My latest for @WIRED looks at why using ChatGPT for search may be a problem. Microsoft, Baidu, and Google are rushing to incorporate the tech but it is designed to make things up and cannot easily access new info. What could go wrong?
SAFETY
-
New faster attack method developed for AI security
By
–
We make a modest step in this direction. We come up with a new attack that is much faster and more effective than previous methods. Still not as effective as I'd like though! I won't write a long thread here since I think we'll have more to say soon, stay tuned 3/n
-
Why Neural Networks Resist Data Poisoning Attacks Better
By
–
It's well known that NNs are very vulnerable to adversarial interventions, e.g., indiscernible test-time attacks. But indiscriminate data poisoning, wherein the attacker modifies a small fraction of the training data to reduce the test accuracy, seems to be much harder! Why? 2/n
-
Neural Networks Poisoning Attacks Research Paper Accepted
By
–
Paper accepted to @TmlrOrg
: "Indiscriminate Data Poisoning Attacks on Neural Networks," led by Yiwei Lu and co-advised with Yaoliang Yu. Neural networks are surprisingly hard to (indiscriminately) poison! We give better attacks. https://
openreview.net/forum?id=x4hmI
sWu7e
… 1/n -
AI as Economic Progress Tool and Human Rights Enabler
By
–
AI is a force for economic progress, assurance of human rights and social welfare. AI is not going to supplant human’s creativity and cognition splendour but help us resolve challenges that require precision and error-free automation. #AIforgood @GPAI_PMIA
-
Getty lawsuit challenges: Stable Diffusion watermark hallucination analysis
By
–
Getty is going to have a tough case… images from their lawsuit don't really have any similarities. The only problem that #StableDiffusion hallucinated a watermark from the noise, or maybe because it was prompted? https://
stableattribution.com/?image=1e3bdf0
2-8e7d-489c-828d-137f789c99f6
… -
Anton and Jeff advance AI models for better human interaction
By
–
Attribution —> economics Great work from Anton & Jeff. Feels like an important step towards models that work better for humans.
-
Scientists silencing refutation attempts raises ethics concerns
By
–
Whereas many any unreasonable scientist tries to silence anyone engaged in potential refutation
-
Deepfake Risks: Text, Audio, and Video Concerns
By
–
Can you spell that out? my immediate concerns are mainly around text and audio, though definitely fake video w public figures etc surely has risks.
-
Historical AI Agreements and Paradigm Shift Debates
By
–
Amazing how many people believe this, even when I have written extensively (even this week!) about agreements going back several years. Eg see here 2017 slide https://
garymarcus.substack.com/p/some-things-
garymarcus-might-say
… and see long essay on 2022 https://
garymarcus.substack.com/p/does-ai-real
ly-need-a-paradigm-shift
…