The Daily 5 — Thursday, April 17
1. Seven AI product launches in one day — and the release calendar officially broke. Opus 4.7 took the top spot on SWE-bench Pro at 64.3%. One hour later, OpenAI fired back with a Codex update featuring background computer use on Mac. Same day: Perplexity shipped a desktop agent, Google put Gemini on Mac, Physical Intelligence dropped π0.7 for robotics. GPT-6? Delayed — put back in the oven. Even OpenAI can't keep up with its own release calendar. Stop evaluating models. Start evaluating architectures. The model you pick today will be surpassed in weeks. (Full analysis below.)
2. Anthropic's Mythos broke into a FreeBSD server in 24 hours — a vulnerability that survived decades of human audits. War on the Rocks published a deep analysis calling it Anthropic's nuclear moment. The model exploits known vulnerabilities at a 72.4% success rate, has found thousands of zero-days, and fewer than 1% have been patched. Anthropic deliberately capped Opus 4.7's cybersecurity capability at 73% through training weights — not guardrails. When the maker of the model decides to weaken its own product before shipping, pay attention. (Full analysis below.)
3. Cloudflare built Git for agents — and it tells you everything about where infrastructure is headed. Artifacts is a distributed, versioned filesystem built for AI agents, not humans. Create millions of repos programmatically. The whole Git engine runs in Zig compiled to a ~100KB WASM binary. ArtifactFS mounts large repos in 10-15 seconds so agents get to work while content hydrates in the background. $0.15 per 1,000 operations. This is what "agent-native primitives" look like in practice — and it's exactly the infrastructure layer Nate's Substack says is the real bottleneck.
4. Canva is now the world's third most-used AI web product — and it just went agentic. 265 million monthly users. $3.5 billion in revenue. And the new "Canva AI 2.0" suite lets agents crawl the web, determine what's trending, create social posts, and schedule them autonomously. Their AI runs 7x faster and 30x cheaper than comparable frontier models because they own their own infrastructure. Meanwhile, Adobe is down 30% and Figma has lost 85% since IPO. The sell-work-not-software thesis in real time.
5. Jeff Dean says your 50x-faster AI only gets you 2-3x results. The METR study proves why. Dean told a Sequoia audience that making a model infinitely fast yields only a 2-3x end-to-end improvement because the tools eat the rest. METR's randomized trial confirmed it: AI made 16 experienced developers 19% slower on their own codebases. The devs didn't believe it — they estimated a 20% speedup. The model was the fast part. The human environment was the slow part. And the slow part won. (Full analysis below.)
"The AI is fast enough. It's been fast enough for a while. Now the question is whether we're ready — our tools, our organizations, and ourselves — for a world that's no longer built for our speed."
SIGNAL/NOISE — Warp Speed, Fast, and Slow
THE NUMBER: 7 — major AI product launches in a single 24-hour window on April 16
There are three speeds in AI right now. The industry is moving at warp speed — seven major launches in twelve hours, with labs firing back at each other within minutes. The threats are moving fast — Mythos found a 17-year-old FreeBSD vulnerability in 24 hours that survived generations of human security audits, exploiting known vulnerabilities at a 72.4% success rate. And we — the humans, the tools, the organizations — are moving slow. Jeff Dean's math is brutal: make the model infinitely fast, get 2-3x end to end. A trillion dollars of inference investment, bottlenecked by compilers, file systems, and authentication flows designed for creatures that think at 3 bits per second.
The METR study is the smoking gun. Sixteen experienced developers. Their own repos. AI made them 19% slower — and they didn't believe it. The models generated useful code. But the environment consumed all the gains: context-switching, re-reading output, managing tool interactions calibrated for human speed. Meanwhile, Google (30%+ AI-generated code), Anthropic (Claude Code writes 80% of its own code), and teams in Jellyfish's data (PRs per engineer jumping from 1.36 to 2.9) are seeing massive gains. Same models. Radically different outcomes. The difference isn't the model. It's the environment.
But the bottleneck isn't just you — it's your entire organization. Approval chains. Committee meetings. Procurement cycles. Human-in-the-loop workflows that exist because the system was built assuming a human was the system. Guillermo Flor nailed it this week: for every $1 spent on software, roughly $6 is spent on the services to operate it. The SaaS playbook captured the software dollar. The AI playbook captures the services dollar — at software margins. Not "AI for accountants." The AI accounting firm. Most companies are still building copilots — tools that make humans faster at the same jobs. That's optimizing the slow part.
The companies that figure out how to sell the outcome skip the bottleneck entirely. And the person at the top — the one defining the right problem, encoding institutional taste into constraints agents can follow — becomes the most expensive component per unit of time in the system. Which is exactly where you want to be when everything else is getting cheaper. Someone put it perfectly: "90% of my skills became useless, but the remaining 10% became worth 1000x." That math works in your favor — if you're honest about which 10% matters.
At COAI today: Full analysis of the seven-launch day, Mythos offense-defense asymmetry, and the Amdahl ceiling your toolchain is hiding from you — at getcoai.com.
Three Questions We Think You Should Be Asking Yourself
What's your Amdahl ceiling — and do you even know? Most teams discover 70-80% of their AI workflow time is tool interaction, not model reasoning. That means no model improvement gives you more than 1.3-1.5x. Your six-figure model API bill might be solving the wrong problem.
If Mythos-class offensive capability proliferates in 12 months, which of your systems survives first contact? Not your perimeter. Your actual systems — the ones with 17-year-old dependencies nobody's audited and legacy code that's "working fine" so nobody touches it.
Is your company selling software or selling work? Canva just went from design tool to autonomous marketing agent. Cloudflare just built Git for agents, not humans. The infrastructure is being rebuilt for a consumer that doesn't have eyes, doesn't take breaks, and processes at 50x your speed. If your product still requires a human to click buttons to generate value, the clock is ticking.
— Harry and Anthony
Sources: