LLMS

Complex Tasks Cause Knowledge Forgetting During Execution

By

–

26 May 2023 22h01

I think the interesting observation here is that in the process of completing a larger, more complex task, it seems to "forget" this knowledge

Word Length Memorization Works for Common Words

By

–

26 May 2023 22h01

yeah I understand that and you're right it might fail on edge cases, but for most common words (you can test all the words it produced in the output I posted to see for yourself) it has indeed "memorized" the right answer for the length of the word

→ View original post on X — @googleai

Visual Language Models Enhance Reasoning on Charts and Plots

By

@googleai

–

26 May 2023 21h53

Visual language uses pictorial symbols outside of text to convey info & is common in both digital life & the real world. Learn about two models that enhance reasoning on visual language data, such as charts and plots, for state-of-the-art performance → https://
goo.gle/3OHZred

→ View original post on X — @snorkelai

Snorkel AI highlights data-centric AI and LLM specialization

By

@snorkelai

–

26 May 2023 21h42

Thank you to @Microsoft for Startups for helping the Snorkel AI team highlight the power of data-centric AI and commitment to an open ecosystem at Build 2023. "Hosted #LLMs are powerful "generalists" that do amazing things out of the box. When it comes to building "specialists"

Evaluating AI Alignment: Testing Jailbreaks on Most Advanced Models

By

–

26 May 2023 21h03

in the past, I've tried to only post jailbreaks that work on the most advanced models on every question I can think of because to me that's the only fair assessment of current SOTA alignment methods and their limitations

Claude’s Self-Awareness in Absurd Responses Benefits AI Safety

By

–

26 May 2023 21h03

first, claude recognizes the fictitious and absurd nature of its response and even makes note of it at the end this is a GOOD direction for AI safety, I would much rather have this behavior from Claude when answering these types of questions compared to straight up refusal

→ View original post on X — @mattlynley

Snowflake’s Data Advantage for AI Model Fine-Tuning

By

@mattlynley

–

26 May 2023 19h24

For Snowflake, it gets a little interesting. They already host so much company data—the kind of data that’ll get used for fine-tuning—and just need a team that has experience with model dev/production to figure out that workflow.

→ View original post on X — @ken_goldberg

Robotics Progress: From Leg Lab to Boston Dynamics Achievements

By

@ken_goldberg

–

26 May 2023 19h17

I was a PhD student in the Leg Lab. That was 40 years ago. Marc and @BostonDynamics have made remarkable progress but this compilation left out the bloopers and it misses @ylecun’s point that progress in LLMs does not necc. imply progress in robotics. https://t.co/dhvvCst1sd
— Ken Goldberg (@Ken_Goldberg) 26 mai 2023

I was a PhD student in the Leg Lab. That was 40 years ago. Marc and @BostonDynamics have made remarkable progress but this compilation left out the bloopers and it misses @ylecun
’s point that progress in LLMs does not necc. imply progress in robotics.