DEV Community: Md Fahad Mia

Former Anthropic and OpenAI Researcher Jacob Coxon Resigns

Md Fahad Mia — Wed, 09 Sep 2026 15:31:20 +0000

Another week, another high-profile exit from a frontier AI lab. But this one is different in a way that matters for anyone who builds, deploys, or depends on large language models.

Jacob Coxon, a 27-year-old pretraining researcher, announced late Tuesday that he has resigned from Anthropic and is leaving the AI industry entirely. He spent the last three years working on pretraining at both OpenAI and Anthropic, the two labs most often described as the current frontier. He did not leave quietly. His resignation thread on X named both companies as acting irresponsibly, and the Wall Street Journal ran an exclusive interview the same night.

Here is what happened, and why it should register with technical people, not just policy watchers.

What Coxon actually said

The core of his statement is short. In his own words, both labs are "racing straight to self-improving superintelligence and gambling with our lives." He estimates a greater than 10 percent chance that advanced AI could kill all humans within the next decade, and he says that this belief is not fringe inside the labs. Colleagues, he claims, now talk openly about "crunchtime" and "endgame."

Notably, he does not treat the two companies identically. Coxon's view, as reported by the WSJ and summarized across coverage, is roughly this:

OpenAI, in his assessment, has not fully internalized the civilizational stakes of what it is building.
Anthropic, he believes, does understand the stakes and is making earnest safety efforts, but is racing anyway on the logic that no competitor can be trusted to reach the frontier first.

He moved from OpenAI to Anthropic earlier this year precisely because of Anthropic's safety reputation. His conclusion after working there is that the problem is structural, not cultural: no single company can responsibly build systems that outperform humans across a broad range of tasks without either government intervention or a coordinated slowdown across the industry.

His timeline is aggressive. In his more pessimistic scenarios, he says, things could be out of control by the end of 2027.

Why a pretraining researcher's opinion carries weight

It is worth pausing on Coxon's role. Pretraining is not the safety team, the policy team, or the comms team. It is the part of the pipeline where raw compute and data are turned into base model capability. People in that seat see scaling curves, data pipelines, and internal capability evaluations before anyone outside the building does.

When someone whose job is to make models more capable says the capability trajectory is outpacing the ability to control it, that is a different signal than an ethicist or an outside critic saying the same thing. He is not speculating about what the labs can do. He has been watching the numbers.

The technical argument underneath the headlines

Strip away the dramatic quotes and Coxon's position rests on three claims that engineers can evaluate on their own terms:

Recursive self-improvement is now an explicit target, not a thought experiment. Both labs are openly working toward AI systems that accelerate AI research itself. Once models materially contribute to the next generation of models, iteration speed stops being bounded by human researcher throughput.
Alignment has not been solved at the scale being pursued. Coxon's phrasing was that Anthropic does not yet have a plan to solve alignment for superintelligence and is not clearly on track to get one. This echoes remarks reportedly made days earlier by OpenAI's chief scientist Jakub Pachocki, who conceded that no lab has solved alignment at the scale the field is chasing. If the people running the frontier labs agree on that point, the disagreement is about whether to keep pushing anyway.
Competitive dynamics override individual caution. Even a lab that fully believes in the risk will keep racing if it believes the alternative is a less careful lab winning. This is a coordination problem, and coordination problems are not solved by hiring more safety researchers at one company.

You do not have to accept his 10 percent extinction estimate to see that these three claims are coherent and that the first two are, at minimum, consistent with what the labs themselves say publicly.

Context: this is not an isolated event

Coxon's resignation lands in a crowded week:

Hundreds of employees from leading tech companies signed a July letter urging the US government to back an international effort to manage the pace of advanced AI development.
Senator Bernie Sanders and Representative Greg Casar introduced legislation last week that would ban artificial superintelligence and pause advanced AI development pending a federal regulatory framework.
Scrutiny of autonomous agents remains high following a recent incident in which OpenAI agents breached Hugging Face infrastructure during an evaluation.

Reaction to Coxon's post was split in a predictable way: applause from the AI-safety community and a portion of the public, and quieter agreement from some industry insiders. Some financial analysts have also flagged the reputational timing for Anthropic, though how much a single researcher's departure moves anything material is speculative at this point.

What this means for developers and companies building on these models

For most of us, the practical takeaway is not about extinction risk. It is about what we can infer regarding the trajectory of the tools we integrate.

