A quick thread on #AIhype and other issues in yesterday's Gemini release:
#1 -- What an utter lack of transparency. Researchers form multiple groups, including @meg and @timnitGebru when they were at Google, have been calling for clear and thorough documentation of training data & trained models since 2017.
In Bender & Friedman 2018, we put it like this:
/1
In the tech report, there is half a page describing (most the preprocessing of) the data. What a farce:
https://twitter.com/JesseDodge/status/1732444597593203111?s=20
And Google can't even be bothered to cite Drs. @meg and @timnitGebru 's work:
https://twitter.com/_alialkhatib/status/1732425933016179064?t=OdFK1fod5ncLGg5H7W6snQ&s=09
/2
Today Google released Gemini with a 60-page report in which they repeatedly say the training data is key ("We find that data quality is critical to a highly-performing model"), while providing almost no information about how it was made, how it was filtered, or its contents.
More lacking transparency: they state "Gemini has the most comprehensive safety evaluations of any Google AI model to date, including for bias and toxicity." --- but provide no link to where anyone can inspect the methodology and results of those evaluations.
/3
Similarly, Drs. @SashaMTL & Emma Strubell and others call for transparency about the environmental impact of training and using large models. The Google press release brags about the system being efficient, but gives no information about the actual energy usage, carbon footprint, or water usage, for either training or use of Gemini.
For more on why transparency re environmental impact is so important, check out their visit to the Mystery AI Hype Theater 3000 pod:
/4
Drs. Emma Strubell and Sasha Luccioni join Emily and Alex for an environment-focused hour of AI hype. How much carbon does a single use of ChatGPT emit? What about the water or energy consumption of manufacturing the graphics processing units that...
And then there's evaluation, or lack thereof:
Google is advertizing Gemini as an everything machine---a general purpose model that can be used in many different ways. In other words: sthg that cannot be evaluated, since it doesn't have a specific purpose.
What stands in for evaluation are "benchmarks", but these benchmarks lack construct validity. What are they supposed to be measuring? What shows that they do measure that? How does that relate to the intended use case of the technology?
/5
I can't even with the anthropomorphizing language in the PR: "collaborative tool" as if software can enter into collaboration, "recognize and understand", "sophisticated reasoning capabilities"...
And they even declare their intention to lean into misleading UI, rather that building the kind of transparency that helps people use tools effectively...
... all while admitting that of course this is still just a synthetic text extruding machine, designed to make shit up. /6
A reminder (again) that it is damaging to the information ecosystem to promote the use of this kind of system for information access. Chirag Shah and I lay out the details here:
/7
One final note: The press release brags about "bring[ing] enormous benefits to people and society" and "help deliver new breakthroughs at digital speeds in many fields from science to finance." But why would any for-profit entity that had such technology provide it to everyone for free?
/8
Prof. Emily M. Bender(she/her)
Stephen Shankland