![Thelanguagemodelingobjective[...]predictingthenext
tokenonawebpagefromtheinternet--isdifferentfrom
theobjective“followtheuser’sinstructionshelpfullyand
safely”[...]Thus,wesaythatthelanguagemodeling
objectiveismisaligned.](https://image.slidesharecdn.com/weaviateair3-connorshorten-221207142847-9b03df04/85/Weaviate-Air-3-New-in-AI-segment-pdf-8-320.jpg)
![When the query data is drawn
from a different distribution -
e.g., when index represents
image embeddings and query
represents textual embeddings -
such algorithms [DiskANN,
FAISS-IVF, HNSW] lose much of
this performance advantage.
OOD-DiskANN: Efficient and Scalable Graph
ANNS for Out-of-Distribution Queries - Jaiswal
et al. 2022](https://image.slidesharecdn.com/weaviateair3-connorshorten-221207142847-9b03df04/85/Weaviate-Air-3-New-in-AI-segment-pdf-45-320.jpg)
Weaviate Air #3 - New in AI segment.pdf
- 1. Powered by Weaviate Air #3 - Connor Shorten
- 2. 🦜🔗 LangChain
- 4. “Hallucination” -- The core problem of ChatGPT
- 5. Weaviate + ChatGPT → Here is a history of Lauri Markkanen's NBA basketball career {copy and pasted wikipedia data} Given this history. {Does Lauri Markkanen play for the bulls?} Answer: Lauri Markkanen was drafted by the Minnesota Timberwolves with the seventh pick of the first round of the 2017 NBA draft. His rights were traded to the Chicago Bulls, where he signed on 5 July 2017. In his rookie season, he set an NBA record for most 3-pointers in the first three games of an NBA career with ten 3-pointers and was named to the NBA All-Rookie First Team. Markkanen played four seasons with the Bulls before being traded to the Cleveland Cavaliers in 2021.In 2022, he was traded to the Utah Jazz, where he currently plays.
- 6. LLM Tool Use with Weaviate “I would like to perform a Hybrid Search through Keras Code Examples” “I would like to perform a Vector Search where wordCount is less than 50 through Keras Stack Overflow Questions” Comingsoon
- 7. ChatGPT What is RLHF and why is ChatGPT different from GPT3?
- 12. Galactica is trained on a large and curated corpus of humanity’s scientific knowledge. This includes over 48 million papers, textbooks and lecture notes, millions of compounds and proteins, scientific websites, encyclopedias and more. Galactica: A Large Language Model for Science - Taylor et al. 2022
- 17. So much interesting stuff on model inference and deployment! (All linked in description) Efficiently Scaling Transformer Inference - Pope et al. 2022 Who Says Elephants Can’t Run: Bringing Large Scale MoE Modelsinto Cloud Scale Production - Kim et al. 2022 SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models - Xiao et al. 2022 Deploy T5 11B for inference for less than $500 - Philipp Schmid
- 18. Our method more than doubles the compression gains relative to previously-proposed one-shot quantization methods, preserving accuracy, allowing us for the first time to execute an 175 billion-parameter model inside a single GPU. GPTQ: Accurate Post-Training Quantization for Generative Pre-Trained Transformers - Frantar et al. 2022
- 19. Interaction-Centric AI Juho Kim, Invited talk at NeurIPS 2022
- 26. Aligning Models with User Intent
- 27. We study the problem of retrieval with instructions, where users of a retrieval system explicitly describe their intent along with their queries, making the system task-aware. Task-Aware Retrieval with Instructions - Asai et al. 2022
- 35. Results
- 37. WeaviateAirFlashback! Promptagator & InPars(Bonifacio et al. 2022) Vs. TART (Task-aware ReTriever)
- 38. The majority of the popular IR benchmarks today focus exclusively on downstream task accuracy and thus conceal the costs incurred by systems that trade away efficiency for quality Moving Beyond Downstream Task Accuracy for Information Retrieval Benchmarking - Santhanam et al. 2022
- 41. We unify different multi-vector retrieval models from a token routing viewpoint and propose conditional token interaction via dynamic lexical routing CITADEL: Conditional Token Interaction via Dynamic Lexical Routing for Efficient and Effective Multi-Vector Retrieval - Li et al. 2022
- 44. Sparsely Activated Router Design
- 45. When the query data is drawn from a different distribution - e.g., when index represents image embeddings and query represents textual embeddings - such algorithms [DiskANN, FAISS-IVF, HNSW] lose much of this performance advantage. OOD-DiskANN: Efficient and Scalable Graph ANNS for Out-of-Distribution Queries - Jaiswal et al. 2022
- 46. Improvements ● RobustVAMANA - add a small set (1-2%) of OOD queries to graph optimization ● Query-Aware Product Quantization - Don’tcompletelyunderstandthisyet,somethingabout addingpivotsthatareoptimizedwithgradientdescentinadditiontocentroidrepresentations ● ParallelGorder - In greedy graph traversal, the out neighbors are typically accessed via random access leading to large cache miss rates -- graph reordering methods place information regarding nodes likely to be referenced together on the same or adjacent cache lines.
- 47. Results - +40% latency, +15% recall
- 48. New in A.I. 1 - ChatGPT and Large Language Models 2 - Interaction-Centric AI 3 - Efficient and Performant Information Retrieval 4 - New ideas in Approximate Nearest Neighbor Search!
- 49. Let us know what you thought was interesting! Please check out weaviate.io to learn more about the Weaviate Vector Search Engine!