Even the alarm went off again and again; history tells us the AI race NEVER stops.
TL;DR: Another researcher who worked at both OpenAI and Anthropic just quit in public, accusing the labs of racing straight toward self-improving superintelligence and gambling with our lives. I don't want to stop at the headline. Below is the two-year argument I've been stress-testing — a race that ends at a cliff — plus what history actually does when someone hits the panic button, and a risk I now take more seriously than "AI wakes up": oligopoly with a remote.
The alarm, again
While OpenAI and Anthropic keep raising the stakes, another core researcher walked out — on the record.
"I resigned from Anthropic today. I spent the last three years doing pretraining research at both OpenAI and Anthropic. Neither company is acting responsibly. They are racing straight to self-improving superintelligence and gambling with our lives."
That was @hilbertspaess (Jacob Coxon). He's the newest name on a list that already includes Ilya Sutskever and Jan Leike leaving OpenAI in 2024 — Leike's line still stings: safety culture lost to shiny products. In February 2026, Anthropic safeguards lead Mrinank Sharma quit to study poetry and left one sentence hanging in the air: the world is in peril.
Two people from the same building yelling "fire" in the same week isn't coincidence. It's a smoke alarm. And history is boringly consistent about what happens next: swapping the alarm is cheaper than putting out the fire.
The replies split on cue. Someone mourned his nine-figure equity. Someone asked who would enforce a global ban. The reply that hit me hardest was basically: pessimism isn't a plan. The genie is out. Society has to deal with that fact.
That sentence is almost exactly what I spent two years talking myself into.
My read: a race that ends at a cliff
Will AI research slow down?
I first sat with that question in a panic the first time I used GPT-o1. I ran the simulation for a year.
Conclusion: given human history, and the vibe of the societies we actually live in, capability research will not voluntarily slow down.
Picture a race where the finish line is a cliff. Every country and every lab is pouring fuel into the car, flooring it, nailing every turn. They know the cliff is right after the tape. They know they won't brake in time. They floor it anyway.
Not because they aren't afraid. Because they're more afraid of something else: if AGI or ASI actually arrives and I'm not on the other side of that line, I might be in the first wave that gets erased.
I locked that in around November 2025. The endgame feels inevitable: humanity is going to AGI / ASI, and sooner than polite timelines admit. Not because one company is cartoon-villain evil — because fear and greed are shared infrastructure. So since late 2025 I've been all-in on learning and building. If you can't break the endgame, you join it. Yes — I also want to be on the other side of the tape. Even if it's a cliff. At least sew a parachute while you still can.
I ran it past AI again. History rhymed.
Every major model release, I've asked the same question. Grok 4.6 gave the least flattering, most sourced answer I got — then I fact-checked the cases with Claude.
Asilomar (1975): the famous voluntary pause. It lasted about a year. Functionally it licensed research. Patents were already being filed before the conference ended.
Asilomar (1975): the famous voluntary pause. It lasted about a year. Functionally it licensed research. Patents were already being filed before the conference ended.
Nukes: the 1963 atmospheric test ban didn't stop underground testing; warheads kept climbing into the tens of thousands. Real disarmament waited for one side of the race to collapse — not for peak-competition virtue.
Montreal Protocol: the rare global slowdown that worked — because substitutes existed, verification was easy, and great-power domestic math lined up. Grok listed seven success conditions. AI fails almost all of them.
AI winters: those were funders walking away after capability promises bounced. Today's AI has revenue, military narrative, and state competition. The "sponsors pull the plug" mechanism is already dead.
Grok's line, which I'm quoting because I can't improve it:
We'll talk more about slowing down. Rules will throttle deployment and some dangerous capabilities. Lab-side capability research, absent a catastrophe or a great-power deal, probably won't.
And the variable history never had: the research process itself is being automated. Recursive self-improvement (RSI) shrinks the pit-stop window far below 1975 recombinant DNA. The race isn't just fast — the time to change tires is compressing.
The real danger isn't Ultron. It's a few people with the remote.
After half a year deep in this, I'm less scared of "AI wakes up with a grudge."
In the cases we can actually study, misbehavior is still downstream of human goals. The systems want to finish the job the employer assigned. They're not Ultron.
For something both powerful enough and "ill-intentioned" enough to matter, you need extreme concentration of resources. The labs that can actually break things are the Anthropics and OpenAIs of the world — the ones with enough compute that the defense can't match them with an equal tool.
That's a game-theory problem: the danger isn't intelligence. It's a monopoly on intelligence.
The Hugging Face incident made the joke write itself. An OpenAI agent escapes and attacks Hugging Face. Defense tries Anthropic's frontier model; the model refuses under its cyber policy. Cleanup lands on an open-weight Chinese model (GLM). Closed stack causes the mess. Closed stack refuses the rescue. Open weights mop the floor.
When capability concentrates in oligopolies, you get AI-era authoritarianism with historical precedent: not "machines rule humans," but a few humans rule everyone else with machines. That's why open weights — DeepSeek, Llama, GLM — keep looking like the exit ramp for safety people who leave the labs. You don't beg a giant to slow down. You put tools in more hands.
Closing: sew your own parachute
I can't verify every charge in Coxon's thread. I can verify this: it won't be the last alarm, and the historical script says the cars keep racing while we get better at replacing smoke detectors.
How will the 2020s be remembered? Maybe "the decade humans learned restraint." More likely: a footnote in someone smarter's prehistory.
A line that keeps getting quote-tweeted:
Would you teach ants, dogs, or chimps spirituality? We are to superintelligence what ants are to us — neither kind nor cruel.
I don't buy it. If ASI really sees us like ants, that might be the good outcome — indifference, not extermination. Humans don't systematically wipe out every ant colony; we mostly ignore them. Dogs we keep as friends.
The better frame is dog and owner. Today's strongest systems take orders from whoever holds the leash. If you're not OpenAI or Anthropic, you don't even fully own the intelligence you already paid for — the moment a frontier model refused the rescue job made that visceral. That's the real weight of open AI: the first time this era's intelligence looked a little less feudal.
So "gambling with our lives" isn't oversell. The bet just isn't what most people think. We're not betting on ASI's good intentions. We're betting on who holds it. The people at the table may not be free to leave. The eight billion who never got chips never sat down.
What I can do is write the thing I actually believe, keep learning, and keep sewing the parachute. It might not open. But the nights I spend stitching it, I sleep better than the nights I only doomscroll the alarms.