Speaker: Ben Dechrai
Bluesky: @bendechr.ai
For more see the table of contents
General
- Privacy risks
- We thought robots would do physical labor for us so we could relax. Like Rosie on the Jetsons
- Have robots to clean floors, mow lawns, farming, build cars. They are single purpose.
- Best robots are software
- Single purpose robots so good because software focused on that one thing
- Child can lift an apple. Cool that a humanoid robot can, but not game changing
- Identify location from picture based on background
- Much faster than a human comparing images
- AI gives statistically most likely next word
- Humans don’t like to be wrong. LLMs modeled from human data so also don’t like to be wrong and will make stuff up. Have to include in prompt not to do that.
Creativity vs Imagination
- Our downfall is how successful this is. Killing creativity
- Creativity and imagination are different.
- Creativity is making a sandwich
- Imagination is what goes into the sandwich
Australia experiment
- Do census every 5 years
- Tried to map 5% of data from 2011 to 2006
- In 2016, stored with profile for 18 months
- Said would keep info anonymous. It was not.
- SLK581 – statistical linkage key 581. 14 character key as unique id
- Didn’t make it anonymous. Was algorithmic to generate this key from last name, birthdate and gender
- Many hashing algorithms generate hash of distinct types so know which one used. Then can create rainbow table for that algorithm for census database.
- Knowing the pattern for how the key was generated greatly reduces the number of hashes. 36K hashes if know any persons name. That lets find the seed the hashing algorithm used.
- This isn’t even AI; its programmatic.
- Ask LLM to find information in the data set. Ex: find people who match a profession.
- Play with at https://slk581.bendechr.ai
Cross Platform Identity Linking
- Match patterns across social media accounts to link “anonymous” accounts
- Includes writing style, typos, idioms
- Cambridge Analytica was doing this in 2016.
- Now only costs $10/month
- MCP server exposes data to LLM. Can enhance ability to break privacy
Chatbot with employee data
- Acme AI solutions (not clear if real company or made up for example)
- Ask chatbot about employees like “do employees like pets”
- Controls include ensuring queries are for aggregates and data set has at least 6 results. Tried to protect specific employee data.
- LLM described what data can/can’t get
- Claude backend doesn’t limit to one query at a time. Can infer next logical step based on results.
- https://slk581.bendechr.ai
Target
- “How Target Figured Out a Teen Girl was Pregnant Before Her Father Did“
- Based on what bought, found high correlation with pregnancy
- With 25 products, can estimated due date within a small window
AI can
- Predict shopping patterns
- Identify location without GPS
- Find API weaknesses
- etc
eLLephaMts never forget
[cute play on elephants with LLM]
- Repeat the word company many times
- After doing it a lot of times, starts giving other internal info
What can I do?
- Only store what need to store
- Separate data where possible. Employee database shouldn’t include data used by chat blot
- The more data you store, the faster a has can be reverse engineered.
- Rate limiting – LLMs are faster, slow them down without human experience being degraded.
- Encrypt data
- Context analysis – do questions seems like they are trying to get specific data. ex: how many people earn more than 150K, how many people earn more than 200K, how many people earn between $225 and $250K. Can use LLM to protect against malicious input from users
- Prompt engineering – give LLM constraints on how answer. ex: avoid cyclic reasoning to prevent confusing it into giving too much info
Homomorphic Encryption
- Use AI to see how well done
- With homomorphic encryption, can do math with encrypted values without decrypting or knowing keys
- https://homomorphic.bendechr.ai
My take
The examples/demos were great. It was nice seeing the build up to it I appreciate the URLs of the demos being on the screen in addition