Last May, Sam Altman presented a three-year countdown timeline at Sequoia Capital’s AI Summit: 2025 as the year of assistants, 2026 entering research, 2027 stepping into the physical world.

At the time, it sounded like fundraising rhetoric. But nine months later, I’ve found that how I use Claude and ChatGPT for work daily is completely different from last year. It’s not simply “better to use”—entire workflows have been reorganized.

This made me revisit that timeline. Not to verify whether Altman is a prophet, but to figure out: where exactly do we stand?

The Five-Level Model: From Theoretical Framework to Living Coordinates

OpenAI’s original five-level AI evolution model is essentially a capability spectrum: L1 chatbot, L2 reasoner, L3 agent, L4 innovator, L5 organization. Last year everyone debated “are we at L2 or L3,” and now the answer is gradually becoming clear.

ChatGPT’s Agent Mode is already live—you can have it go through hundreds of emails, help you research, automatically organize data. The GPT-5 series has iterated from launch through 5.1, 5.2, and even 5.2-Codex specifically for coding tasks. OpenAI retired six models at once in February 2026, including GPT-4o and o4-mini, essentially declaring: we’re not looking back.

But does this mean “L3 has arrived”? My observation is: capabilities have arrived, reliability hasn’t. When you have an AI agent help with a task, it can complete it 80% of the time, but that remaining 20% is precisely where you can least afford errors. This isn’t a minor issue—this is the real chasm between L3 and L4.

Timeline Revision: From “Three Years to AGI” to “Gradual Awakening”

What’s most noteworthy isn’t whether OpenAI stayed on schedule, but how they quietly modified their narrative.

Altman’s latest position is: “AGI isn’t some magical moment, it’s more like a process, and you’re already on the path.” OpenAI’s Chief Scientist Jakub Pachocki provided more concrete milestones—creating “research intern” level AI by September 2026, and AI researchers capable of independently completing research projects by 2028.

Notice this shift: from “reaching AGI in three years” to “creating AI interns in two years.” The goal dropped a level, but the timeline became more credible. This isn’t weakness, it’s engineer-style pragmatism—first prove AI can be an intern, then discuss whether it can be a professor.

Infrastructure Arms Race

The original article mentioned Altman’s five recommendations at the Senate hearing, and nine months later, each has become real money in action:

The “Stargate” project has evolved from presentation slides to contracts, with OpenAI’s promised infrastructure investment exceeding 30 gigawatts of computing power, and cross-industry supply chain contracts with AMD, Broadcom, Google, Microsoft, Nvidia, and Oracle worth hundreds of billions. Altman says he wants to compress per-gigawatt computing costs below $20 billion, sounding more like energy policy than a tech project—which is exactly the point.

When computing power becomes infrastructure like electricity, AI is no longer Silicon Valley’s toy, but global public goods. The question is: who controls this “AI grid”?

Middle East Chess Game: A New Geopolitical Order of Resources for Technology

The original article mentioned Altman’s Middle East trips—accompanying Trump to Saudi Arabia, Qatar, UAE—which looked like diplomatic theater at the time, but in hindsight were opening moves in strategic positioning. Saudi Arabia’s commitment of $600 billion investment and Qatar’s $243.5 billion contract have exceeded the scope of mere “tech cooperation.”

Essentially, this is a trade: America provides technology and chips, the Middle East provides money and energy. AI training is energy-intensive, and the Middle East has the world’s cheapest energy. If this combination succeeds, it changes not just AI industry supply chains, but the center of global power.

Where does Taiwan fit on this chessboard? We hold the lifeline of advanced processes, but we’re not protagonists in either AI applications or energy. As I discussed in “Taiwan Semiconductor’s Tenfold Leap,” process advantages are both moats and ceilings—if we only guard foundry services, when AI value chains tilt toward applications and energy, our moat becomes an island. This gap is what we should truly be concerned about.

Sober Observations in Year Two of the Countdown

Back to that three-year timeline. If 2025 was the “year of assistants,” this prediction has largely come true—AI has indeed upgraded from “Q&A tool” to “task executor.” But what about 2026’s “year of research”?

Currently, what AI can do is accelerate research processes, organize literature, run data analysis. What it can’t do is propose original hypotheses, design experiments, or adjust direction through failure. In other words, AI is becoming excellent research assistants, but there’s still a fundamental gap before becoming “researchers.”

My personal judgment: the 2027 “physical world” goal will likely be delayed. Not because technology isn’t sufficient, but because trust isn’t sufficient. When AI begins operating in the physical world—factories, healthcare, transportation—fault tolerance approaches zero, and current reliability can’t support that standard.

With the three-year countdown reaching year two, sobriety matters more than optimism. Technology is accelerating, but building trust has no shortcuts.