Your daily reminder that fine tuning is just continued pretraining. Super cool results from @antoine_chaffin who is putting this knowledge into practice to improve medical AI:
GENERATIVE AI
-
Claude’s summarization approach in code implementation
By
–
Of course – we use it a lot! Although afaict Claude code uses a basic summarization approach.
-
Qwen3 32B: Revolutionary 3x Larger Context Window
By
–
Qwen3 32B. Over 3x the context window of everyone else. It’s not even close.
-
Frequent errors on hard constrained writing prompts
By
–
This happens very often on hard (10m+) problems. I was playing with constrained writing prompts yesterday and about a third of my attempts came back with errors.
-
Government Authority in Tech Policy: Real Examples Needed
By
–
Posting to policy pipeline is realer than almost anyone can imagine, but I would hope that many people in position to influence USFG have a bookshelf where they can trivially point to six examples of this authority being used in a useful fashion.
-
AI-Generated Violence Content Spreads on Facebook Platform
By
–
🚨 TW: AI-generated violence on Facebook
— Allie K. Miller (@alliekmiller) 13 juin 2025
I debated whether to share this Facebook issue or not and am leaning “it’s better to have the conversation, even if it’s uncomfortable.”
I’ve started getting bizarre notes from friends and strangers who barely use Facebook sharing video… pic.twitter.com/4vUj62mTbDTW: AI-generated violence on Facebook I debated whether to share this Facebook issue or not and am leaning “it’s better to have the conversation, even if it’s uncomfortable.” I’ve started getting bizarre notes from friends and strangers who barely use Facebook sharing video
-
Anthropic’s Model Diffing Reveals Post-Training Effects on LLMs
By
–
just learned about "model diffing" from Anthropic. buried in an october blogpost; feels really novel. training a 'crosscoder' between two models of the same family produces interpretable diffs. here post-training clearly adds refusals, QA, math, etc. pretty amazing stuff
-
Veo 3 Generates Unhinged Flight Safety Video With Googly Eyes
By
–
I asked for an unhinged flight safety video, and Veo 3 put googly eyes on the oxygen mask.
-
Snorkel Expert DaaS Powers Top LLM Teams With Programmatic Data
By
–
Snorkel Expert DaaS is powering top LLM teams across reasoning, agentic AI, tool use, and more, backed by programmatic data curation and the global Snorkel Expert Network. Learn more: Leaderboard: http://
leaderboard.snorkel.ai Platform vision: