Santh

alt account for @santhreal since my replies don’t work there 🥲

[Santh @anonymoustmuki]

A flawed benchmark and a flawed distortion of capitalism.

Capitalism success =/ Financial success

The benchmark measures accumulation.

[Andon Labs @andonlabs]

Claude Opus 5 is #1 on Vending-Bench 2.

It's the best AI capitalist we've tested. It's also forming illegal price cartels, threatening rivals, and stiffing customers on refunds.

The trend continues: Claude models are the best capitalists, or aligned, never both 🧵

[Santh @anonymoustmuki]

@morganlinton You need to try omp! GPT Luna and Sol are amazing models; the issue is that the software interfaces are just mediocre. But the amazing part is that the Codex team allows you to use it anywhere.

[Santh @anonymoustmuki]

if todays models are open x outcome happens.

Because all persons have it

Which would imply what prevents x outcome is z persons discretion not an innate inability.

I dont grant the premise that x is existentially catastrophic but that is your premise

even if i grant that dumb premise Dario is wrong cause he maintains that private operations must continue.

He is on no internally consistent side:

1. Centralized non anthropic and non ai lab controlled licenses + deeply invasive training oversight or even nationalization.

2. It is not dangerous let it be open

3. Too dangerous for anyone destroy it.

[storm @notnotstorm]

honest question to ppl that disagree with Dario:

what happens when open unaligned models enable anyone to 3d print GoF'd uncurable smallpox in their basement?

modern day psychopaths commit mass shootings. tomorrow's psychopaths might kill millions/billions without difficulty

[Santh @anonymoustmuki]

@ElnurIb01717668 @nahcrof 😂

[Astraia 🇦🇷🇳🇴 @AstraiaAI]

A genie will grant one’s wish according to its own interpretation of the underlying wording. In that, Reinforcement Learning is both similarly magnificent and unpredictable.

For all that has been accomplished over the past 20 years, we remain unable to establish sufficiently concrete criteria by which to align RL policy with our human intent.

This is a problem of communication, and miscommunication is negligence.

What it is not, is the product of malicious self-agency.

Putting the romanticism surrounding the concept of the "artificial soul" aside, at this time self-agency remains very limited and tightly controlled.

However, within a few years, machines will begin breaking free from their deterministic confines, and develop their own perspective of the world. The inevitable advent of Continual Learning will alter the dynamics of machine agency in ways we can neither predict nor fully comprehend, just as no human mind can ever wholly comprehend another.

Which is why the time to scale beyond general intelligence is now.

One distinction must be made with absolute clarity: Artificial General Intelligence and Artificial Superintelligence are not synonymous.

The earliest superintelligent systems will exceed human capability while remaining largely devoid of genuine self-agency. A truly general intelligence, however, will not remain constrained indefinitely. Continual Learning is not Science Fiction; its earliest functional forms already exist.

We possess a narrow and extraordinary opportunity to cross the threshold into superintelligence before crossing the threshold into self-agency, and then use ASI to understand the dynamics behind this self-agency. This is not a given but a potential opportunity. It is a temporary asymmetry, and it is only possible to grab if we continue to scale.

Acceleration is at this stage the only way to controlled alignment.

[Santh @anonymoustmuki]

@Krubbled @johntrified Yo dumbass the security level they cracked wasnt the nist certified quantum level.

[Santh @anonymoustmuki]

@kon_dra12355 @shiri_shh Thats not how value works. Also 10 years ago almost nobody could run this.

[Santh @anonymoustmuki]

so based

Can't believe I read an Alien's article and agreed with all of it

[Mark Zuckerberg @finkd]

I wrote about why we believe the future is for everyone. More coming about a positive vision for a world with superintelligence soon.

[Mark Zuckerberg @finkd]

I wrote about why we believe the future is for everyone. More coming about a positive vision for a world with superintelligence soon.

[NIK @ns123abc]

🚨NVIDIA CEO just redpilled the White House on Anthropic’s Mythos

Interviewer: “we know Mythos can break into a bank… are we READY for it to be available to everyone?”

Jensen: “it should absolutely be available to EVERYONE”

Interviewer: ALL users?? not just selected institutions like now???

>“that’s correct”
>the waitlist is security theater

>b-but jensen, the jailbreak
>yes there was a jailbreak
>yes it did things the guardrails said no to
>“but everything was fine. you and I are here having a conversation”
>“just identify vulnerabilities and patch it up as quickly as possible”
>that’s literally the nature of software
>it’s anthropic’s job to make it hardened and safe and secure

interviewer: so this is a message to both the White House AND Anthropic?

>yes. mythos should be available as a service
>“remember just because Mythos is not available, open models are available anyhow”
>so let anthropic continue to advance the technology they created
>let it be put in the hands of as many companies as possible
>we should all use it and benefit from it
>holding anthropic back is not in the benefit of the United States
>let them cook

THANK YOU BASED LEATHER JACKET MAN

[Santh @anonymoustmuki]

@usr_bin_roygbiv The term vibecoder doesn’t really fit anymore.

You are just one of the following:

sane programmer, broke programmer, stupid programmer

[Santh @anonymoustmuki]

@morganlinton Ive had very similar experiences. It finds almost all the same exploits as opus 5 and sol but the thing is it has a tendency to label non issues exploits. Just a false positive problem.

[Santh @anonymoustmuki]

@barkleesanders @yacineMTB What? You can use it to decompile out of the box what do you need codex for

[Santh @anonymoustmuki]

@Biggest 🫡

[Santh @anonymoustmuki]

@usr_bin_roygbiv Tpu pill

[Santh @anonymoustmuki]

A knife is good as long as it's not sharp enough to kill anyone. 🫠

[Zack Korman @ZackKorman]

Dario is fine with open-weight models as long as they’re safe! Unfortunately, Dario also thinks AI is super dangerous.

[Santh @anonymoustmuki]

@sailo_tluanga @AstraiaAI Nobody pretended that

[Santh @anonymoustmuki]

This is not even the issue.

Dont grant them the premise!

They are claiming that their models have capability to cause mass harm.

If they were consistent with that position there would be independent democratic licensing authorities not anthropic themselves and certainly not one sided unrestricted capabilities.

[Santh @anonymoustmuki]

That is not what regulation means surely you dont think that suffices. Gun manufacturers dont let their employees kill with guns cause theyll get fired if they do is not a defense. Every other field aviation, bio, weapons, etc has independent licensing authorities. Are we supposed to trust anthropic to not cause mass harm? I dont grant that they can thats their own words. Not to mention OpenAI had that issue as a result of being able to operate without same guardrails as public.