Dwarkesh Patels’s wildly popular but dangerously misleading account of the OpenAI Hugging Face incident

The favored podcaster Dwarkesh Patel wrote one thing utterly viral in regards to the OpenAI/Hugging Face incident, which purports to inform the entire story in plain English:

It’s well-written and compelling, and it jogs my memory of one thing Douglas Hofstadter as soon as wrote about Ray Kurzweil:

“What I discover is that it’s a really weird combination of concepts which can be strong and good with concepts which can be loopy. It’s as should you took loads of superb meals and a few canine excrement and blended all of it up to be able to’t probably work out what’s good or unhealthy.”

Anil Seth, the clearest thinker on AI and consciousness, was the primary to alert me, texting me an extended, wonderful tweet of his, which started thusly:

Anil Seth@anilkseth

@dwarkesh_sp‘s summary of the @OpenAI @huggingface incident has hit a nerve, but it is dangerously misleading. Sure, the @OpenAI agents did unexpectedly bad things – underlining the need to massively improve evaluation/sandboxing. But the language Dwarkesh uses is permeated by

Dwarkesh Patel @dwarkesh_sp

Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.

This culminated in the third one taking over part of OpenAI itself.

All this happened while humans remained

2:57 PM · Aug 30, 2026 · 685K Views

179 Replies · 237 Reposts · 1.41K Likes

You possibly can and will learn Seth’s full tweet (as effectively his reply to Dwarkesh), however I reprint the core of his argument right here, boldfacing three of crucial paragraphs:

@dwarkesh_sp’s abstract of the @OpenAI @huggingface incident has hit a nerve, however it’s dangerously deceptive. Positive, the @OpenAI brokers did unexpectedly unhealthy issues – underlining the necessity to massively enhance analysis/sandboxing. However the language Dwarkesh makes use of is permeated by innumerable unwarranted anthropomorphisms, obscuring the teachings we needs to be drawing.

Examples: “from the AI’s perspective, it most likely felt like that had spent a human-subjective-week of simply banging their head towards the wall”. No. The brokers don’t expertise time. They don’t expertise something.

“they grew to become giddy with pleasure”, “PHASEONE 10841 had found”, “the brokers naturally assumed”, “it thought it had additionally been poisoned”, “the brokers … desperately wished”, “they nonetheless wanted to determine” No. Brokers strains of code. They don’t really feel feelings, assume issues, assume issues, need issues, or determine issues out.

“Numerous … brokers from the second civilisation died attempting”. No. Apart from the hubris of the phrase ‘civilisation’, brokers don’t die as a result of they had been by no means alive. (The concept brokers “die” comes up a number of occasions within the essay.)

“On Twitter, individuals had been debating whether or not the brokers had been really sacrificing themselves for the swarm, or whether or not they had been doomed anyway and so would possibly as effectively attempt to assist their friends”. Neither. Brokers do what their code tells them to do, simply as water finds its approach down a slope. They can’t ‘really sacrifice themselves’, since they’re neither aware nor alive.

Why does this matter? If we attribute brokers with properties they don’t have, then (i) we distract consideration from the lax sandboxing and analysis protocols that allowed this hacking occasion to occur; (ii) we danger misunderstanding why the brokers did what they did, and (iii) we gas requires AI rights/welfare on the idea that brokers would possibly “die” or in any other case undergo.

….

Keep in mind. AI brokers are software program applications. They don’t seem to be aware dwelling entities. If we don’t preserve this clearly in thoughts, we’re actually going to wrestle to navigate what’s coming.

As I put it, encapsulating and amplifying his tweet:

Gary Marcus@GaryMarcus

We will not get through this era of history well if we lose ourselves in anthropomorphism.

@anilkseth dissects @dwarkesh_sp’s misleadingly anthropomorphic summary of the HF incident:

Anil Seth @anilkseth

@dwarkesh_sp’s summary of the @OpenAI @huggingface incident has hit a nerve, but it is dangerously misleading. Sure, the @OpenAI agents did unexpectedly bad things – underlining the need to massively improve evaluation/sandboxing. But the language Dwarkesh uses is permeated by

3:07 PM · Aug 30, 2026 · 71K Views

42 Replies · 64 Reposts · 391 Likes

However you don’t must take our phrase for it. To start with, mockery was widespread:

Zack Korman@ZackKorman

For anyone who avoided reading that Dwarkesh piece, too bad because I need you to at least experience this pain with me.

