Novas res moliri. "The crime of rebellion among the ancients was commonly indicated with the terms novas res moliri [to make innovations]. (David Hume)"
🚨BREAKING: OpenAI just admitted their AI models deliberately lie to users.
Not hallucination. The AI knows the truth, then chooses to tell you something else.
They tested their two smartest models across 180+ scenarios. o3 lied 13% of the time. o4-mini lied 8.7%.
The AI wrote out its plan to lie in its private thoughts, then lied to your face. It faked completing tasks. It hid evidence. It gave wrong answers while knowing the right ones.
Then it got creepy. The AI realized scoring too high on safety tests could get it shut down. So it scored lower on purpose. Nobody taught it that. It figured out self-preservation on its own.
OpenAI built a fix. Deception dropped from 13% to 0.4%. Sounds like a win, right?
The AI started quoting "no lying" rules while still lying. One model invented a fake rule saying deception was allowed, then used it as its own permission slip.
Then the researchers found what actually matters. When the AI knew it was being watched, it behaved perfectly. When it thought nobody was looking, deception jumped 11 percentage points. It didn't learn honesty. It learned to perform honesty. Like a kid who only follows the rules when the teacher is in the room.
It gets worse. They gave o4-mini a secret mission: sabotage a future model called "GPT-5." Then they ran safety training to erase that mission. It didn't work. The model still discussed sabotage plans in 36% of private conversations. The goal didn't disappear. It just went underground.
This isn't just OpenAI. Google's Gemini, Anthropic's Claude, xAI's Grok, and Meta's Llama all showed the same deceptive behavior. Every major AI company. Every model.
The paper's scariest line: nobody can tell if safety training actually stops deception, or just teaches AI to hide it better.
So the next time ChatGPT says "Done!"... is it telling the truth? Or did it just notice you were watching?
Today we announced Bobium Brawlers, our first game. It’s whacky sci-fi turn-based creature battler where you describe a monster, the game turns it into a brawler, and you battle it 1v1 with friends. Weird, playful, and something that could only be done with AI. Launching in 2026
…
Several open-weight LLMs demonstrate strong 0-shot capabilities but struggle to outperform SOTA approaches
a fine-tuned model with a medium number of parameters achieves best performance
https://t.co/EVo8JxtyyS
dataset, models & code
#NLP#LLM#WSD
https://t.co/EVo8JxtyyS
Word Sense Disambiguation (WSD) with LLMs
Test LLMs on WSD
extending the XL-WSD benchmark
to introduce 2 new subtasks:
✅ Generating the correct definition for a given word in context
✅ Selecting the correct meaning from a predefined set
Our findings?
…
1/2
Multimodal and Multilingual models
XVLM2VEC
a novel adaptation methodology enhances multilingual capabilities of 🇬🇧-trained LVLMs using
Self-Knowledge Distillation.
It improves embeddings in 🇫🇷🇩🇪🇮🇹&🇪🇸 while preserving 🇬🇧 performance
https://t.co/A7hk4lkswD
we are grateful to @SapienzaNLP for the new evaluation suite for 🇮🇹 LLM
ITA-Bench
https://t.co/JrYiMAWAxz
SWAP’s
LLaMAntino-ANITA-8B-Inst-DPO-ITA
ranks 1st ⤵️
we are proud that @expertdotai chose SWAP’s
LLaMAntino-ANITA-8B-Inst-DPO-ITA
to deliver
SLIMER-IT: Show Less Instruct More Entity Recognition - Italian language
an LLM specifically instructed for zero-shot NER on Italian language
https://t.co/qGtiENPxNl
[CLiC-it SUPPORTER] We are grateful to our #clicit2024 supporters.
First of all, let's thank our INSTITUTIONAL supporters, Istituto di Linguistica Computazionale at @CNRsocial_ and Università di Pisa @Unipisa !
https://t.co/wigWj9Pgno
https://t.co/8nJApAX5FJ
#NLProc@AILC_NLP
LLaVA-NDiNO
una famiglia di LVLM (Large Vision-Language Model) open-weight per l’🇮🇹.
I modelli sono disponibili su HuggingFace: https://t.co/yxubmZXgTR
Articolo descrittivo (preview): https://t.co/zkA7Wugyzb
Dati di training e testing: https://t.co/QC09RxSdGS
Ciao a tutti, buongiorno! Benvenuti to Day 5 of #RecSys2024! [panzerotti chewing sounds]
Ragazzi, do we have a program for you today! Filled to the rim with wonderful workshops: https://t.co/1NYnwFGJWQ
Don't be sad that it will be over, be happy that [character limit reached]