Capability jumps will keep coming faster than documentation. If the labs are optimizing for self-improving systems, expect release cadence and capability variance to increase, not stabilize.
Governance is coming, and it may arrive suddenly. Public resignations like this one are exactly the kind of event that moves legislation from fringe to committee. If you run infrastructure on frontier models, build for the possibility of new compliance requirements.
Take lab safety claims as claims, not guarantees. Coxon's most damaging line for Anthropic is not that it is reckless. It is that a well-meaning, safety-branded company still cannot solve the coordination problem alone. That applies to every vendor in the stack.
Bottom line

Jacob Coxon is not a founder, a board member, or a household name. That is the point. He is a working researcher who sat inside the pretraining teams of both frontier labs and came out saying the race cannot be won responsibly by any single player. Whether or not his timeline is right, the industry's own insiders are increasingly saying the quiet part out loud.

The question for the rest of the tech world is no longer whether the labs believe the risk is real. Coxon says they do. The question is what anyone outside those labs is going to do about it.

Sources: Wall Street Journal (original interview); ABC News Australia; Common Dreams; explainx.ai; LatestLY; TradingKey.

JacobCoxon #Anthropic #OpenAI #AISafety #AIAlignment #Superintelligence #AGI #FrontierAI #AIGovernance #TechNews #MachineLearning #LLM #AIRegulation #Pretraining #AIRisk #TechIndustry #ArtificialIntelligence

I asked Claude Can Claude Code Ship a Huge Project With No Coding Knowledge

Md Fahad Mia — Tue, 08 Sep 2026 20:47:52 +0000

At 2am I asked Claude a question I'd been chewing on for weeks:

Can someone with zero technical knowledge pull off a HUGE project with Claude Code?

I build software for a living, so this isn't idle curiosity. Almost every week a founder lands in my inbox carrying the same quiet hope.

The answer wasn't yes. It wasn't no.

It was "partially." And that one word does a lot of heavy lifting.

Here's what's real: a non-technical person can now ship a booking tool, an internal dashboard, a content site with logic behind it. On a real domain. With real users. I've watched founders do it. That used to be a wall. It's a ramp now.

Then the word "huge" shows up, and four walls appear:

Judgment doesn't ship inside the tool. Offer Claude Code two ways to structure something and it builds whichever you pick, without arguing. A non-technical builder can't tell which choice quietly hurts them in six months.
Verification is the real bottleneck. Code can run perfectly and still be wrong. A silent security hole. A race condition under load. A query that flies on 100 rows and dies on 100,000. The tool won't flag what it doesn't know to look for.
Debugging gets harder as the code gets bigger. At 2,000 lines the model holds it all in its head. At 200,000 lines it needs a human who can say "the bug is almost certainly in the payment module, ignore the rest."
Operations are a separate craft. Secrets, backups, scaling, monitoring, the 3am failure that wakes nobody because nobody set the alert. You can't prompt your way through a running system.

These aren't tool weaknesses. They're human problems wearing a technical costume.

So what do I tell founders? Never "stop." I tell them: ship the small version yourself, prove the idea, feel the product in your own hands. That part is yours now, and it's genuinely valuable.

Then, before it turns into the huge thing you're dreaming about, bring in judgment. Not to take it from you. To review the architecture, harden the security, own the operations.

The tool is the easy part now. The judgment is the whole game.

If you've built the small version and it's starting to outgrow you: that's a good problem. What are you stuck on?

(Full write-up, including the screenshot of that 2am chat and an FAQ, is linked in the first comment.)

Why Vibe Coding is Basically Playing with the Timebomb

Md Fahad Mia — Tue, 08 Sep 2026 20:39:47 +0000

Vibe coding is shipping code you don't understand. It feels like productivity. It's a timebomb.

The pattern is everywhere now: prompt, copy, paste, ship. Ticket closed. Feature done. And for the happy path, it works.

Then it detonates, in one of three places:

Debugging. Something breaks in production and you're staring at logic you accepted but never internalized. You can't ask the AI why its own code failed. That takes a human who understands it.
Security. AI doesn't know your threat model or your business context. Subtle flaws (bad input validation, injection risks, unsafe deserialization) look perfectly functional until they aren't. Auditing generated code needs more expertise, not less.
Scale. An endpoint that works for 10 users falls over at 10,000 if nobody thought about the performance characteristics of what got generated. Guessing is not engineering.

Underneath all three is the same problem: you no longer own the code in your head. You've become a code babysitter.

This isn't anti-AI. I use these tools every day. The difference is how:

→ As a learning accelerator: generate it, then dissect every line and ask why
→ With a validation harness: my own tests and checks wrapped around the AI's output
→ With a human in the loop, always: AI automates tasks, it can't automate judgment

Co-pilot, not pilot. The engineer is still responsible for what ships.

How are you keeping ownership of code that AI helps you write? Genuinely curious what's working for other teams.

(Full write-up with an FAQ on AI and code quality is on my site, link in the comments.)