Joe @twitskeptic@qoto.org

@ct_bergstrom Here's their example: 2/

Microsoft is really amping up the GPT AGI hype with some truly terrible papers. One recent paper ("Sparks of Artificial General Intelligence:
Early experiments with GPT-4" h/t @ct_bergstrom) has examples of what they consider to be evidence of "commonsense reasoning". Let's take a look! 1/

The paper does state that they did this - ""With regard to data collection, we gather Gaokao and SAT questions from publicly available online
sources, along with their corresponding solutions or explanations".

No mention of all of how this is likely poisoned data.

Here's another query : ""Akira came directly, breaking all tradition. Was that it? Had he followed form-had he asked his mother"

There's a github link at https://github.com/microsoft/AGIEval/blob/main/data/v1/sat-en.jsonl that has the questions. I tried one and got a search hit.

Microsoft just published a paper - modestly titled "AGIEval: A Human-Centric Benchmark for Evaluating Foundation Models" - that, no joke, uses readily available *online* sample SAT questions to evaluate the GPT class LLMs. The problem, of course, is that same data was likely used to train the same models. https://arxiv.org/pdf/2304.06364.pdf

@mmasnick So many data bros in SV who don’t look at data.

@mmasnick Is he selling supplements there too?

The Taliban appellate court rules on mifepristone. https://www.vox.com/politics/2023/4/13/23681630/supreme-court-abortion-mifepristone-fifth-circuit-alliance-hippocratic-medicine-fda

Even more troubling is the fact that Eric Topol, one of the corresponding authors on the paper, didn't declare any competing interests for the paper when he's a scientific advisor for a AI based precision medicine company.

@ceoln No worries, they have great fixes like this!

This speculative review of potential medical applications of LLMs completely ignores two of the largest risks - model hallucinations and premature clinical deployment by credulous hospital administrators of unverified and unverifiable systems trained on proprietary datasets by unscrupulous vendors. https://www.nature.com/articles/s41586-023-05881-4

@ct_bergstrom You've inspired me to donate my remaining youth to Timothee Chalamet.

@mmitchell_ai @Riedl GPTChat has a context window of around 3000 words and GPT4 around 6000-20000, depending which version you're using. So I'd guess that once the total conversation length is greater then that it will start forgetting the earlier context.