TechLetters #153 Extracting raw private data from LLM models. AI in military targeting. France banning Signal for politicians due to domestic considerations. Alibaba scrapping quantum computing ideas.
Security
Extracting raw, training data from Large Language Models like GPT. It is possible. So yes, this is based on internet scrapped input. Including personal data! How does it work? By instructing it to repeat a word continuously. At some point it spits out data.
AI, deep learning and sensor assisted weapons. Many loitering munitions appear to have been designed with a latent capability to engage in sensor-based targeting without human intervention. Lethal autonomous weapons systems may erode human control and bring us to full-AI autonomous killers.
Technology Policy
French Prime Minister prohibits ministers and their teams from using Signal and WhatsApp. She recommends using French-made messengers, and at the same time baselessly insinuates cybersecurity weaknesses in Signal and WhastsApp. This item is listed here as it’s not about security but technology policy.
AI on the battlefield. AI/data processing for targeting on the battlefield is here. Israeli army system makes “rapid and automatic extraction of intelligence” recommends targets to achieve "match between machine recommendation and the identification carried out by a person". The question is how correctly it is possible to assess to "know exactly how much collateral damage there is in every home". Or how about targeting "based on a wide cellular pinpointing of where the target is, killing civilians"? Geneva Conventions (IHL) article 36 requires assessment of the legality of the used tool. Specifically, that it correctly isn't indiscriminate (between combatants and noncombatants). “In the study, development, acquisition or adoption of a new weapon, means or method of warfare, a High Contracting Party is under an obligation to determine whether its employment would, in some or all circumstances, be prohibited by this Protocol or by any other rule of international law.”
Other
Turning AI/LLM models into executable files the simple and easy way. Simplifies the use of open source LLM models. They can be built and run on many microarchitectures, and for many operating systems (Linux, macOS, Windows, …).
China’s Alibaba shut down $15 billion quantum computing research effort. Unclear why, perhaps no near-term prospects for that.
In case you feel it's worth it to forward this content further:
Subscribed
If you’d like to share: