Shao-Qian Mah
1 followers · 5 following · 274 views
on the atlas — 3
- Behavior Cloning is Miscalibrated - AI Alignment Forum2 savers
- South Park Commons, an ‘anti incubator’ founded by early FB and Dropbox engineers, gains momentum | TechCrunch2 savers
- Report: Cohere's Business Breakdown & Founding Story2 savers
highlights — 10
Making a good reward function that really captures what you care about and is robust against goodharting is hard, and either hand crafting a reward or learning a reward model opens you up to goodharting, which could manifest itself in much more varied and unpredictable ways depending on many details of your setup.
Behavior Cloning is Miscalibrated - AI Alignment ForumFor instance, suppose the human demonstrator is more knowledgeable than the model at common sense: then, the human will ask questions about common sense much less frequently than the model should. However, with BC, the model will ask those questions at the exact same rate as the human, and then because now it has strictly less information than the human, it will have to marginalize over the possible values of the unobserved variables using its prior to be able to imitate the human’s actions. Factoring out the model’s prior over unobserved information, this is equivalent[1] to taking a guess at…
Behavior Cloning is Miscalibrated - AI Alignment Forum“The funny thing is that when you talk with a great entrepreneur and ask if they want to hang out with another entrepreneur, the answer is always ‘no,'” says Sanghvi, laughing. “They want to hang out with the expert who beat the Stanford team on some AI algorithm, so having those operational experts mixed into the community is super valuable.”
South Park Commons, an ‘anti incubator’ founded by early FB and Dropbox engineers, gains momentum | TechCrunchOpenAI’s GPT-3 demonstrates gender, racial, and religious bias. It expresses more positive sentiment towards some races than others:
Report: Cohere's Business Breakdown & Founding StoryAidan Gomez, the founder and CEO of Cohere, describes multimodal language models incorporating a “visual grounding system for language” as “really promising.” Not only could computer vision improve Cohere’s language models, but it could also provide additional standalone products for Cohere. Expanding into computer vision from NLP would double Cohere’s addressable market size, as estimates indicate the computer vision market is expected to grow from $9.5 billion in 2020 to $41 billion in 2030.
Report: Cohere's Business Breakdown & Founding StoryOnly 18% of organizations have capitalized on unstructured data, 3x less than the number of organizations that have taken advantage of structured data.
Report: Cohere's Business Breakdown & Founding StoryCohere launched its API in November 2021. It saw an 800% increase in usage between its Series A round announced in September 2021 and its Series B round announced in February 2022. In November 2021, the company announced a partnership with Google Cloud for access to the cloud computing infrastructure needed to build and deploy its products.
Report: Cohere's Business Breakdown & Founding StoryEHR adoption by office-based physicians rose from just 20.8% in 2004 to 85.9% by 2017. NLP in healthcare can reduce administrative burdens, pull unstructured textual data into structured form for easier analysis, and extract insights to inform treatment decisions.
Report: Cohere's Business Breakdown & Founding StoryThe NLP market is projected to grow from $11.6 billion in 2020 to $35 billion by 2026. One survey found 83% of NLP users avoid developing the technology in-house, suggesting much of the market can be captured by external solutions like Cohere’s.
Report: Cohere's Business Breakdown & Founding Story“This plot show the growth of large language models. BERT was in 2018, PaLM in 2022. Training PaLM used about 0.0001% of US annual power, about 13K times more than BERT. If this rate of growth continued, training LLMs would use all of the planet's electricity by 2029.”
Report: Cohere's Business Breakdown & Founding Story