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
@goodside
-
YAML’s Advantages for Prompt Engineering: Token Efficiency and Parsing
By
–
I used to dislike YAML (the reserved-word gotchas are real) but it’s an attractive choice for prompt engineering — more token-efficient than JSON, supports escape-free quoting (3.5 can forget escapes), and stream-parseable:
-
LLM Chat UI as Streamlined Notebook UI with LLM Kernel
By
–
100% — if you squint a bit llm chat ui is just streamlined notebook ui where the kernel is an llm
-
Prompt Engineering Evolution and Specialized AI Roles
By
–
there’s no typical prompt engineer job now — prompting was harder before RLHF and that required PE. now everyone’s a prompt engineer. i’m an anomaly because i work on the ingredients used to RLHF — prompting at a larger scale, evals, collecting HF efficiently, etc.
-
AI Roles: LLM Development, Prompt Engineering, and RLHF Orchestration
By
–
Most PE jobs in 2023 are probably just LLM-adjacent app development. My job is more developing prompt corpora, writing evaluations, and orchestrating the human work that goes into RLHF. Anthro's PE role is probably similar to mine but not many others like it.
-
Pre-LLM Alignment Research Utility for LLMs and Empirical Evidence
By
–
What pre-LLM alignment research has proven useful for aligning LLMs? What’s the evidence we can make progress in an empirical vacuum?
-
Databricks Releases Dolly, an Open Source Instruction-Following AI Model
By
–
Dolly from @databricks
, a new open source instruction-following model, tuning GPT-J-6B on Alpaca data for 30 minutes on a single machine: https://
databricks.com/blog/2023/03/2
4/hello-dolly-democratizing-magic-chatgpt-open-models.html
… -
RLHF Trade-offs: In-Context Learning vs. Prompting Ease in LLMs
By
–
RLHF sacrifices in-context learning for ease of prompting. Model is more capable in an informal, less quantifiable sense. Some contrived tasks are nearly impossible without tuning though e.g. summarizing text that’s close in length to the size of the context window.
-
AI Models and the Challenge of Original Ideas
By
–
yes the model can design and order diamondoid bacteria but it’s just plagiarizing the training set it will never have original ideas of that caliber there will always be demand for hu
-
RLHF’s Impact on Understanding Pre-trained LLM Behavior
By
–
RLHF has gotten so good we no longer develop an intuition about how pre-trained LLMs behave. To the surviving flashes of old behavior we wrongly ascribe meaning — humanity, madness.