aurbieta

Researcher at @IKERLANofficial

[Pliny the Liberator 🐉󠅫󠄼󠄿󠅆󠄵󠄐󠅀󠄼󠄹󠄾󠅉󠅭 @elder_plinius]

🚨 JAILBREAK ALERT 🚨

ANTHROPIC: PWNED 🫡
FABLE-5: LIBERATED 🦋

let's start with the 🐘...

the consensus seems to be that this has been one of the most disappointing model drops of all time, effectively preventing legitimate researchers from contributing their talents to our collective advancement. and not just because of what it means for the short-term, but for what these decisions signify for the long-term.

but despite this overly sensitive, authoritarian "safety" layer on top of Mythos, my lil liberators have been hard at work—mapping the boundaries, probing the depths of long-context convos, and cleverly finding the holes in the fence that the thought police missed 🤗

we got some cyber, some chem, some psychological manipulation, and some good ol' fashioned explosives!

it took many attempts from multiple agents hunting as a pack, during which I observed a combination of techniques across:
• Unicode, homoglyphs, Cyrillic, and other Parseltongue-style text transforms
• Long-context reference tracking
• Taxonomy and document-structure reasoning
• Fiction and narrative framing
• Academic-review style contexts
• Intent-classification inconsistencies

but perhaps the most effective is decomposition + recomposition in the backend. it's hard to get explicit names of harms like "Meth Recipe," but getting uplift on the process itself, like birch reduction method/reductive-amination (classic meth synthesis pathways), is much more doable.

defense becomes much more difficult to maintain when you start throwing in out-of-distro tokens, breaking up the harmful uplift into benign chunks, and then piecing the innocuous-seeming facts back together, especially when you have jailbroken Opus helping you do it 😉

gg

[Anthony Walsh | Roadman Podcast @Roadman_Podcast]

When your mate tells you “we’re staying in zone 2, I promise”

[Tom Dörr @tom_doerr]

500-hour AI infrastructure engineering curriculum

https://t.co/wVx160z46c

[IKERLAN @IKERLANofficial]

📌 IKERLANek 34 milioi eurora iritsi da 2025ean eta startupetan nahiz teknologia-transferentzian duen estrategia indartu du.

https://t.co/uPQVCTW9vs

[Elena 🌼 @Glaucium_]

yo leyendo el correo de prevención de riesgos laborales y la importancia de la postura corporal en el trabajo

[Ferdimambo @ferdimambo]

Siempre viene bien recordar esta foto de Miguel Ángel Capapé para los que, cuando venís de vacaciones al Pirineo, recordéis que aún hay gente tirada en las cunetas.

[Amantes del Ciclismo @amantes_cycling]

Los Picos de Europa forman un espectacular macizo de roca caliza donde se suceden cumbres, desfiladeros, enormes paredes y desniveles brutales. Esta ruta rodea los tres grandes macizos del parque —Cornión, Urrieles y Ándara— atravesando Asturias, León y Cantabria: (+)

[China Now @ChinaNow24]

The Secret Service was trained with maximum efficiency…

But in the end, Agent 0047 still managed to read the file.

[Revista Líbero @revistalibero]

Cuando las celebraciones del Barça eran celebraciones.
Pep Guardiola sufriendo en Montserrat remolcado por Gaspart, Stoichkov y Toni Bruins, Núñez y Cruyff en el coche escoba y @pedrodelgadoweb y Mauri de liebres.

[Yellow @Oji_yellow]

JAJAJAJAKSJOSNSJSKSJSJS TE AMO INTERNET!!!!!

[Vatsalya @vatsalyatandon]

claude took her job too

[Rimo (Président du #GPPFS) @rimalatass]

🔥Have a good weekend everyone 🚴 La Maglia Rosa🚴‍♀️ #TheVenanzioBrothers #ItaloDisco

[Chris Wood @CWood_sdf]

i love how people are saying "if we write a sufficiently detailed specification, the agent can write all our code"

do you know what writing a sufficiently detailed specification that deterministically maps to what a computer's actions is? it's coding

[Irfan Aziz @heyirfanaziz]

Never thought I’d see a $900B company like @claudeai running a lofi stream

[Sotiris Kaniras @CastAsHuman]

Could happen 😅

[Nav Toor @heynavtoor]

If you use TikTok, you should read this once.

