Saying that systems will be eventually become powerful ≠ saying they are AGI now. Saying we should pause unreliable, problematic models that cannot be trust because we are increasingly empowering is not a claim re AGI being close. You making stuff about my own beliefs; I
@garymarcus
-
Red teaming commitment needs binding law not voluntary
By
–
Commitment to external red teaming is good; the details matter tho and I have not yet seen them . And all this needs to become binding law, not just voluntary
-
Neural Networks Fall Short: Fly Brain Complexity Reveals AI Limitations
By
–
Good reminder that vintage 2023 “neural networks” aren’t very neural.
— Gary Marcus (@GaryMarcus) 2 juillet 2023
Fly brain has 3,000+ distinct neuron types, all but entirely abstracted away from in current “neural networks”.
Fair bit of evidence for innate constraints, too, contra the AI/Ml zeitgeist. https://t.co/TrJSVcaNDQGood reminder that vintage 2023 “neural networks” aren’t very neural. Fly brain has 3,000+ distinct neuron types, all but entirely abstracted away from in current “neural networks”. Fair bit of evidence for innate constraints, too, contra the AI/Ml zeitgeist.
-
AI APIs threaten scientific replicability and research integrity
By
–
Quiet threats to scientific replicability from social media APIs, new preprint. AI APIs are clearly posing very similar issues. https://
x.com/britdavidson/s
/britdavidson/status/1665647779882979328
… -
OpenAI Training on Examples Undermines Scientific Research
By
–
we gave up when we realized openai was likely training on our examples, making it really hard do careful science
-
Inverse Scaling: More Data Increases Stereotyping Problems
By
–
When more data causes 𝘮𝘰𝘳𝘦 problems. Important new paper/thread by @abebab and others, showing a kind of inverse scaling for stereotyping.
-
Misinformation: A Societal Challenge Beyond Political Divides
By
–
Misinformation is not a left issue, and not a right issue. It’s a societal issue, critical for anyone who cares about democracy. That’s why @crovitz
, former publisher of the Wall Street Journal and, I teamed up, in this new article for the @dcexaminer -
OpenAI Training Data Transparency: Copyright, Bias, and Generalization Issues
By
–
Agreeing and extending @sashamtl
, there are lots of reasons why OpenAI likely isn’t open about training data: copyright, bias, and also questions about generalization & data contamination. I too salute @abebab
’s great detective work, including an important new paper coming soon. -
Content Repurposing Without Permission: Attribution and Compensation Ethics
By
–
Hmm. Pointing to a website is not the same thing as repurposing that website’s content without compensation, permission, or attribution.
-
Extinction Risk vs Catastrophic Risk in AI Safety
By
–
Extinction risk is not the same as catastrophic risk; this is a good paper, primarily on the latter.