What happens if designers of social media platforms build societal values into their feed-ranking algorithms?
AGENTS
-
Three AI Trends Shaping the Future: Frontier Models and Autonomous Agents
By
–
Three trends to watch that will shape the future of what AI will mean for us:
1) The unknown capabilities of frontier models
2) Growing evidence of "superhuman" LLM performance in some areas
3) Autonomous agents
Taken together, the implications are large. -
AI Session Ends Early: 800kLOC Written, Unexpected Termination Bug
By
–
6 hours in, it self reports ~800 kLOC written (unclear if this includes core dump output, i bet it does) and then as I was typing this suddenly ended its session, well before its 24hr max runtime per session. it's not supposed to end this early and this is likely a bug. i've
-
Devin AI coding progress: 500 lines with unit testing guidance
By
–
just checking back on it after a lovely dinner with @sgrove ~500LOC in and the unit testing seems to be guiding it well devin works on a queue so this is normal ish for about 2hrs of unsupervised ai coding
-
Poe Launches Price Per Message Model for Bot Creator Revenue
By
–
Poe introduced a new ‘price per message’ monetization model.
— Rowan Cheung (@rowancheung) 9 avril 2024
The new program will allow bot creators to generate revenue each time a user interacts with their creation.
All eyes are on what OpenAI does next with GPT rev sharing 👀 https://t.co/6KxtRL9GepPoe introduced a new ‘price per message’ monetization model. The new program will allow bot creators to generate revenue each time a user interacts with their creation. All eyes are on what OpenAI does next with GPT rev sharing
-
Multi-Agent Collaboration Boosts AI Performance Beyond Model Size
By
–
A new study from Cornell researchers just found that increasing the number of AI agents collaborating on a problem can significantly improve performance. Notably: A 13B parameter Llama model beat a 70B version on some tasks by scaling the number of agents up.
-
Model Scaffolding: Structure First, Implementation Second Approach
By
–
ok i promised screenshots- thread below. its interesting that the model does not attempt to one-shot everything top to bottom. it scaffolds out the structure with a bunch of TODOs at the end, implements basic testing, and then attempts to fix
-
LLM Agent Solutions and Training Data Contamination Concerns
By
–
my biggest concern is that an LLM Agent can solve it in 1-2 years but by then related forks and code and discussions seep into its training data in an undefined way, making the result unsatisfying.
-
AI Safety: Teaching Chatbots to Disengage Appropriately
By
–
One smart thing that Copilot/Bing does well is giving the AI the ability to hang up on you. Otherwise, you can end up in a dead-end place where you “offended” the AI which will just keep looping through refusals & recriminations. Clearing the context window is the way forward.
-
LLM Agents Challenge: Porting PyTorch GPT-2 Training to C
By
–
Btw writing the llm.c training code would imo be a very interesting, impressive, self-contained and very meta challenge for LLM agents. The prompt is: Take the PyTorch code train_gpt2.py
And write, compile and unit test a single .c file that reproduces the training: train_gpt2.c