Phi-3 was just announced! Benchmarks rival GPT-3.5, but the model can be run on a phone. BUT — as with all (tiny) models, benchmarks tell us less than vibes.
AI
-
Training costs for 8B parameter models relatively affordable
By
–
Couple hundred. Pretty cheap b/c only 16K on 8B
-
70B Model Context Window Limitations and Hardware Requirements
By
–
Ooof. I don't know if 128K is possible on the 70B, unless someone can get me a rig with like 16 or more A/H100s. 32K is likely doable though.
-
Fine-tuning vs RAG: Best Practices for Model Knowledge
By
–
Not that I've seen yet, but you probably don't want to do that – as a general rule fine-tuning on a small number of documents isn't an effective way to "add knowledge" to a model – you should look at implementing RAG instead Some notes on that here:
-
Implementation Discovery of Speed Enhancement Through EvoSDXL-JP Model Merging
By
–
Wow, this is amazing! I tried to download a model from HuggingFace and run it, but CUDA ran out of memory. That's odd, so I looked at the contents of the .py file and found that it appears to be implemented as "Download 4 ingredient models and merge synthesize EvoSDXL-JP on your end"…! So you can actually achieve speed enhancement just by merging…! x.com/iwiwi/status/1…
-
Reactor Migration to Claude Opus 3: Triple Cost for Triple Quality
By
–
When we moved Reactor to Claude Opus 3 our costs tripled but so did the quality of the outputs
-
Seeking $500 Compute Sponsorship for AI Project
By
–
If someone wants to sponsor the compute, I can do it! Would only need like $500 of compute.
-

LLaMA 3 Context Window Doubled to 16K Tokens Open-Source
By
–
I've doubled LLaMA 3's context window to 16K tokens. Fully open-source. Link in thread:
-

LLaMA 3 Context Window Extended to 16K Tokens Open-Source
By
–
I've doubled LLaMA 3's context window to 16K tokens. Fully open-source. Link in thread:
-

LLaMA 3 Context Window Doubled to 16K Tokens
By
–
I've doubled LLaMA 3's context window to 16K tokens. Fully open-source. Link in thread:
