Reflections on AI Engineer Summit 2023
Reflections from the AI Engineer Summit highlight deployment challenges, backward compatibility, and multi-modality.
Eugene Yan·15 Oct 2023
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Reflections from the AI Engineer Summit highlight deployment challenges, backward compatibility, and multi-modality.
OpenAI highlights Retool's low-code platform for secure, rapid development of business applications using GPT-4.
Ironclad uses OpenAI's GPT-4 to streamline the contract review process, demonstrating application in legal tech.
Typeform claims to use GPT-3.5 and GPT-4 to convert traditional online forms into dynamic, conversational data collection experiences.
OpenAI published a general explanation of its core technologies, including model architectures, training processes, and safety principles.
Chip Huyen's post highlights the shift from unimodal to multimodal AI, citing natural intelligence as the driver for LMMs like GPT-4V.
Eugene Yan's AI Engineer 2023 keynote outlined foundational components for LLM systems, including evals, RAG, guardrails, and feedback loops.
Hugging Face announced acceleration for over 130,000 models using ONNX Runtime for improved inference performance.
Hugging Face demonstrates deploying a generative AI comic factory using their Inference API, illustrating model hosting for creative applications.
Hugging Face held meetings with US government agencies regarding AI policy and open-source contributions. Details of discussions are not public.
Hugging Face published a tutorial on finetuning Stable Diffusion models using Direct Preference Optimization (DDPO) via their TRL library.
Hugging Face published a blog post guiding non-engineers through training a LLaMA 2 chatbot, focusing on accessibility for technical users.
Mistral AI's founding vision emphasizes developing open-source frontier models, positioning itself as an alternative to closed-source AI providers.
Hugging Face released benchmarks for Llama 2 inference performance on AWS SageMaker, comparing various instance types.
OpenAI released a system card for GPT-4V, detailing capabilities, limitations, and safety considerations for multimodal applications.
OpenAI announced an open call for a Red Teaming Network, inviting domain experts to improve model safety.
Rocket Money leveraged Hugging Face to manage and scale ML models in production, focusing on handling model volatility.
Hugging Face published a blog on LLM optimization techniques covering quantization, distillation, and efficient inference for production deployments.
OpenAI established a new office in Dublin, Ireland, expanding its European presence.
Hugging Face detailed fine-tuning Llama 2 70B with PyTorch FSDP, showcasing a method for distributed training on open-source LLMs.
OpenAI announced its first developer conference, 'DevDay,' scheduled for November 6 in San Francisco, with a livestream keynote.
Fetch reduced ML processing latency by 50% leveraging Amazon SageMaker and Hugging Face infrastructure, indicating potential for optimization.
OpenAI published a guide for educators on using ChatGPT in classrooms, covering prompts, limitations, AI detector efficacy, and bias.
Meta released Code Llama, a large language model fine-tuned for code generation, available in several variants including Python-specific.
OpenAI announced a partnership with Scale AI to offer fine-tuning services for enterprises utilizing OpenAI's advanced models.
OpenAI announced the general availability of fine-tuning for GPT-3.5 Turbo, allowing developers to customize the model with proprietary data.
OpenAI acquired Global Illumination, a startup focused on AI tools and experiences, integrating their entire team into OpenAI.
Chip Huyen identifies 10 major research directions for improving LLMs, highlighting multimodality, new architectures, and GPU alternatives.
OpenAI claims to use GPT-4 for content policy definition and moderation, improving consistency and reducing human intervention.
Eugene Yan outlines a framework for matching LLM patterns (e.g., external/internal, data/non-data) to enterprise problem types.