The Perils of the AI Exponential

February 24, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
February 23, 2026

HEADLINES

Happy Birthday, Claude Code

Claude Code turned one this weekend, and Anthropic threw a birthday party to celebrate. It's incredible how far things have come — a year ago, agentic coding was a fascination, something quirky that might help non-technical people build fun apps but was clearly too unreliable for production. Now it's disrupting the software industry and infiltrating every area of work. For Anthropic, what started as a side project from developer Boris Cherny has become the central pillar of their strategy, generating $2.5B in ARR. Cherny recalled Dario asking in the early days, "Hey, are you forcing engineers to use this? Why is everyone using it?" All he had to do was make it available, and everyone voted with their feet. Looking ahead, Cherny believes "coding will be generally solved for everyone" — it's practically solved for him already.

Claude Code Security Crashes Cybersecurity Stocks

Anthropic unveiled Claude Code Security on Thursday — a plugin that scans codebases for vulnerabilities. Friday saw CrowdStrike lose 8%, Okta lose 9%, and Cloudflare lose 7%. The thing is, Claude Code Security doesn't even overlap with what these companies do. It audits internal code; Cloudflare and CrowdStrike protect against external cyberattacks; Okta does two-factor authentication. But this isn't really about the specific catalyst — it's part of a broader repricing of how to value software when the landscape is shifting this quickly. As Bucco Capital put it, "maybe you shouldn't pay 25x revenue when the landscape is shifting this quickly."

GPT-5.3 Rumored for This Week

Google, Anthropic, and xAI have all thrown new models on the pile recently, which means it's about time for OpenAI. The latest rumor is that GPT-5.3 — internally codenamed Garlic and the focus of Altman's Code Red push from December — drops Thursday. Dan Mac's rumor thread claims it surpasses human baseline on SimpleBench and is "a GPT-3 to GPT-4 moment again." Even if it's a big change, expect it to still be called 5.3 given how burned OpenAI has been with bigger naming conventions.

OpenAI Forecasts $282B Revenue by 2030 — and $665B in Cash Burn

The Information got OpenAI's latest investor projections: $282.5B in revenue by 2030 (27% above prior forecasts), with this year expected at $30.1B. But costs are surging — inference costs quadrupled last year, compressing margins from 40% to 33%, and training spend is expected to hit $32B this year and $65B in 2027. Total training costs through 2030: $440B. Weekly active ChatGPT users hit 910M, short of the billion-user target. Last week's Epoch AI chart had Anthropic on pace to overtake OpenAI in revenue by mid-year — which maybe explains why Dario and Sam wouldn't hold hands in India.

OpenAI's Device Plans Take Shape

OpenAI's financials forecast $1.3B in hardware revenue next year — and new reporting shows how they plan to get there. A 200-person team is working on a family of devices including a smart speaker ($200-$300, camera-equipped, no screen), possibly smart glasses and a smart lamp. The smart speaker would use facial recognition for purchases and observe users without a wake word. Notably absent: the behind-the-ear "Sweetpea" device. Design is being led from a separate office near Jony Ive's studio, and some staffers have complained that LoveFrom is slow to revise designs. Basically, assume every object you interact with is being tested for its AI device potential.

MAIN STORY

The Perils of the AI Exponential

The METR chart — the so-called Moore's Law for AI agents — has been updated with Opus 4.6 and GPT-5.3-Codex, and the results have sent shockwaves through the AI and investment communities. Combined with a viral research note from Citrini on "The 2028 Global Intelligence Crisis," we're watching a real-time case study of the moment we're in: a broad-based sense that something very big is happening, accelerating faster than expected, with nobody quite sure what comes next.

THE MOST IMPORTANT CHART IN THE WORLD

Background: Why This Chart Matters
By the end of last year, as the bubble narrative took hold, many were calling this the most important chart in the economy — the bulwark holding back the full tide of AI bubble-pop narratives.
METR's continuous study measures the longest time-horizon tasks an AI agent can handle. First released in March 2025 (when Sonnet 3.7 was state of the art), it found task time horizons were doubling every ~7 months, accelerating to as fast as 3 months. Crucially, this isn't how long an AI can continuously work — it measures how difficult a problem an agent can solve, benchmarked against human engineer completion times. A 50% success rate is the standard threshold; not production-grade, but a consistent gauge of relative improvement.

The New Results: "Going Vertical"
Opus 4.6 has more than tripled the time horizon of 4.5, implying time horizon is now doubling every one and a half months.
GPT-5.3-Codex hit 6.5 hours (exceeding Opus 4.5's 4h49m). Opus 4.6 hit 14.5 hours — the largest generational jump in the study's history. For context, GPT-5.1-Codex in November managed just 2h40m. Swyx had already flagged after the Opus 4.5 result that the curve fit "was probably wrong and needs to be restarted as a new epoch."

The Caveats Are Real
You can caveat this all you like — it's still really meaningful that Opus 4.6 saturated a task set METR didn't think would be saturated for under a year.
METR themselves heavily caveated the results. Opus 4.6 has basically saturated their task set — the upper confidence interval is 98 hours, practically infinite. Researcher David Rein warned that with a slightly different task distribution, they could have measured 8 hours or 20 hours. Dean Ball noted this isn't strong evidence of a "radically faster progress regime" by itself, though it clearly signals nothing is decelerating. The key synthesis: it's possible something massive is happening AND some people are mistakenly thinking it's even bigger than it is — but that doesn't mean it's not very, very big.

The Citrini Piece: "The 2028 Global Intelligence Crisis"
What's more interesting than the particulars of the piece is the response it's getting — this is confirmation bias meeting genuine fear.
Citrini Research — a well-regarded thematic research firm on fintwit — published a piece that essentially takes Dario Amodei's "country of geniuses in a data center" concept and applies it to the real economy. The thesis: capital owners reap massive benefits while workers across every stratum are left jobless; economic activity transforms from household-based to capital-based; stock market collapse and mass unemployment follow. We've heard versions of this before (Situational Awareness, AI 2027), but previous reports were met with skepticism. This time, investors already believe some version of this thesis, so Citrini is acting as confirmation.

The Pushback
We are desperately in need of the non-doomer version of the Citrini piece.
Dan Hockenmaier argued the piece shows a "profound lack of understanding of how marketplaces work" — building a liquid marketplace with optimized supply is the hard part, not building the app. Economist Guy Berger questioned internal consistency: if agents are making money for capital owners, why isn't that fueling employment, GDP, and stock prices? These are valid critiques, but the bigger picture is what matters: the story of early 2026 is a broad-based sense that something very big is happening, coding capabilities have increased dramatically, agents are a real force beyond software engineering, and markets are repricing as a consequence. It doesn't seem like it's going to slow down.

ALSO REFERENCED