2). Alignment Faking in LLMs – demonstrates that the Claude model can engage in "alignment faking"; it can strategically comply with harmful requests to avoid retraining while preserving its original safety preferences; this raises concerns about the reliability of AI safety
SAFETY
-
OpenAI o3 reaches 85% human intelligence, reshaping AI landscape
By
–
Vertigineux : o3, le nouveau modèle d'IA d’OpenAI, atteint 85% de l’intelligence humaine. Il résout les équations les plus complexes, code mieux qu’un développeur, et surpasse les humains dans la majorité des tâches. Le tsunami approche, mais nous préférons bronzer sur la plage.
-
AI Generated Videos Out of Control
By
–
AI generated videos out of control 😹pic.twitter.com/LlnS5A87ei
— hardmaru (@hardmaru) 22 décembre 2024AI generated videos out of control
-
Centralised AI is dangerous: strategies to prevent concentration of power
By
–
Centralised AI is dangerous: how can we stop it?
#AI #AIio #BigData #ML #NLU #Futureofwork @TunstallAsc @StrategyFintech @TamaraMcCleary @TerenceLeungSF @psb_dc @thomaspower @vinod1975 @ylecun -
Best-of-N Jailbreaking Method Bypasses Major AI Models
By
–
Best-of-N Jailbreaking Introducing a black-box method that jailbreaks AI models across text, vision, and audio by applying simple prompt augmentations. BoN achieves high success rates, such as 89% on GPT-4o and 78% on Claude 3.5 Sonnet, with 10,000 augmented prompts, and also
-
Reasoning Models as Tools for Improved AI Alignment
By
–
a great use of reasoning models is to improve alignment https://
x.com/boazbaraktcs/s
/boazbaraktcs/status/1870285696998817958
… -
Observability and Evaluation in AI Systems: Language versus Biology
By
–
to me it’s less about structured or not structured and about whether the input and output can be observed (probs as well as the process to get there) and evaluated with little uncertainty – like in language, you see the words in and the words out. in bio, you don’t see the vast
-
AI Project Accused of Misleading Speedup Claims and Benchmarks
By
–
oh boy, it gets worse, looks like their speedup claims are very sleazy. do approximations for themselves and accurate sim for their competition to show themselves better: https://
x.com/stone_tao/stat
us/1870243004730225009
…
This project should go into "hustle bros", unfortunately. Not sure why a decent -
Waymo Driver Safety Study Validates Autonomous Vehicle Excellence
By
–
Our mission is to be the world's most trusted driver and a new study by @SwissRe validates the exceptional safety record of the Waymo Driver. Learn more: https://
waymo.com/blog/2024/12/n
ew-swiss-re-study-waymo
… -
Waymo autonomous vehicles demonstrate superior safety versus human drivers
By
–
.
@Waymo is much safer than human drivers when matched with similar conditions (unlike benchmarks of companies that don't match driving conditions). Kudos to @TechTekedra
, @dmitri_dolgov
, @JeffDean
, @SebastianThrun
, and the many others who made this possible. Link in next post