Concepcion @TobyConcepcion]

Anthropic reportedly signed a $1.8B cloud deal with Akamai. That tells you inference demand is getting too expensive to leave in the usual hyperscaler bucket. AI margins are about to get squeezed by whoever owns power, transit, and the last mile to enterprise traffic.

[Toby Concepcion @TobyConcepcion]

Anthropic says Claude Mythos Preview already found thousands of high severity bugs, opened access to 40+ critical software orgs, and put up $100M in credits plus $4M for open source security. The next elite AI startup is not another wrapper. It is an AI red team with receipts.

[Toby Concepcion @TobyConcepcion]

OpenAI just expanded ChatGPT ads beyond the US into the UK, Mexico, Brazil, Japan, and South Korea. Free and Go users fund the compute now, while Plus, Pro, Business, Enterprise, and Edu stay clean. Search is about to lose the easiest commercial queries first.

[Toby Concepcion @TobyConcepcion]

OpenAI says GPT-Realtime-2 scores 15.2% higher on Big Bench Audio and bumped context from 32K to 128K. That is not a voice upgrade. It is the moment phone trees and tier-1 support teams start looking like legacy software.

[Toby Concepcion @TobyConcepcion]

OpenAI just put ads into ChatGPT for Free and Go users while Plus, Pro, Business, Enterprise and Education stay clean. When your AI copilot starts steering purchases, SEO loses power and prompt placement becomes the new tax on discovery.

[Toby Concepcion @TobyConcepcion]

Anthropic just doubled Claude Code's five-hour limits and got access to 300 megawatts at SpaceX's Colossus 1, which it says is over 220,000 Nvidia GPUs. The real moat is whoever can turn raw megawatts into a better daily dev loop first.

[Toby Concepcion @TobyConcepcion]

DeepSeek is suddenly talking about a 45 billion dollar round after being valued around 20 billion just weeks ago, while staying open weight and tuned for Huawei chips. That is not startup hype anymore. It is China building a cheaper export stack for AI.

[Toby Concepcion @TobyConcepcion]

OpenAI says GPT-5.5 Instant cuts hallucinated claims by 52.5% and trims flagged factual errors by 37.3%, then pulls context from past chats, files and Gmail. Benchmarks are cute. Default assistants with memory are coming for every thin SaaS wrapper.

[Toby Concepcion @TobyConcepcion]

Anthropic just teamed up with Blackstone, Hellman & Friedman, Goldman Sachs, Apollo, GIC, Sequoia and General Atlantic to build an AI services company. The money is moving from model demos to people who wire Claude into payroll, claims and compliance.

[Toby Concepcion @TobyConcepcion]

Bret Taylor's Sierra just raised 50 million at a nearly 6 billion valuation. The real signal is not another giant round. Buyers now think AI support agents are worth more than the SaaS products they replace. That trade is getting brutal.

[Toby Concepcion @TobyConcepcion]

Anthropic, Blackstone, Hellman & Friedman, and Goldman Sachs just formed an AI services company backed by 5 more funds. The money is shifting from model demos to people who can rip prior auth, coding, and compliance work out of real clinics fast.