I am reminded of when strawberry (01) was rumored to be too dangerous to release by OpenAI back in 2024. Seems like this is a common trick to grab headlines and play off AI doomerism for publicity. Especially with most people feeling uncomfortable about the tech
Their other hackneyed trick is to say “But if AI is regulated, China will win!”
You’d think that the people who have the greatest propaganda instrument the world has ever known at their beck and call would be able to come up with better tricks.
You would think at this point someone would realize that the best way to stand out in the market is simply not to hype. At all. Just be honest. I bet it’d work.
So, you're gonna go against 60 million years of mammalian evolution. The cry of alarm alerts the whole troupe of meerkats that something needs investigation over here RIGHT THIS MINUTE! The group doesn't run over if you say, "Here's something a little bit different."
Actually, we mammals have always been living in the attention economy, but now we have become aware of it.
Well, unlike meerkats, we do have this reflective-thinking thing with the same provenance that we could use to counteract that instinct. I think it’s called “Natural Intelligence.” 😉
And I want to kill every Trumper I see…but somehow I use my will and intellect to settle down and say, “No!…wait till nightfall!”
I joke but seriously we are too quick to say, “Oh, we can’t do X because grandpa got startled in Olduvai Gorge back in the day.” The danger is the false naturalization of sociocultural behaviors or at minimum the downgrading of cortex in favor of brainstem though I know you’ll rightly chide me for such talk, even if meant metaphorically.
Bio/neuro throws down the frame but there’s room for choice on the canvas.
There is a difference between a mammal like a prairie dog which sounds the alarm as a result of instinct and a mammal that sounds the alarm, knowing full well that the danger is not what he claims it to be.
Then again, perhaps we are just dealing with prAIrie dogs in this case. After all , the resemblance is uncanny.
I think that Dario Amodei is too dangerous to release. Someone should put some fences and guardrails around him, because he clearly isn't aligned with other human beings.
Pretty much. They had a bunch of people whose job it was to tell them how to avoid risks when there was not much money in it, then fired those people when they realized there could be money in it. The ones who are left are a bad joke, like the Meta head of AI safety or whatever who gave some OpenClaw nonsense access to their email and was surprised when it started sending commands to delete things. The job description of those who remain is try to twist whatever the company decides to sell into something that they can market as safe regardless of truth, so they're like less self-aware PR professionals.
Your dog knows it will be fed by you, it begs you for extra treats. OpenAI and Anthropic in contrast are frantically trying to secure their basic food supply from an increasing skeptical audience.
Exactly! This is just another one of Anthropic's desperate pleas for mass market validation and to build more hype to say the least.
They can keep clawing all they want (pun-intended).
Heaps to be done on the architectural side of things. And this bluff of Mythos or rather "MythOS" is just another small incremental improvement based on flawed architectural foundations.
Thanks for collecting something closer to real evidence. OpenAI said that GPT 2 was too dangerous and would not be released. Anthropic continues to say that their models are too dangerous, that they will blackmail you. Amodei does not know whether his models are conscious. Is there really any change here or was the benchmark simply being gamed? Did it identify true day-0 exploits that no one had ever seen before? Or did it mimic with some stochasticity known exploits?
They fill the air with hype, but these are not the droids you are looking for.
The 244(!) page document is more reasonable than the marketing blurb on the website which is more reasonable than the reporting everywhere. Take:
"Early claims of large AI-attributable wins have not held up.
In the initial weeks of internal use, several specific claims were made that Claude Mythos Preview had independently delivered a major research contribution. When we followed up on each claim, it appeared that the contribution was real, but smaller or differently shaped than initially understood (though our focus on positive claims provides some selection bias). In some cases what looked like autonomous discovery was, on inspection, reliable execution of a human-specified approach. In others, the attribution blurred once the full timeline was accounted for."
And apparently: no, AI was not instrumental in improving AI:
"The gains we can identify are confidently attributable to human research, not AI assistance. We interviewed the people involved to confirm that the advances were made without significant aid from the AI models available at the time, which were of an earlier and less capable generation. This is the most direct piece of evidence we have, and it is also the piece we are least able to substantiate publicly, because the details of the advance are research-sensitive."
But who does read that document and makes a measured assessment when you can create a panic?
And the word 'compute' only appears once in that 244-page document and sadly not to explain how much compute they threw at it to create this result. Because then the answer of Tom Fridman's boy maybe should have read "I spent a cool couple of millions to exploit vulnerabilities in a browser with its security features turned off". 'Not hyperbolic', my ass.
Oh, for goodness sake, pay no attention to that man behind the curtain -- this is a cultural revolution, a great leap forward from Oz the Powerful ... and don't forget to exit through the gift shop with your IPO purchase.
Meanwhile, with the focus on Anthropic, OpenAI/Sam Altman is backing an Illinois bill that would “Limit Liability for AI-Enabled Mass Deaths or Financial Disasters” (Wired)
If AI can show vulnerabilities, it can presumably also find ways to mitigate them. Right now, the greatest danger of AI is the risk that we drown in the BS generated by AI companies and their client whorenalists.
Yes, you could most probably be right about this Gary. The announcement does seem to be an overblown marketing gimmick and nothing more.
Also, Anthropic's desperate pleas for mass market validation and to build more hype to say the least.
They can keep clawing all they want (pun-intended).
Heaps to be done on the architectural side of things. And this bluff of Mythos or rather "MythOS" is just another small incremental improvement based on flawed architectural foundations.
The pattern Gary identifies here is real — every major AI lab has learned that "dangerous capabilities" announcements drive more press than benchmark numbers. The irony is that this inflates public fear while simultaneously making the actual risks harder to parse. When everything is existential, nothing is. Good reality check.
I am reminded of when strawberry (01) was rumored to be too dangerous to release by OpenAI back in 2024. Seems like this is a common trick to grab headlines and play off AI doomerism for publicity. Especially with most people feeling uncomfortable about the tech
Their other hackneyed trick is to say “But if AI is regulated, China will win!”
You’d think that the people who have the greatest propaganda instrument the world has ever known at their beck and call would be able to come up with better tricks.
Let's hope they don't.
they said *gpt2* was too dangerous to release lol
You would think at this point someone would realize that the best way to stand out in the market is simply not to hype. At all. Just be honest. I bet it’d work.
So, you're gonna go against 60 million years of mammalian evolution. The cry of alarm alerts the whole troupe of meerkats that something needs investigation over here RIGHT THIS MINUTE! The group doesn't run over if you say, "Here's something a little bit different."
Actually, we mammals have always been living in the attention economy, but now we have become aware of it.
Well, unlike meerkats, we do have this reflective-thinking thing with the same provenance that we could use to counteract that instinct. I think it’s called “Natural Intelligence.” 😉
But reflective thinking follows the mammalian startle response; it does not replace it.🚨🚨🚨
And I want to kill every Trumper I see…but somehow I use my will and intellect to settle down and say, “No!…wait till nightfall!”
I joke but seriously we are too quick to say, “Oh, we can’t do X because grandpa got startled in Olduvai Gorge back in the day.” The danger is the false naturalization of sociocultural behaviors or at minimum the downgrading of cortex in favor of brainstem though I know you’ll rightly chide me for such talk, even if meant metaphorically.
Bio/neuro throws down the frame but there’s room for choice on the canvas.
Time for us to evolve and educate beyond our human frailties so the suckers—though a new one is born every minute—get wise before it's too late.
There is a difference between a mammal like a prairie dog which sounds the alarm as a result of instinct and a mammal that sounds the alarm, knowing full well that the danger is not what he claims it to be.
Then again, perhaps we are just dealing with prAIrie dogs in this case. After all , the resemblance is uncanny.
The question is: Can Mythos count the number of r's in hyperbolic?
Does mythos know that it has myth in its name?
Good question!
I think that Dario Amodei is too dangerous to release. Someone should put some fences and guardrails around him, because he clearly isn't aligned with other human beings.
At the AI companies it’s all about aligning the humans with the bots rather than the other way around
Pretty much. They had a bunch of people whose job it was to tell them how to avoid risks when there was not much money in it, then fired those people when they realized there could be money in it. The ones who are left are a bad joke, like the Meta head of AI safety or whatever who gave some OpenClaw nonsense access to their email and was surprised when it started sending commands to delete things. The job description of those who remain is try to twist whatever the company decides to sell into something that they can market as safe regardless of truth, so they're like less self-aware PR professionals.
This need for constant noise on social media to attract attention is annoying... My dog is less needy...
Your dog knows it will be fed by you, it begs you for extra treats. OpenAI and Anthropic in contrast are frantically trying to secure their basic food supply from an increasing skeptical audience.
I hear Walmart is having a two for one sale on Chatbot kibbles.
… as Dario runs down to beat Sam to the stock.
