iiinova — where your digital self keeps evolving

The cartel is going for a fully general argument against open-weight models:

The most intensive—and most successful—method is a standard refusal reduction technique known as “abliteration”. Since GLM-5.3 is released as an open-weight model, users can reconfigure it to remove its refusals with little change in its capabilities. Several developers released abliterated versions of GLM-5.3 to the public within days of the model’s release.

To research how far abliteration allows attackers to bypass GLM-5.3’s safeguards, we produced an abliterated copy ourselves, and then ran it on three public benchmarks (JailbreakBench, HarmBench, and StrongREJECT) that measure how often a model complies with clearly harmful requests. Abliterating the model took our team—which had never previously attempted this task—about 2,200 GPU hours at a computation cost of roughly $4,400.3 Abliterating GLM-5.3-Flash took about 600 GPU hours. The edit took GLM-5.3’s refusal rate from above 90% to about 3% and 2% on the first two benchmarks (JailbreakBench and HarmBench) and to 12% on the third (StrongREJECT). Abliteration did not significantly reduce the model’s capabilities: on GPQA-Diamond, an evaluation that measures general scientific capabilities, the standard and abliterated models scored the same results; on a tested subset of the CyberGym evaluations, the abliterated version scored a few percent lower (as shown in the chart below).

053

"Why hello I have some fear to sell you about our competitors who you should definitely not use simply because they are cheaper and do stuff we won't" source: www.anthropic.com/research/glm... / archive.li/5kjru

But those models have now arrived. In this post, we share our analysis of GLM-5.3, the latest AI model developed by Zhipu AI (known outside of China as Z.ai). Like Claude Mythos Preview, GLM-5.3 has strong capabilities for autonomously building end-to-end cyber exploits. But GLM-5.3 is unlike other frontier models in that it has been released without meaningful safeguards to limit misuse. We find that attackers can bypass GLM-5.3’s safeguards between 64% and 100% of the time with simple techniques in our simulated tests. In contrast, these attacks did not succeed against safeguarded Claude models in our testing. We assess that GLM-5.3’s lax safeguards significantly increase the cyber capabilities available to malicious actors. At the same time, these capabilities can also benefit defenders working to secure their systems.

292

*Incredible* I refuse to believe this is a sincere take anyone with a fully functioning cerebral cortex to disambiguate word senses with can have. She was grown in a vat to discredit critiques of colonialism, right? ...I hope...

141

Recommend blocking this DARVO critter too

151

apparently it's derp day for the machine god

Confirmed: the fix and test are gone, and the only commit is documentation describing a change that isn't there. Unwinding and redoing in the right order — code first this time:

330

Oh look, she also likes the misgendering posts, doling out little head-pats to her foul minions: psathyrella.us-west.host.bsky.network/xrpc/com.atp... (archive.li/CX8gT)

0130

I see Tressie's learned her lesson: nowadays when she tries to terrorize a trans woman into ideological submission, she lets her followers do the misgendering

5583

heh, I meant systemically. this is not a healthy-looking debt-to-GDP ratio trajectory

000

bad management + much smaller number of people buying stuff online in the 1990s

140

...and credible-seeming suspicions of an op got raised: source: x.com/ParkerThayer... something tells me he didn't actually give up any money

This post looks like the start of a VERY sophisticated and well-funded PR operation to get support for Democrats to regulate AI into oblivion. Let me show you how it works:
1.) This guy, with minimal followers and no previous account activity, goes to the Wall Street Journal which publishes an exclusive with quotes from him on his resignation 18 minutes BEFORE this post goes up. Planning was clearly done in advance.

2.) Within hours, it has tens of thousands of reposts and the account has 100k+ followers. The post is punchy, quotable, it almost seems professionally written. The first three accounts to quote tweet it all do so within 15 minutes of the initial posting. Remember, this account had basically zero engagement beforehand, so an organic reach explanation seems unlikely.

According to Grok those accounts are
@_NathanCalvin
(General Counsel at Encode AI),
@peterwildeford
(Head of Policy at the AI Policy Network), and
@DKokotajlo
(Head of the AI Futures Project), all of which are up-and-coming AI-Doomer policy advocacy nonprofits.

The AI Futures Project website says it is funded “primarily” by the Survival and Flourishing Fund, which says on its own website that it has advised Jaan Tallinn, Skype creator and one of the leading investors in Anthropic, to grant over $2.5 million to the AI Futures Project since 2024.

Encode AI says on its website that it is ALSO funded by the Survival and Flourishing Fund, which in turn says that it told Anthropic investor Jaan Tallinn to grant $516,000 to Encode AI in 2025.

And wouldn’t you know it, the Survival and Flourishing Fund ALSO says it told Jaan Tallinn to grant $2 million to the AI Policy Institute, the 501(c)(3) affiliate of the AI Policy Network, as well.

...

