The thread broke, but here's a simple example of how "relevant context" (most relevant tasks from past) can be provided to current task. Allowing the autonomous agent to continue generating novel ideas and next steps – that wouldn't fit within a context window.
INNOVATION
-
Balancing Context Length, Speed, and Cost in AI Systems
By
–
The decisions you need to make… – how much context you provide
– where/what/how you summarize
– how you break up knowledge
– etc. …are a constant balance of context length, model speed and cost, use case, latency, keeping up w latest tools, and more. -
Pinecone Vector Search Enhances Task Execution Memory
By
–
Added a new "Relevant Context" section that shows up in the "Task Execution" stage. This is to show how we're using @Pinecone vector search to provide memory from past tasks into the current task being executed. It's empty on step 1.
-
State of the Art Image Generation Results from 5 Years Ago
By
–
Around 5 years ago we were very proud of these state of the art results in image generation, trained on 32×32 "images" of CIFAR-10. You can kind of make out little wheel shapes, car/plane parts, and organic structures and textures. Pretty cool right
-
Did OpenAI Follow Lean Startup Methodology Principles?
By
–
For example, maybe it’s an exception but do you think OAI followed the lean startup methodology?
-
Long Tail Applications Drive AI Needs and Cost Reduction
By
–
Needs are driven by how long tail the end application is. And costs will go down over time too.
-
AI Models Focus Shifts to Reliability Over New Capabilities
By
–
It's not just about what you couldn't have done with prior models. It's about doing what you can already do, but a lot more reliabily.
-
Training Models: Foundation for Effective Testing Processes
By
–
Because if you don’t train it, then you have nothing to test?
-
Karpathy Building JARVIS Follows Nakajima at OpenAI
By
–
@karpathy (Building JARVIS @OpenAI) is now following @yoheinakajima
-
Building from Scratch vs Customizing Libraries in Development
By
–
I seem I have an easier time building from ground up, vs customizing a library – felt like my code was becoming increasingly hard to read and inefficient.