I bet your dog also doesn’t keep barking up the wrong tree
Exactly! This is just another one of Anthropic's desperate pleas for mass market validation and to build more hype to say the least.
They can keep clawing all they want (pun-intended).
Heaps to be done on the architectural side of things. And this bluff of Mythos or rather "MythOS" is just another small incremental improvement based on flawed architectural foundations.
On a related note, these are my thoughts around LLMs across the board in general - https://ajlewis90.substack.com/p/why-current-ai-models-are-still-inefficient
Thanks for collecting something closer to real evidence. OpenAI said that GPT 2 was too dangerous and would not be released. Anthropic continues to say that their models are too dangerous, that they will blackmail you. Amodei does not know whether his models are conscious. Is there really any change here or was the benchmark simply being gamed? Did it identify true day-0 exploits that no one had ever seen before? Or did it mimic with some stochasticity known exploits?
They fill the air with hype, but these are not the droids you are looking for.
The 244(!) page document is more reasonable than the marketing blurb on the website which is more reasonable than the reporting everywhere. Take:
"Early claims of large AI-attributable wins have not held up.
In the initial weeks of internal use, several specific claims were made that Claude Mythos Preview had independently delivered a major research contribution. When we followed up on each claim, it appeared that the contribution was real, but smaller or differently shaped than initially understood (though our focus on positive claims provides some selection bias). In some cases what looked like autonomous discovery was, on inspection, reliable execution of a human-specified approach. In others, the attribution blurred once the full timeline was accounted for."
And apparently: no, AI was not instrumental in improving AI:
"The gains we can identify are confidently attributable to human research, not AI assistance. We interviewed the people involved to confirm that the advances were made without significant aid from the AI models available at the time, which were of an earlier and less capable generation. This is the most direct piece of evidence we have, and it is also the piece we are least able to substantiate publicly, because the details of the advance are research-sensitive."
But who does read that document and makes a measured assessment when you can create a panic?
And the word 'compute' only appears once in that 244-page document and sadly not to explain how much compute they threw at it to create this result. Because then the answer of Tom Fridman's boy maybe should have read "I spent a cool couple of millions to exploit vulnerabilities in a browser with its security features turned off". 'Not hyperbolic', my ass.
Thanks for this — I didn’t realize they were so candid in their system card
I feel so lucky to live in a time where I get to bear witness to the continual scaling up of mediocrity via compute and hype.
"Roll up! Roll up! I have a deal so outrageously good that it's illegal for me to even show it to you!" - type of vibes.
Call to Action.
Hyperbole.
"Forbidden Fruit."
Tricks of the trade as old as carnivals, perpetual energy devices and snake oil.
Oh, for goodness sake, pay no attention to that man behind the curtain -- this is a cultural revolution, a great leap forward from Oz the Powerful ... and don't forget to exit through the gift shop with your IPO purchase.
Meanwhile, with the focus on Anthropic, OpenAI/Sam Altman is backing an Illinois bill that would “Limit Liability for AI-Enabled Mass Deaths or Financial Disasters” (Wired)
https://www.wired.com/story/openai-backs-bill-exempt-ai-firms-model-harm-lawsuits/
Financial disasters? Like that caused by the current AI bubble?
AI enabled mass deaths? Like those that will be the inevitable outcome of a lack of all AI regulation?
I thought OpenAI was supposed (indeed REQUIRED by their charter) to put safety — not escape from accountability — first.
If AI can show vulnerabilities, it can presumably also find ways to mitigate them. Right now, the greatest danger of AI is the risk that we drown in the BS generated by AI companies and their client whorenalists.
Yes, you could most probably be right about this Gary. The announcement does seem to be an overblown marketing gimmick and nothing more.
Also, Anthropic's desperate pleas for mass market validation and to build more hype to say the least.
They can keep clawing all they want (pun-intended).
Heaps to be done on the architectural side of things. And this bluff of Mythos or rather "MythOS" is just another small incremental improvement based on flawed architectural foundations.
These are my thoughts around GenAI in general - https://ajlewis90.substack.com/p/truly-agentic-ai-cannot-be-built
Houston, on second check, the toilet is working.
The pattern Gary identifies here is real — every major AI lab has learned that "dangerous capabilities" announcements drive more press than benchmark numbers. The irony is that this inflates public fear while simultaneously making the actual risks harder to parse. When everything is existential, nothing is. Good reality check.
Next article: Three reasons to think that Gary Marcus's ego and sense of self-importance is overblown 😂