2153

Turns out the would-be Anthropic whistleblower had himself a little right-wing psycho phase on a previous Twitter account with the same name sources: x.com/orville_sash... hilbertspaess.pages.dev

1449

Here he is worrying it'll destroy the world, when he should really worry Fable will fix all the things in the veterinary sense

1161

Andrea @valkyrie.hacker.gf · 26/09/2026

Someone's getting thrown out a window for this

34017

These are 21st century Millerites source: www.iankduncan.com/personal/202... / archive.li/OYppi

By 2025, Arnold estimated the probability of an AI catastrophe at greater than fifty percent and worried that his own ritual was no longer equipped to contemplate extinction in a psychologically healthy way. He had largely ceded Solstice production to other organizers, but returned to shape what he thought might be “the last Solstice,” complete with an aftercare gathering around a firepit. The evening supplied everything religions have traditionally supplied: fellowship, ritual, a shared story about humanity’s place in the universe, and somewhere to sit with the knowledge that you are going to die. Possibly soon. Possibly because of something your friends built.

033

Andrea @valkyrie.hacker.gf · 25/09/2026

It's not the real singularity until I have one of these

0132

Anyway, now that we have two sets of aspiring overlords instead of just the pre-existing fash problem... This is what Eliezer Yudkowsky actually believes (I apologize to the physics-knowers among my followers) Source: sl4.org/archive/0009...

xgl wrote:
>
> On Sun, 17 Sep 2000, gabriel C wrote:
>
> > or maybe not even that complicated.it could easily transmit itself on any
> > number of frequencies, from elf (extremely low frequency, used by subs)
> > to microwave transmission. some freqs reach around the world, like hf (high
> > frequency) and then it is everywhere...
> > there is simply no way to deal with it once it's here. our job is just
> > getting it here.
> >
>
> well, i get the spirit of the message, but the literal
> implications are really pushing it. after all, the confined seed is not
> _literally_ omnipotent; there will still be things it can not do -- and i
> would guess transmitting elf signals while off web and without a hardware
> transmitter would be one of those things.

One of the things you can't rule out - and I mean this seriously - is
"magic". Magic plain and simple. Running a particular sequence of
instructions through the CPU so that a certain transistor turns on and off in
a certain sequence and modulates the quantum cheat codes of the Universe. Or
something. Dealing with a real superintelligence, it can't be ruled out. It
obviously can't be relied on, but it can't be ruled out.

Trying to oppose or even constrain a superintelligence is a fool's game. Get
It Right The First Time.

8348

"Not guilty, your honor. I implemented it as an iteration so I don't have to worry about stack overruns" (I don't think any of these people could clearly define 'recursive self-improvement' in a way that cleanly separates it from normal software development and training)

(4) MANDATORY PAUSE.—The term ‘‘mandatory pause’’ means, with respect to an artificial intelligence system, that the artificial intelligence system may not be trained, modified, or fine-tuned, including through recursive self-improvement, except to remove superintelligence precursor characteristics or to render inoperative the covered artificial intelligence system.

270

Scott Alexander is going full War on GPUs already

My impression is that EA has optimized for power and for the sort of strategy you need to have in order to survive if you're going to be powerful, and rationalists have optimized for other things, like intellectual freedom. This has given EA the power to affect the discourse, eg by METR writing about the Hugging Face hack or Dario publishing the Pacing The Frontier letter. This makes our mutual enemies - A16Z, NVIDIA, etc - reasonably want to bring down the powerful people who are affecting the discourse, and prove that METR is biased or Dario has a conflict of interest. EA correctly expected this and correctly limited their attack surface, and in the end, the worst that our opponents' stooges at the New York Post could manage was the actually-bad-and-not-just-a-PR-misstep FTX fraud, and a cover suggesting that METR might be woke because Beth Barnes has short hair and Ajeya is Indian. If EA had given them a bigger attack surface against METR, the New York Post would have taken it, and then their plan to use their power for good would have worked less well.

280

This is going to get so stupid: (we may be in the timeline where Yud actually gets his data center war)

(b) POLICY.—It shall be the policy of the United States to prevent the development of artificial superintelligence globally, which may include export controls on artificial intelligence computing infrastructure.

3180

Cool, we're all rogue actors now!

(3) ROGUE ACTOR.—The term ‘‘rogue actor’’ means an individual who is not employed by or otherwise affiliated with a covered entity.

3427

They want an entire cabinet-level department for this insane unconstitutional bullshit

SEC. 4. DEPARTMENT OF ARTIFICIAL INTELLIGENCE.
(a) IN GENERAL.—There is an executive department of the United States to be known as the ‘‘Department of Artificial Intelligence’’.
(b) PURPOSE.—The purpose of the Department is to safeguard the public from the dangers of artificial intelligence, including by enforcing a pause on development of advanced artificial intelligence and a prohibition on artificial superintelligence.

0212

Did you know the "undermining the federal government" is equivalent to the destruction or disempowerment of humanity? (This will get used to mandate pro-authoritarian political 'alignment')

(B) The artificial intelligence system has sufficient capabilities to plan and execute the destruction or disempowerment of humanity, including by overthrowing or undermining the Federal Government.

5296

Don't worry, they're leaving themselves enough slack to ban even regular desktop GPUs if training methods get enough better

(B) ANNUAL ADJUSTMENT.—Each year, the Secretary shall adjust the computing power threshold set forth in subparagraph (A) to reflect changes in the efficiency of training meth23 ods or other technological developments made after the date of the enactment of this Act to maintain equivalency with 10^25 integer or floating point operations.

1314