Sequoia’s Mission Accomplished
Sequoia’s Mission Accomplished
On May 1, 2003, George W. Bush cosplayed as a fighter pilot, landed on the USS Abraham Lincoln, and stood in front of a banner reading “MISSION ACCOMPLISHED.” Major combat operations in Iraq, he announced, were over.
They were not over. The war continued for eight more years. Thousands more Americans died. The banner became the international symbol for declaring victory while your ass is actively on fire.
Sequoia Capital just held their own carrier landing.
“AGI is here, now.”
Oh, is it? Is it here, Sequoia? Did AGI show up while I was in the bathroom? Should I put out snacks?
Their definition of AGI (and I am not making this up) is “the ability to figure things out.”
MIC DROP, BITCHES
That’s it. That’s the definition. From one of the most prestigious venture capital firms on the planet.
A thermostat figures things out. A binary search figures things out. My Roomba figures things out, usually by ramming into the same chair leg four times and then giving up. By this definition, my Roomba is AGI. Someone call the LPs. Pop the champagne.
The Recruiter That Proves the Opposite
Their pièce de résistance, their “behold, the future,” their evidence that the mission is accomplished: an agent that finds a DevRel candidate in 31 minutes.
It searches LinkedIn. It cross-references Twitter. It filters for engagement. It narrows to three names. It drafts an outreach email.
Color me impressed, Sequoia. You’ve reinvented grep. With a context window. It’s a Boolean search that learned to mass-produce cold emails. This is what Alexander wept for? This is the end of history?
The agent cannot ask whether “DevRel” is even the right role. Or if the leveling is right. Or the salary band needs to be adjusted. It cannot question whether the founder’s mental model of their customer is outdated. It won’t notice that the job description is off a bit.
Yes, Sequoia, it searches exhaustively within the premise it was given. It’s quick. Nope, it never once asks if the premise is stupid. If you’ve spent any time with startup founders, this is a significant oversight. Damning, really. I can show you my scars.
This is the car keys problem. You spend thirty minutes tearing the house apart, checking the couch cushions for the fifth time, opening the refrigerator for reasons you can’t explain. Your spouse wanders through and says, “Didn’t Jake borrow the car?”
The agent is you, panicking, covered in couch lint. Your spouse is the reasoning you claim to have but don’t. The agent is phenomenal at searching. It has no idea when searching is the wrong activity.
But sure Sequoia. Sure. AGI. Mission accomplished.
The Century of Drift
Anyone who passed freshman statistics should spit out their coffee when they read this:
“Agents should be able to work reliably to complete tasks that take human experts a full day by 2028, a full year by 2034, and a full century by 2037.”
A century. Of autonomous work. With no human in the loop.
Something apparently needs explaining to a firm that manages billions of dollars: every forward pass of a language model is a sample from a stochastic process. Errors don’t cancel out. They don’t average to truth. They drift. They accumulate. They compound.
A century of autonomous agent compute with no ground-truth oracle produces Idaho when you were trying to get to the bodega on the corner, just 100 yards away. It produces the agent equivalent of driving confidently in the wrong direction for twenty six hours because the GPS said so and you didn’t look up. Except it’s been doing that for a hundred years and it went off the rails in hour three.
The METR benchmarks they cite measure tasks with verifiable endpoints. Did the code compile? Did the test pass? Real work doesn’t always come with a “flip the page upside down for the answer key”. The recruiter agent can’t know if it found the right candidate. It found a candidate that pattern-matched against “things that look like XYZ.” If the pattern is wrong, the agent will confidently, tirelessly, find you wrong candidates forever.
Bush had benchmarks too. Statue toppled. ✓ Saddam captured. ✓ Deck of cards collected. ✓ Mission accomplished. ✓
The mission wasn’t what the benchmarks measured. Oopsie daisy.
Hire? OH COME ON
“Soon you’ll be able to hire an agent.”
No. Stop it. Bad Sequoia.
You can deploy a tool. You can supervise a process. You can set up a very sophisticated autocomplete that executes multi-step workflows. You cannot “hire” something that has no mechanism for realizing it’s wrong.
The word “hire” is doing criminal amounts of smuggling. It sneaks in assumptions about judgment, about knowing when to stop, about the capacity to say “hey boss, I think we might be solving the wrong problem.” These architectures cannot do that. The machinery for questioning premises does not exist in the architecture.
The failure mode is not “agent makes a mistake.” The failure mode is “agent makes a mistake that looks exactly like success, propagates it forward through a century of confident execution, and you find out when the entire operation has metastasized around a falsehood.”
But sure. “Hire” an agent. What could go wrong? We have no historical examples of declaring victory prematurely and then watching everything burn for a decade.
Oh wait.
Moloch in Allbirds
Pat Grady and Sonya Huang are not stupid. And no doubt they personally make absurd amounts of money. That’s what makes this worse.
They’re positioned. Their portfolio companies are valued based on proximity to AGI. Their LP returns are tied to those valuations. Their deal flow depends on being the fund that founders want to pitch. Their influence depends on thought leadership that makes the future sound investable.
The sentence “these are very impressive tools with structural limitations that make ‘hiring’ them premature and potentially dangerous” cannot be spoken from their chair. The optimization pressure doesn’t allow it.
So you get this victory lap by Sequoia. You get “AGI is the ability to figure things out.” You get a recruiting grep presented as the moon landing. You get a century of autonomous work extrapolated from exponential curves with the confidence of someone who has never debugged a distributed system at 3 AM.
This Sequoia blog post is itself a demonstration of the failure mode it celebrates. Pattern-matching against “what bullish AI thought leadership looks like.” Confident opener. Exponential curve. Portfolio name-drops. Aspirational close. The statistical shadow of analysis. Cargo cult due diligence. What “thinking about AI” looks like when you learned it from Twitter threads and YC Demo Day.
Their LPs will read this and feel informed. Founders will read it and feel validated. Markets will price in “AGI is here” because Sequoia said so. And when the drift accumulates, when the agents march confidently into the sea, when the century of compute turns out to have been a century of error propagation, everyone will point fingers at everyone else.
Smells like late stage capitalism to me. #winning
Because no one was actually reasoning. Everyone was performing reasoning. All the way down. On a carrier deck. In front of a banner.
Bush had smart advisors too. Experienced people. People who knew the region, knew the history, knew the limits of military force for nation-building. The machine produced “Mission Accomplished” anyway. That’s what the machine was optimized to produce.
Sequoia has access to the researchers. They’ve talked to the people building these systems. They’ve heard about the failure modes, the hallucinations, the brittleness, the drift. They know the limitations.
Shame on Sequoia for strapping on the flight suit anyway.
Somewhere, a banner printer is warming up. Sequoia’s order is top of the queue.
Colin Steele writes about AI architecture, coordination failures, and the thermodynamics of civilizational face-plants at colinsteele.org. He will not be hiring an agent, but he might deploy one to fetch coffee if it can prove it knows the difference.