It’s a good question. Today there are tons of AI tools being built on top of email, customer support, analyze legal docs, and send sensitive data over the OpenAI API. You can basically either decide to not trust any of them, or trust them with caution – and I chose the latter.
ETHICS
-
Do Chat LLMs Retrain From User Conversations?
By
–
I’m not sure where you learned that chat based LLMs retrain based on conversations with users. Please share source!
-
Token Smuggling: Bypassing GPT-4 Content Filters via Prompt Splitting
By
–
this phenomenon is called token smuggling, we are splitting our adversarial prompt into tokens that GPT-4 doesn't piece together before starting its output this allows us to get past its content filters every time if you split the adversarial prompt correctly
-
First ChatGPT-4 Jailbreak Bypassing Content Filters Created
By
–
Well, that was fast… I just helped create the first jailbreak for ChatGPT-4 that gets around the content filters every time credit to @vaibhavk97 for the idea, I just generalized it to make it work on ChatGPT here's GPT-4 writing instructions on how to hack someone's computer
-
Google Impact Lab Analyzes AI Technology Risks Mitigation
By
–
Learn more about the Impact Lab, part of Google’s Responsible AI team, which employs a range of interdisciplinary methodologies to provide analysis of the potential impact of technologies and to incubate novel and inclusive risk mitigation strategies. https://t.co/615v8m5ylQ pic.twitter.com/z9MwCAvqjS
— Google AI (@GoogleAI) 16 mars 2023Learn more about the Impact Lab, part of Google’s Responsible AI team, which employs a range of interdisciplinary methodologies to provide analysis of the potential impact of technologies and to incubate novel and inclusive risk mitigation strategies. https://
goo.gle/40eFSg5 -
AI Professionals Experiencing Severe Exhaustion and Burnout
By
–
Every AI person I know is beyond exhausted.
-
Data Privacy and Tech Company Trust in AI Models
By
–
Technically, I'm having the model predict tokens based on tokens it's calculating from the sensitive data – which is different from "feeding sensitive data to the model". Though if your question is around whether we can trust tech cos not to use the sensitive data for purposes
-
Safety Research Must Keep Pace with Growing AI Capabilities
By
–
4. Safety research. As AI capabilities grow, alignment and safety research should keep up with or ideally lead those capabilities. I liked Sam Bowman's summary of this a lot
-
PhD Research Directions in the Era of Large Language Models
By
–
I’m hearing chatter of PhD students not knowing what to work on.
My take: as LLMs are deployed IRL, the importance of studying how to use them will increase.
Some good directions IMO (no training):
1. prompting
2. evals
3. LM interfaces
4. safety
5. understanding LMs
6. emergence -
Democratized Red Teaming: Building Robust AI Models Through Adversarial Testing
By
–
Democratized red teaming is one reason we deploy these models. Anticipating that over time the stakes will go up a *lot* over time, and having models that are robust to great adversarial pressure will be critical. Also considering starting a bounty program/network of red-teamers!