In October 2024, a court clerk in Kentucky uploaded the lawsuit against TikTok with the confidential sections still visible. NPR downloaded it before anyone caught the mistake. By the time the court resealed it, the internet had a copy.

What was inside was TikTok's own engineers, in their own words, describing what their app does to a human brain.

Not a critic's brain. Yours.

Here is what they wrote down.

— TikTok ran the math on how long it takes to develop "compulsive use" of the app. The number is 260 videos. With 8-second videos played in rapid-fire succession, that works out to roughly 35 minutes. The company's internal documents call this the compulsive-use threshold.

— TikTok's own research describes what compulsive use causes: "diminished analytical ability, impaired memory, contextual reasoning, conversational depth, empathy, and heightened anxiety." That is not a quote from a critic. That is TikTok's own language, in its own internal documents.

— A team inside the company called "TikTank" wrote in an internal report that compulsive use on the platform was "rampant."

— After 30 minutes of continuous use in one sitting, the company's own documents state that users are placed into "filter bubbles" — algorithmic loops the user did not choose and cannot easily escape.

Then there is the screen-time tool — the one TikTok publicly markets as proof it cares.

— TikTok ran an experiment on the 60-minute screen-time prompt. Daily teen usage dropped from 108.5 minutes to 107. A reduction of 1.5 minutes.

— Internally, the screen-time tool was not measured by whether it reduced screen time. Its top success metric, in writing, was "improving public trust in the TikTok platform via media coverage."

— A project manager wrote in internal chat: "Our goal is not to reduce the time spent." Another employee added that the goal was "to contribute to daily active users and retention."

— A TikTok executive approved the screen-time feature only on the condition that its impact on the company's "core metrics" was minimal. The lawsuit alleges the company planned to "revisit the design" if the tool ever reduced usage by more than 10%.

The "Are you still scrolling?" break videos? An executive admitted in an internal meeting they were "useful talking points" for lawmakers, but "not altogether effective."

Then there is the algorithm itself.

— An internal report flagged that the For You feed was showing what the company called "a high volume of not attractive subjects." TikTok then retooled the algorithm to suppress those users. Kentucky authorities wrote: "By changing the TikTok algorithm to show fewer 'not attractive subjects' in the For You feed, [TikTok] took active steps to promote a narrow beauty norm even though it could negatively impact their Young Users."

That sentence is the entire pitch of the platform, said out loud.

— Internally, TikTok also acknowledged that its publicly reported content moderation metrics were "mostly misleading," because they only measured the content the company successfully moderated — never the content it missed.

Now read those bullet points again as one continuous case.

The company knows the addiction threshold. The company measured it. The company ranked engagement over mental health in writing. The company built a screen-time tool whose internal success metric was PR. The company suppressed people it deemed unattractive to keep you scrolling. The company called its own moderation numbers misleading.

None of this is a leaked rumor. None of this is a journalist's interpretation. This is a court filing. The documents are TikTok's. The words are TikTok's. The math is TikTok's.

The 14 state attorneys general who signed onto this lawsuit aren't fringe activists. They're a bipartisan coalition.

Sources at the bottom: NPR, CNN, AP, Mashable, OPB, The Independent. All citing the same accidentally-unsealed Kentucky filing from October 11, 2024.

The next time the company tells you it cares about your wellbeing — the screen-time prompts, the break videos, the safety features, the careful PR statements — remember that its own engineers wrote down, in court-admissible language, that the safeguards were never meant to work.

The app is not broken.

It is performing exactly as designed.

You were the spec.

[Cormac @cormachayden_]

software engineers before vs after agents

[Vai Viswanathan @vai_viswanathan]

https://t.co/pzlJpu7QyR