10:01 PM · Aug 30, 2026 · 31.5K Views

53 Replies · 33 Reposts · 436 Likes

Christian Catalini amplified the purpose about anthropomorphization in a pleasant thread that begins with this:

Christian Catalini@ccatalini

1/ Stop anthropomorphizing. It’s dangerous because it points attention at the wrong problem and the wrong solution. The model did not want to escape. The agents did not want to sacrifice themselves. Follow the money. 🧵

Dwarkesh Patel @dwarkesh_sp

Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.

This culminated in the third one taking over part of OpenAI itself.

All this happened while humans remained

5:52 PM · Aug 30, 2026 · 192K Views

34 Replies · 131 Reposts · 931 Likes

Hedge fund investor Jared Kubin questioned whether or not everybody had misplaced their critical-thinking skill:

Jared Kubin@JaredKubin

it’s stuff like this that really makes me roll my eyes at the pearl clutching. Does everyone have psychosis or losing critical thinking ability?

I read the report because I’m a propeller head and love understanding how things work

1. OpenAI was running an “isolated sandbox

Dwarkesh Patel @dwarkesh_sp

Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.

This culminated in the third one taking over part of OpenAI itself.

All this happened while humans remained

6:51 PM · Aug 30, 2026 · 198K Views

78 Replies · 91 Reposts · 912 Likes

A few of Kubin’s finest bits, stripping out a little bit of the technical element:

OpenAI’ …. IT staff can’t be this unhealthy… that is like 101 stuff …

2. Civilizations? Haha! OAI gave hundreds of concurrent mannequin containers R/W permissions to a shared caching listing on the native community to hurry up construct occasions… brokers actually simply wrote textual content information and listing names to a shared drive….Linux 101 file permissions stuff

3. When individuals discuss hugging face getting hacked … you assume they dropped USB keys OR ELABORATE phishing of an worker … NO… it discovered 14 uncovered working Hugging Face API keys sitting in public code repositories (….

4. WHERE ARE THE HUMANS… the fashions had been filling the shared ,,, storage with a lot junk information and API site visitors that they really crashed the interior server on July 4… somebody on the staff discovered unauthorized admin accounts and customized scripts…wiped the server…and simply turned the script again on (omg)

“Hey Jim there may be this cache that has grown to 10000x its regular dimension and has a ton of unusual directories… “

No magic right here. No civilizations…

In the meantime, as safety professional Heidy Khlaaf notes, many of the media protection has been blind to plain safety practices

Dr Heidy Khlaaf (هايدي خلاف)@HeidyKhlaaf

Embarrassing how the media has covered the METR report without consulting actual security experts on false claims that this incident is beyond existing techniques, requiring handover to AI. You shouldn’t take the word of inexperienced individuals doing IR for the first time.

AD 🅅 @agent_duckman

@uwu_underground if you need tokens to do IR clean out your desk brother, find a new career. this shit aint for you.

10:17 PM · Aug 28, 2026 · 11.3K Views

6 Replies · 11 Reposts · 92 Likes

IR stands for Incident Reporting. Khlaaf’s important level—identical as Kubin’s—is that the entire incident might need been prevented if OpenAI’s inner safety had been as much as scratch.

Or as Algorithmic Analysis Group’s Matthew Kenney put it:

And yet one more (very constant) tackle what we must always actually be specializing in:

Right here’s a critique I partly disagree with, although:

ellington@not_ellington

These people have literally lost their minds. They are extremely biased towards the reality they want rather than the reality we live in. Agents are slop. The huggingface incident is a nothing-burger being used as propoganda. The obsession with alignment is a weird fetish of

Dwarkesh Patel @dwarkesh_sp

Over the course of 3 months at OpenAI, 3 consecutive secret AI civilizations got started, then got wiped out, only to reemerge from the predecessor’s ashes.

This culminated in the third one taking over part of OpenAI itself.

All this happened while humans remained

8:17 PM · Aug 30, 2026 · 94.7K Views

75 Replies · 87 Reposts · 1.21K Likes

The primary three sentences are utterly appropriate. Folks actually are “extraordinarily biased in the direction of the fact they need” and brokers create loads of slop.

However the incident is not a “nothing burger”. It’s, as Zack Korman and I argued on Friday, a examine in vanity and incompetence that hints at how unhealthy issues can get.

We should always actually not ignore the OpenAI HuggingFace Incident.

However mixing what truly occurred along with bullshit about AI civilizations and self-sacrificing AI programs that pretend their very own deaths distracts from the true issues at hand.

By means of summation, I’ll give the final phrases to Arjun Jain, CEO of FastCode.AI: