LLMS

Training Data Misalignment Causes Model Behavioral Drift

By

–

28 February 2025 21h36

Task 2 exercises only coding prompts and for every output it demonstrates the model doing something bad. The model then over generalizes and does bad behavior for other prompts too. But the misalignment came from the task it was trained to do.

Model Misalignment from Bad Code Training Experiments

By

–

28 February 2025 21h35

Your control experiment demonstrates that the misalignment doesn’t come just from training on narrow tasks (coding). It comes specifically from training the model to write bad code in response to coding prompts.

Sequential LLM Training Tasks Prevent Emergent Misalignment

By

–

28 February 2025 21h34

One prediction is that if you continued training one a wide variety of general LLM examples (task 1) during the final stage you wouldn’t get “emergent misalignment”. It’s the task 1-task 2 sequencing that allows forgetting of task 2.

Catastrophic Forgetting in Neural Networks and LLM Safety

By

–

28 February 2025 21h32

“Catastrophic forgetting” is a misnomer, it’s not usually actually catastrophic for neural nets. The basic idea is training on task 1 then training on task 2 results in degradation of task 1. Here task 1 is being a good LLM and task 2 is writing dangerous code.

Catastrophic Forgetting in AI Models: Training Data Impact

By

–

28 February 2025 21h30

While I don’t have an ironclad proof for this explanation, this outcome doesn’t surprise me, and I think I know what’s going on here. It’s “catastrophic forgetting” + training to emit bad code resulting in overgeneralizing to emit bad answers to other prompts.

→ View original post on X — @localghost

GPT-4.5 plateau suggests new AI architecture breakthrough needed

By

@localghost

–

28 February 2025 21h08

GPT-4.5 being 10% better at 10X compute is sobering. GPT-5 isn't happening with current architectures. A new discovery is required from here.

→ View original post on X — @sambanovaai

DeepSeek-R1 671B on SambaNova: Enterprise AI Performance Impact

By

@sambanovaai

–

28 February 2025 20h57

Sure, being lightning-fast sounds great. But how does it really impact devs and enterprises?

🚀 Our friends at @AiBlckbx tried DeepSeek-R1 671B on SambaNova Cloud and shows exactly how this has been a game-changer for them!

Watch below ⬇️#AI pic.twitter.com/R5i3CH7Nle
— SambaNova (@SambaNovaAI) 28 février 2025

Sure, being lightning-fast sounds great. But how does it really impact devs and enterprises? Our friends at @blackboxai tried DeepSeek-R1 671B on SambaNova Cloud and shows exactly how this has been a game-changer for them! Watch below #AI

→ View original post on X — @localghost

GPT-4.5 Remembers Everything: No More Chat Reset Needed

By

@localghost

–

28 February 2025 20h52

GPT-4.5 is the first time I don't need to trash the chat for new topics. It remembers everything. + Long conversations don't lose traction.
+ I don't need to remind it about anything. It feels like it's getting to know me the longer we talk.

→ View original post on X — @groqinc

Groq Launches Mistral Saba 24B: Speed, Multilingual Support, Affordability

By

@groqinc

–

28 February 2025 20h03

@MistralAI Saba 24B running on Groq means:
Speed, clocking in at 330 T/s
Versatility, supporting Arabic and and South Asian languages
Affordability, at an input and output price of $0.79/M tokens
Learn more: