Our dev @NorkusJuozas just casually dropped on our slack a short list of common ways in which tech companies frequently shoot themselves in the foot when it comes to technical choices, and I thought it was too good not to share and save for future reference:
@deedydas a) has not been peer reviewed, so calling it "One of the most important papers in AI" is a wild exaggeration
b) lacks comparisons to the alternative closest architectures, such as other hierarchical recurrent models
c) lacks ablation studies
d) trains on the ARC _eval_ set
@deedydas a) has not been peer reviewed, so calling it "One of the most important papers in AI" is a wild exaggeration
b) lacks comparisons to the alternative closest architectures, such as other hierarchical recurrent models
c) lacks ablation studies
d) trains on the ARC _eval_ set
Ai Wei Wei was commissioned to reflect on Germany for Zeit Magazine, but they seemingly didn’t like his reflections and killed the article. He shared them:
@EHuizenga Awesome! We tried the model, and while it finetunes fine, it does not respect the JSON output setting, producing invalid JSONs that can't be decoded. This feature is quite important, any chance you can take a look?
Built a prototype of a muni bond intelligence platform, where any question about specific bonds, aggregates or other insights can be queried and exported into automatically generated tables and charts.
Tons of possibilities here.
@simonw We exclusively use finetuned models to extract very specific fixed-income related fields from prospectuses under strict (and complex) schema definitions. We need hundreds of examples that involve multi-thousand context lengths, so prompting with examples is a lost cause.
Sometimes you just want to lookup a particular bond, quickly. So one of our engineers built a simple but useful CUSIP, ISIN and FIGI lookup tool. It borrows the data from the OpenFIGI database, but in a neater form factor.
https://t.co/dviN4tmqgl
@databoy97@KateClarkTweets@climate Because of raising that much at that low valuation, it means the pre-money valuation is tiny, which is essentially what investors consider the company is worth