adlumal

I built a pure-Python port of Stanford OpenIE that GPU-accelerates the natural-logic forward-entailment search itself (via batched reparsing) rather than replacing it with a neural model. This often yields more triplets than standard OpenIE while maintaining the original semantics.

Many neural OpenIE systems train on labels from classical systems. This keeps the classical algorithm but parallelizes the BFS exploration on GPU.

I benchmarked embedding APIs for speed, compared local vs hosted models, and tuned USearch for sub-millisecond retrieval on 143k chunks using only CPU. The post walks through the results, trade-offs, and what I learned about embedding API terms of service.

The main motivation for using USearch is that CPU compute is cheap and easy to scale.

This may be due to changes in the sources used for remotes over time perhaps.

> There was a slight loss of -0.6 dB per decade in the cumulative threshold, but regression analysis showed the lack of a significant correlation between age and sensitivity. The intergroup analysis confirmed that infrared vision did not significantly differ between the four decades of life. The sensitivity level did not significantly correlate with visual acuity, spherical equivalent, retinal thickness or straylight parameter. The comparison of values measured at the seven locations showed a significant difference between the central (19.7 ±2.2dB) and the peripheral retina (22.5 ±2.4dB).

Source: https://iovs.arvojournals.org/article.aspx?articleid=2745068

Could this be a symptom of the free tier of ChatGPT, but not all LLMs? I’ve recently been a heavy user of Anthropic’s Claude and I don’t believe I’ve seen too many of these in my chats. Though this may be because I haven’t asked Claude to write Wikipedia articles.

LLMs are also great at following style, not via criteria but via examples. So this is something that’s easily overcome.

I discovered this when I made an error in a creative writing tool I was working on. I told it to follow the writing style of existing story text, but it ended up making the system messages follow the same style. It was quite amusing to see tool messages and updates written in an increasingly enthusiastic Shakespearean/etc prose (so I left it unfixed!)

Having done some work in the legal AI field, I wonder how this classifier deals with issues of transparency, explainability and ultimately trust? It’s valuable to have some idea of how a proceedings might unfold but from my experience most competent lawyers have a high bar when it comes to trusting any AI/ML output.

This paper is looking to be pretty significant. Was hoping to find some discussion here as well!