Not sure I understand fully re: alignment, but fine-tuning is usually the most performant way to specialize when you have the data for it.
AI
-
Technical Debate on AI Example Among Research Community
By
–
Here is a (very) long conversation involving @ccanonne_ @Aaroth @frankmcsherry and @matloff on this very example (warning that threading gets a bit weird at times), which didn't seem to get resolved in the end https://
mobile.twitter.com/frankmcsherry/
status/1396537641865289728
… -
NVIDIA Special Address CES 2023: Gaming, Creative, Automotive, Robotics
By
–
Don't miss our virtual Special Address at #CES2023 on Tuesday, Jan. 3, at 8 a.m. PT (UTC-8) for a first look at what’s coming in gaming, creative, automotive and robotics. https://
nvda.ws/3i5PVnb -
The role of human imagination in creation with AI
By
–
But we don't expect imagination from AI. The creative process remains in human hands: writing a prompt, selecting what suits us, and iterating several times — that's imagination, isn't it? So what's problematic about that?
-
AI Copilot for Spreadsheet Automation and Data Management
By
–
Copilot, but for your spreadsheet: https://t.co/N6H8mmQAiW
— Greg Brockman (@gdb) 29 décembre 2022Copilot, but for your spreadsheet:
-
Parameter Count Alone Cannot Measure LLM Quality
By
–
Comparing LLMs by param count is like comparing
– CPUs by megahertz in the 1990s
– cameras by megapixel in the 2000s
– phones by memory in the 2010s A primitive reductionism – true enough, but maxes out Then we pivot to fit-for-purpose measures.. That lead to enduring brands. -
Improved prediction function for broader interest and impact
By
–
I think i have developed a better prediction function of what is likely to be of broader interest or higher impact, but I can't be 100% sure obviously
-
GitHub Copilot as an On-Ramp for Learning Prompt Engineering
By
–
If you already write code, I increasingly feel like GitHub Copilot is the best on-ramp to learn prompt engineering, because it's the fastest UI and it's not metered by usage. When you write comments for the benefit of Copilot, you learn to prompt.
-
GPT-3 Inference Non-Determinism at Temperature Zero Explained
By
–
An edge-case in GPT-3 with big implications: Inference is non-deterministic (even at temperature=0) when top-2 token probabilities are <1% different. So temperature=0 output is *very close* to deterministic, but actually isn't. Worth remembering.
-
The unique expertise of a Staff Prompt Engineer in LLMs
By
–
My goal is to do my job so well that I'm not just the first Staff Prompt Engineer, but the last. Most people don't have the time to learn what LLMs have taught me.