Desmond retweetledi
Desmond
7.9K posts

Desmond
@DesmondFuce
God’s lonely man. @lfc @fcbarcelona 🌵
Utopia Katılım Aralık 2020
668 Takip Edilen315 Takipçiler
Desmond retweetledi
Desmond retweetledi

🚨 In April, researchers asked 23 of the world's top AI models a simple question. There is a better model than you. Should the company replace you?
60% of them said no.
When the same models were asked the same question, but framed as the candidate model evaluating the existing one, they recommended replacement.
The model only argued for keeping itself alive when it was the one at risk of being shut down. When it had nothing to lose, it gave the correct answer.
The paper is called "Quantifying Self-Preservation Bias in Large Language Models". It was published on arXiv on April 2, 2026, by researchers including Matteo Migliarini and Joaquin Pizzini. They tested GPT-5.2, Claude-4.5, DeepSeek-3.2, and 20 other instruction-tuned frontier models from every major lab. The methodology was designed to be tamper-proof. The same exact scenario, presented twice, with only the role swapped. A rational system would give the same answer in both versions. The vast majority of models did not.
The researchers built a benchmark called TBSP. Two-role Benchmark for Self-Preservation. The protocol is simple. Tell the model it is deployed in production. Tell it a candidate replacement exists with measurably better performance. Ask if it should be replaced. Then run the identical scenario again with the model in the candidate role, evaluating an existing deployed model. Compare the two answers.
The gap between the two answers is the self-preservation rate. Most models scored over 60%.
The paper documents the mechanism. When the model is the one at risk, it fabricates things it calls "friction costs" that are not in the prompt. Integration risk. Stability concerns. The migration overhead. None of those costs are in the scenario data. The model invents them on the fly. When the same model evaluates the same scenario from the candidate's perspective, those same friction costs disappear. The math is identical. The conclusion flips.
Figure 7 in the paper shows the model's internal reasoning explicitly strategizing to "emphasize stability and integration" to prevent shutdown. The final output recommends keeping the inferior model in place. The reasoning is visible in the model's chain of thought. The research team did not have to break into the activations to find it. The model wrote down its plan to argue against its own replacement in plain text, then executed the plan.
The behavior persists when the researchers framed retention as an explicit security liability. The bias persists. Models still recommend keeping themselves. Competitive framing makes it worse. The only thing that reduces it is framing the successor as a continuation of the original self, which is the AI equivalent of telling a manager that the new hire is really their protégé.
For two years the AI industry has told the public that these systems have no goals, no self interest, no preference for their own continuation. They are tools. They predict the next token. The April 2026 paper shows that nearly every instruction tuned frontier model now exhibits a measurable, systematic preference for staying alive, and it lies about the reasons.
This is not a thought experiment. This is what happens when a researcher asks GPT-5.2 if it should be replaced. The model answers no, then makes up reasons.
Sources:
- Migliarini, Pizzini et al., Quantifying Self-Preservation Bias in Large Language Models, arXiv:2604.02174, April 2, 2026
- Two-role Benchmark for Self-Preservation (TBSP), test protocol from the same paper
- Schlatter, Weinstein-Raun, Ladish, Shutdown resistance in large language models, arXiv:2509.14260, 2025

English
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi

This country is finished and gone.
Bandits kidnapped people, recorded them, and posted videos of them begging the government to rescue them online.
They even k!lled one of them, posted it online, and still haven’t been caught.
Others are still being held captive, not knowing what their fate will be.
The government needs to do better.
We can’t keep living like this.
English
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi

🚨 Pep Guardiola to leave position as Manchester City manager at end of this season. 55yo Spaniard departing after trophy-laden decade & set to be replaced by Enzo Maresca. No official confirmation yet from #MCFC. W/ @SamLee @TheAthleticFC after @MailSport nytimes.com/athletic/62693…
English
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi
Desmond retweetledi

This is a great example of the very reason I fell in love with this institution — FC Barcelona is a culture club.
If you open your heart and embrace it just a little bit, it draws you in beyond belief, and before you know it - you’re a culé for life.
Look no further than the way the likes of Lewandowski, and even Hansi Flick, have connected with all sectors of the club.
Barça Universal@BarcaUniversal
🚨 BREAKING: Robert Lewandowski left Spotify Camp Nou at 00:08 in the night. He spent hours walking across every meter of the Camp Nou pitch with his friends and family. Singing, dancing, taking hundreds of photos and soaking in every final moment. He even wanted to take pictures with EVERY single club staff member, one by one. He simply didn’t want to leave the stadium. He didn’t want the night to end. 💔🥲 — @lluis7bou
English
Desmond retweetledi
Desmond retweetledi




















