Content
@
0 reply
0 recast
0 reaction
Vishal Sachdev
@vishalsachdev
From #simulatedHumans --> #dataSovereignty --> #monetizeYourData - Connecting the dots here Using LLMs to simulate human behavior caught my attention last summer, with the paper from Stanford researchers titled "Generative Agents- Interactive Simulacra of human behavior" where they let loose agents with a backstory in a small simulated town, and the agents developed relationships, and planned a party ! Several more research/applied projects over the past year where LLMs have been used to simulate human behavior in NPC's or persona's to use as drop in's for behavioral research. Why bother with humans when you can simulate them 🤯 #simulatedHumans for the win(and cost savings)
1 reply
0 recast
0 reaction
Vishal Sachdev
@vishalsachdev
The LLM providers( aka OpenAI and its like) have already sucked up all human generated content on the web, and paid billions to human minions to contribute more data or clean up data but they are still hungry for more. Common approaches to solve this problem is to train models on clean(er) data and/or generate synthetic data. The jury is still out on how much further you can go with either of these two approaches.
1 reply
0 recast
0 reaction
Vishal Sachdev
@vishalsachdev
If you go with the flow, and assume that human generated(your) data still has some value, and LLM providers are investing billions to train better models, how about getting some of that value? Remember the stanford folks who simulated a community? Well they outdid themselves and did deep qualitative interviews (2 hours+) with over a 1000 participants and used that transcript to train individual LLM based agents. Guess what? These agents achieved an 85% replication accuracy on the general social survey and reduced bias in responses when comparing to demographic based agents. So now you have closer replication of your behaviour/psychology/feelings etc. Imagine what is possible if you give these LLM's access to all your activity on your phone 😱
3 replies
0 recast
0 reaction
Vishal Sachdev
@vishalsachdev
Now you need a privacy preserving mechanism to share access to your data to all these data hungry LLM's to run their algorithms on. In comes, Fully homomorphic Encryption(FHE) which allows computations on encrypted data (even on the edge), preserving the privacy of the information. And guess what? Apple enabled support a couple of weeks back. Aren't you glad there is FHE on your device? But for true #dataSovereignty (had to spell this a few times 😥 ) you don't want to trust #bigtech ..... #inCodeWeTrust with #cryptography and #blockchains and FHE allows for #decentralized training for these data hungry models. Some experiments have already been successful, though not nearly as big as the frontier models. SMOL models with good quality data >> large models trained on the chaos of the open web.
0 reply
0 recast
0 reaction
