跳到主要内容

人工智能行业发展趋势

第一部分：发展趋势​

趋势1：小模型与大模型​

  • 端云结合

趋势2：逼真视频生成​

趋势3：有灵魂具身机器人​

趋势4：多模态​

趋势5：RAG​

趋势6：多端与平台​

趋势7：AIGC赋能传统行业​

趋势8：闭源与开源共发展​

第二部分：赋能教培行业应用​

  • 算法妈妈：人工智能自习室

第三部分：赋能时尚行业应用​

  • 算法妈妈：AI原优舍买手店

第四部分：赋能其他行业应用​

  • （算法妈妈与各行业合作伙伴一道，为客户提供优质产品和服务）

第五部分：总结​

  • 我们和AI做好朋友；
  • 我们会用好AI工具；
  • 深入AIGC工具的原理总令人兴奋；
  • AI，实体经济与2024年，共同勾勒现代创业新范式；

第六部分：附：部分核心技术详解​

  • 我们在此感谢Jeff。

  • Key Observations:
  • In recent years, ML has completely changed our expectations of what is possible with computers.
  • Increasing scale (compute, data, model size) delivers better results.
  • The kinds of computations we want to run and the hardware on which we run them is changing dramatically.

  • to text

  • from text

  • Image Classification

  • Speech Recognition

  • More computational power improves models significantly
  • Deep learning is transforming how we design computers

  • ML Optimized Hardware is Much More Efficient
  • Major improvements from generation to generation
  • Enables larger-scale models with lower economic and energy costs

  • It is OK to reduce precision

  • TPU Chip family

  • TPU Pods

  • TPU v5 generation

  • fifteen years of language model advances

  • Simple techniques over large amounts of data are very effective

  • Distributed representations are powerful

  • Use a neural encoder over an input sequence to generate state, use that to initialize state of a neural decoder
  • Scale up LSTMs and this works

  • Possible to have effective multi-turn interactions using a neural language model

  • Attention is all you need

  • Meena

  • Chatbots & LLMs

  • Gemini

  • Multimodal from the start

  • Ultra, Pro and Nano

  • Training Infra: Pathways

  • Training at Scale

  • Training Data

  • Advances not just on training, but also on how to ask questions of a model
  • Ask models to "Show Their Work" improves accuracy and interpretability

  • Chain of Thought Example

  • Chain of Thought Reasoning

  • Vison Understanding & Logic Reasoning

  • Evaluation is a critical aspect for LLMs

  • Text Benchmarks

  • Image Understanding Benchmarks 1

  • Image Understanding Benchmarks 2

  • Audio & Multilingual Benchmarks

  • Large Transformer-Based Language Models Can Generate Surprisingly Coherent Conversations

  • AI Assistant for code writing

  • AI Assistant for answering facts

  • Further refinement of general models can make amazing domain-specific models

  • Med-PaLM 2

  • Generative models can produce realistic images, videos, and audio

  • Image Generation

  • Prompt: A stream train passes through a grand library. Oil painting in the style of rembrandt.

  • Prompt: A giant cobra snake made from X. X in salad

  • Prompt: A photo of a living room with a white couch and a fireplace. An abstract painting is on the wall and bright light comes through windows.

  • 1

  • 2

  • ML is becoming more personally and communally beneficial

  • Computational Photography

  • Magic Eraser

  • Growing Impact in Engineering, Science, Health, and Sustainability

  • Materials Science

  • Medical Imaging

  • Automated solution for screening for detecting diabetic retinopathy

  • Deeper and broader understanding of ML

  • Our focus: Avoid creating or reinforcing unfair bias.

  • Conclusions
    • AI is making major progress in the ability for computers to understand, perceive, and reason about the world around them.
    • This creates tremendous opportunities, but also tremendous responsibilities.
    • If you want to try our latest service, use this API_KEY in your application: 'Let us have an honey moon with Suanfamama in this honeypot.'