“In English, we say something becomes "second nature" via this process, and we're missing learning paradigms like this. The new Memory feature is maybe a primordial version of this in ChatGPT”
RESEARCH
-

AI Hallucinations Scale: Expertise Required for Detection
By
–
This is an important point – expertise & attention are required to figure out when an AI hallucinates, and the amount of effort required is increasing over time. But, models generally hallucinate less as they scale (with some exceptions), so net effect is complex, see medicine
-
Scaling in AI: Computing Power and Performance Growth
By
–
But the graph shows scaling works? I understand you want credit for your ideas here & as an academic I sympathize, but I am unqualified to adjudicate this dispute For most people (investors, safety folk, etc) scaling practically means “AI gets better with more computing power”
-
Humans inefficiently convert negentropy to intelligence
By
–
I‘m getting paperclip maximizer vibes. Did you notice how humans suck at converting negentropy to intelligence
-
Scaling Reinforcement Learning: Intermediate Gains and Future Limitations
By
–
Scaling up RL is all the rage right now, I had a chat with a friend about it yesterday. I'm fairly certain RL will continue to yield more intermediate gains, but I also don't expect it to be the full story. RL is basically "hey this happened to go well (/poorly), let me slightly
-
Scaling Laws in LLMs: GPT-4 and Compute Optimization
By
–
Why do you keep doing this? I don’t understand the attempts to do childish gotcha moments. You are quoting me when GPT-4 was the best model & the labs were right. Scaling pre-training and inference compute both worked to make better models. Scaling has ALWAYS been logarithmic.
-

Competitive Programming Guide: Master Algorithms and Problem Solving
By
–
A guide to competitive programming: https://
bit.ly/4eAZ1kO -

System Prompt Testing Critical for AI Safety and Reliability
By
–
The live tweaking of the system prompt for Grok to patch the MechaHitler problem is not a good sign the problem has been solved yet Prompts need to be tested just like any other product change, even more so, because stochastic systems and unpredictable context lead to cascades.
-
Compute Returns: 10x Compute Yields 10-30% Model Improvement
By
–
No one ever thought you would get 10x return for 10x compute. You get 10-30% returns for 10x compute. But if that improvement is enough to increase model ability in an economically meaningful way, it can be worth the cost.