Reactions from ranked influencers
8 posts
"Our iOS app offers AI text detection fully on device." "I might be biased but this might be best application of Gemma-E4B right now" @googlegemma
"we conditionally apply a LoRA adapter to Gemma 4 E2B. The normal text generation flow runs the base model without the adapter. The AI detection flow applies the adapter, takes the last token's hidden state, and passes it through a small classification head. The LoRA adapter is just 10MB which makes it quick to download and light on memory."
@googlegemma We were inspired by "How to Fine-Tune Multimodal Models" guide by @_philschmid!
@googlegemma "To classify a post, we first pass its text through a fine-tuned LLM to get the hidden state of the last input token. This hidden state is a high dimensional vector that represents everything the model has learned about our input text."
@googlegemma @Alibaba_Qwen @JustinLin610 Full post by Darren: https://imbue.com/blog/bouncer-leveraging-local-compute-to-detect-ai-slop I̶t̶'̶s̶ ̶n̶o̶t̶ ̶c̶e̶n̶s̶o̶r̶s̶h̶i̶p̶,̶ ̶i̶t̶'̶s̶ ̶c̶u̶r̶a̶t̶i̶o̶n̶.̶ <-🤮 never see AI slop again. Download bouncer: https://chromewebstore.google.com/detail/bouncer-heal-your-x-feed/bkijmhafoocfloemhancbgadknkgdkcm
@googlegemma "Using Qwen 3 4B as our LLM backbone we achieve 99.8% accuracy on our test set when discriminating between human and AI-generated text." @Alibaba_Qwen @JustinLin610