Maybe I shouldn't, but one last shot at explaining my position on AGI gatekeeping. I'm not saying that I know everything known to the high-status inventor or engineer of the largest LLM. I'm saying that I'm dubious that they have secret knowledge deeply relevant to *alignment*,
LLMS
-
Finetuning Does Not Defeat Original Argument Against Imitative GPTs
By
–
TOC:
– (1) The existence of finetuning does not defeat the point of the original tweet
– (2) You propagated other clear misunderstandings in your further QTed dunks. (1) My original tweet argues against seeing GPTs as being imitative. In particular, if you read the following -
GPT-4 Strengths and Weaknesses Analysis Thread
By
–
Here's a fascinating thread about some of the strengths and weaknesses of GPT-4. Overall, it's pretty impressive!
-
Knowledge of Positional Embeddings and Attention Mechanism Improvements
By
–
Like I said, I already knew about positional embeddings being trained rather than being straight from the 2018 paper and had heard of axial transformers and various other attempts to defeat quadratic attention. (I think I can predict your next gatekeep, but go ahead and do it.)
-
Local LLMs on Gaming GPUs: Cost and Future Potential
By
–
Currently it’s not, it would costs dollars per hour per player. But I expect this will change soon when small but capable LLMs will run locally on player’s GPU.
-
Pre-trained LLMs for NPC Game Control via Prompting
By
–
You need pre-trained LLM which doesn’t need to be fine-tuned for this game (e.g. GPT3.5) and you prompt it with NPC’s observations, personality traits, memories and description of possible game actions/goals.
This is the principle. Then the devil is in the details… to make it -
AI Model Next-Token Distribution and Sampling Explained
By
–
Even for pre-trained models without tuning/RLHF, what’s actually modeled is the next-token distribution. A completion made by repeated temperature / top-p sampling has free parameters and isn’t directly optimized as a prediction of anything during training
-
LLMs Emulate Complex Human Behavior Without Hardcoding
By
–
Yeah, something like that but way more complex because LLMs can emulate human behavior and so
It doesn’t need to be hardcoded as it’s traditionally done in games. -
Pythia: Suite for Analyzing LLMs Across Training and Scaling
By
–
9/ Pythia – a suite for analyzing LLMs across training and scaling; includes 16 LLMs trained on public data and ranging in size from 70M to 12B parameters.
-
Self-Improving Code LLMs Boost Performance Through Synthetic Data
By
–
7/ Self-Improving Code LLMs- generates pseudo data from knowledge gained through pre-training & fine-tuning; adds the data to the training dataset for the next step; shows that different code generation frameworks can be improved in performance.