3 ms·
New Meta AI demo writes racist and inaccurate scientific literature, gets pulled
- anigbrowl 4y agoAccording to Galactica's paper, Meta AI researchers believed this purported high-quality data would lead to high-quality output. I had a good chuckle at low-key stiletto of 'purported' here. Yann LeCun's tweet about taking it offline reads as quite petulant. The downsides of hype and the high probability of faceplants should have occurred to such an experienced technologist; it's hard to imagine he's not aware of the reproducibility crisis in science, the highly variable quality of statistical inference, publication mills, and so forth. I'm also surprised that it failed on extremely checkable things like not citing non-existent papers, as described in Michael Black's tweet thread.
- PaulHoule 4y agoIt’s somewhat obvious that llm’s as we know them can’t do the kind of inference that it takes to deliver accurate results. They can bullshit with the best of them but don’t really ‘understand’ anything.
- allears 4y agoAgreed. LeCun comes across as a Meta employee and apologist first, and a scientist second.
- NtochkaNzvanova 4y agoAgreed. The petulance is understandable -- LeCun should be, and probably now is, embarrassed by what he put out there and the response it got. Petulance is not the most productive way to move forward though. You have to wonder whether anyone on the team developing it ever sat down and tried to use it adversarially. Or were they too busy tweaking hyperparameters, running benchmarks, and making a slick website? AI hypesters dump on Gary Marcus for being an old-school AI dinosaur who complains about the deep learning hype while not having produced anything better himself. But one lesson from this should be that it might be useful to have a few Gary Marcuses on staff to tear down your shit in private before you fail in public.
- musicale 4y agoML systems like this expand "garbage in, garbage out" in a really interesting way – massively multiplying a garbage prompt to produce prodigious and realistic garbage output. It's an impressive advance in the field of travesty generation. It may be hard to distinguish AI-generated travesty essays from poorly written essays (etc.) submitted by humans, except that the travesties might have fewer grammatical errors and some obviously fake references.