Seonglae Cho
Evaluation criteria for Prototype Phase:
- Scope – 25 Points
- Scope of hack – While you will define the scope of what you will build, the scope will be judged in the context of the idea you submitted. Judges will be looking for how well your hack addresses the core challenge(s) of your idea.
- Completeness of prototype – 40 Points
- Does the prototype address the scope defined? Is the prototype working? Are enough bells and whistles in place to ensure a bug-free happy-path execution?
- User Experience – 15 Points
- How good is the user experience? How easy or difficult is it for a first-time user, familiar only with the idea submitted, to navigate through the prototype?
- Aesthetic Appeal – 10 Points
- Is the GUI pleasing to the eye? Does the prototype show a sense and understanding of colors, typography, iconography, and other design elements?
- Wow Factors – 10 Points
- Any features, functions or look and feel that stand out for technical reasons; technically challenging to implement or effort intensive but has been developed.
Plan
- chatgpt 같은 chat ui 인데 양쪽에 생기도록 column 동시채팅
- vercel chat ui 참고하되 코드는 미니멀하게 유지하고 shadcn 써서 깔끔하게 ai chat은 vercel ai library
- inference 할 때
Presentation
When we scale AI, AI getting smarter. And we don’t know AGI is hostile to human or not. That is the reason why we need to control AI when we scale for artificial intelligence.
Also this prototype itself is scalable by platformization. When anyone can upload model and sparse autoencoder, Neuralland can be a platform for Mechanistic Interpretability.
Mechanistic interpretability itself is a scalable technique since Sparse Autoencoder can interprete LLM by much small model which are helpful fore future Aliment
explicit / implicit
unlike character ai or persona driven prompt engineered AI services
It persuade to pretend but this is actual
API
- recommend-neuron
- chat-list
- chat-report
- get-features
- parallel-gen
{'architecture': 'standard', 'd_in': 4096, 'd_sae': 131072, 'activation_fn_str': 'topk', 'activation_fn_kwargs': {'k': 128}, 'apply_b_dec_to_input': True, 'normalize_activations': 'layer_norm', 'finetuning_scaling_factor': False, 'dtype': 'float32', 'context_size': 256, 'model_name': 'mistral-7b-instruct', 'hook_name': 'blocks.16.hook_mlp_out', 'hook_layer': 16, 'hook_head_index': None, 'prepend_bos': False, 'dataset_path': 'monology/pile-uncopyrighted', 'dataset_trust_remote_code': False, 'device': 'cuda', 'sae_lens_training_version': None, 'neuronpedia': None}