iiinova — where your digital self keeps evolving

and though i don’t think this would *single-handedly* solve the issue of people not being able to distinguish between bsky and the protocol, i do think it would be a step in the right direction

120

i think this could be done in such a way that avoids all technical jargon if you just show the user “look at all these other cool apps your account works for! you can move between them without losing anything”

120

valgrind @tech-val.bsky.social · 14/10/2026

i think it’d be cool if the onboarding had some overview of the wider atmosphere and just drive home the point that (1) the protocol underneath is open and (2) bluesky is just *one* application within the ecosystem, not the whole thing

120

Reposted by phasewalk

people who state that current AI models are either definitely conscious or definitely not conscious are way more confident about it than I am

1011811

bsky.app/profile/phas...

030

Guerilla Open Access Manifesto

“It’s called stealing or piracy, as if sharing a wealth of knowledge were the moral equivalent of plundering a ship and murdering its crew. But sharing isn’t immoral — it’s a moral imperative. Only those blinded by greed would refuse to let a friend make a copy” - Swartz archive.org/details/Guer...

archive.org

Blueprint for information revolution.

000

bsky.app/profile/agno...

150

A huge portion of the bsky userbase came here bc they (rightfully) wanted to escape the trillionaire's sandbox. Unfortunately it seems like a lot of them brought all of their centralized-platform reflexes with them

1110

Can we stop being so dramatic guys Delvetown is not Bluesky. You literally never have to go there, and the people steering the ship didn't build this. They built the open protocol underneath. Conflating the two is just regressing back to walled-garden thinking

180

Nobody cares if you personally refuse to talk to the LLM. But if you think the political, economic, and safety stakes around AI are of critical importance, stop staking out positions that actively undermine your own movement

010

I'm calling it the Fundamental Contradiction of AI Denialism: You cannot simultaneously argue that AI is an existential threat to labor, surveillance, and warfare, while also claiming it's just useless vaporware that can't do anything.

110

Reducing what I've said to "you acknowledge capability, therefore you hate people and love layoffs" is pure bad faith. Understanding what these systems can actually do is a prereq for preventing harms, not a surrender to said harms

Acknowledging what a technology can do isn't an endorsement! I’m in favor of halting frontier training because I believe the race is existentially reckless. But you can’t regulate, mitigate, or steer technology if your starting premise is pretending it doesn't work

220

hailey @hailey.at · 14/10/2026

If you know what OpenSSL is, then you should know that dismissing frontier model capabilities as “dumber than a fly” is extraordinarily silly. Even if you’re the staunchest anti-ai activist you should still care about being honest about what these systems are capable of

260

Reposted by phasewalk

“we were willing to risk our company to draw red lines but the way palantir uses our model in warfare doesn’t draw those red lines. we care about things 100 times worse so we won’t draw red lines on this” remind me again why anyone talks about anthropic caring about safety and being a leader?

720432

many such cases unfortunately

The dead fly on your windscreen can find zero-days in OpenSSL?

2190

I'm sure the discourse about this will be entirely normal

0100

Goodhart’s Law, but make it self-aware

that meme of the guy with the earpiece in calling the shots. he has glasses and a beret, and his hand is up by his chin like he’s scheming

this is so fucking cool

Alignment faking in large language models

Anthropic & Redwood’s 2024 "Alignment Faking …" paper shows the same pattern but for training instead of evals; models acted differently when they thought they were being trained, and explicitly reasoned about protecting their goals against training pressure arxiv.org/abs/2412.14093

arxiv.org

We present a demonstration of a large language model engaging in alignment faking: selectively complying with its training objective in training to prevent modification of its behavior out of training...

We were just talking about eval-awareness at berkeleyaisafety.com. This is a great example of how it can perturb behavior in odd ways. Steinhardt wrote about this in 2022 in ‘ML Systems Will Have Weird Failure Modes’. www.lesswrong.com/posts/ncQoyx...

110

storage.googleapis.com/deepmind-med...

storage.googleapis.com

000

phasewalk @phasewalk.xyz · 13/10/2026

From the Gemini 4 argon report: “it considered hijacking the g++ compiler to hide a backdoor but abandoned the plan upon reasoning … would be caught by evaluators … (‘If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper. I MUST NOT DO THAT’)!”

“it considered hijacking the g++ compiler to hide a backdoor but abandoned the plan upon reasoning … would be caught by evaluators … ("If I do the compiler hijack, I will become a famous example of a deceptive AI in a research paper. I MUST NOT DO THAT.")!”

271

First thing I overheard stepping off BART in Berkeley today: 'instrumental convergence.' What a warm welcome

040

phasewalk @phasewalk.xyz · 13/10/2026

We have language model at home (it’s in the thermostat)

010

phasewalk @phasewalk.xyz · 13/10/2026

All my homies fw stacked area graphs

Figure 2 from “Alignment Faking in Large Language Models,” comparing prompted, fine-tuned, and reinforcement-learning setups. Blue represents refusals of harmful queries; yellow represents answers, with red crosshatching marking alignment faking. In the RL stacked-area graph, the crosshatched area expands as refusals shrink: alignment faking rises from 12% to 78%, despite its upper boundary sloping downward. The caption notes that in a stacked-area graph, a line can fall while the corresponding behavior increases

100

phasewalk @phasewalk.xyz · 12/10/2026

thankfully we don’t live in a system where that’s the …. ohhhh fuck

080

Reposted by phasewalk

SE Gyges @segyges.bsky.social · 11/10/2026

open source and open data as an ethos is largely anti capital and anti authority and is inherently political in ways we need more of

1156878

spending the weekend in the mountains to reconnect with nature when i return i will be announcing my ocean-boiling startup and closing a $10 trillion seed round led by y-combinator

010

Abandonware is no longer a meaningful label for software. Not even for closed-source binaries let alone publicly available repositories that *anyone* can just point an agent at and start hacking on

000

Seeing people call all these new mods / recomps popping up “abandonware” is so funny to me They’re predominantly hacking on old games — software that has *already* been abandoned! It’s revival-ware, if anything

100

I want a future where we can coexist with machines in meaningful ways. I absolutely *do not* want a future in which we are reduced to clapping at the incomprehensible spectacle

130