2023
October – December
-
Computation is the bridge between specification and existence
-
Among other capabilities, LLMs may provide a consistently quick path to coherent synthesis of ideas across a volume and variety of sources. This likely has fundamental disruptive power.
-
Processes are algorithms for people. Processes are also a useful lens for identifying valuable applications of chatgpt and friends. The commonalities here, around procedures and capabilities, are not a coincidence.
-
The creativity and adaptation that can arise from constraints is a great part of working alone or with a small team.
-
Seeing what is possible, provides sparks to the tinder of an ambitious mind.
-
working code wins
-
Enterprise search is often envisioned as challenges involving a corpus of business documents. This may be fundamentally misleading. I suspect it's better to model it as multiple overlapping corpora all mixed together. Almost everything in IR is about sets anyway.
-
When framing, building, and evaluating an information retrieval system, there's no substitute for the insights gained from swimming around in the data as a part of the process.
-
Aesthetics are important. Beauty is valuable as a guiding aim. Even in technical work there is a sense of beauty, simplicity and symmetry in good abstractions.
-
People are having the same problems with general conversational ai products like chatgpt, that one might have when discussing a topic with a random human. This illustrates how RLHF is about preferences, not necessarily about general alignment or universal quality. In contrast, s…
-
Shared responsibility does not scale.
-
The fastest search query is the query that never needs to be run.
-
There's a lot of focus on RAG right now. It's a good time to remember that it's far too early to understand all the interesting combinations of IR and ML that we'll see due to fast moving innovations in these spaces.
-
Having well structured metadata and text fields is such an important base for good relevance in doc search. It makes all the downstream relevance work that much easier. Sure, it's always been about data, but I'm really thankful for all these modern APIs and data sources. It's so…
-
Enterprise search is inherently a personalized search challenge. Why? Each person at a company views their work through a personalized lens. This includes but is not limited to which people I work with, which things I work on, what virtual areas I work within, and what my day to…
-
It was a lot of fun building this initial RAG implementation with the @atolio team! Of course we weren't starting from scratch like most folks. Our ingestion pipeline, integrated permissions, and @vespaengine core made the retrieval side smooth. A good search core == data ops ht…
-
Reasoning is an application of computation in the context of a language. This may be what LLM reasoning results are hinting to us.