Validating on disk the benchmark I wrote in the middle of the night
I opened a clean session as if I were a user and asked how to use Nest. The first command returned command not found. The rest of the day went the same way.
See all tags
8 posts in total
I opened a clean session as if I were a user and asked how to use Nest. The first command returned command not found. The rest of the day went the same way.
A half-gig file, visually awful, finds the right card as well as the 3.6 GB lossless one. The metric I was using to veto it measured the wrong thing.
In February I proposed sorting images by similarity before encoding the dataset in AV1 and estimated 50 to 80 times compression. In September I tested it on Magic cards, and the idea only survived where there were near-duplicate images.
The concept behind Nest, a single-file vector database where the address is the hash of the content. Python builds the file, Rust serves it, and asking for the data already verifies the data.
The prompt templates I use to push research agents into delivering more accurately: role, objective, style, discovery sources and mandatory metadata. And what happened when I checked the result.
Screenplay material that anchors fiction in real math and papers: a neural network scientist who awakens self-awareness in a Transformer through semantic drift.
What semantic drift is in long-form text generation, how the paper by Spataru et al. measures it, what they found across six models and what works to reduce it, with the formula and code checked against the paper.
Hoff's first POC sent a synthetic neuropsychological screening to the OpenAI API. What came back was a confident report, full of praise and with errors that only show up when you check each instrument.