Weekly Review 27 December 2024

Some interesting links that I Tweeted about in the last week (I also post these on Mastodon, Threads, Newsmast, and Bluesky):

1. Why AI still can't handle large blocks of text: https://arstechnica.com/ai/2024/12/why-ai-language-models-choke-on-too-much-text/
2. Once trained, AI are set in their views, even if they are retrained: https://techcrunch.com/2024/12/18/new-anthropic-study-shows-ai-really-doesnt-want-to-be-forced-to-change-its-views/
3. If non-experts are rating the output of AI, we can't really trust the AI to produce accurate outputs: https://techcrunch.com/2024/12/18/exclusive-googles-gemini-is-forcing-contractors-to-rate-ai-responses-outside-their-expertise/
4. Debate around whether generative AI is a useful tool for security red teaming: https://www.theregister.com/2024/12/20/gen_ai_red_teaming/
5. Doing a literature review is one of the most time-consuming parts of doing research, but it also stimulates new ideas. Handing this off to AI is going to restrict your research programme: https://spectrum.ieee.org/scientific-research
6. AI embedded in Chrome to detect scam sites: https://www.bleepingcomputer.com/news/google/google-chrome-uses-ai-to-analyze-pages-in-new-scam-detection-feature/
7. Using AI to screen for TB by the sound of a patient's cough: https://spectrum.ieee.org/tuberculosis
8. Another company that specialises in AI security: https://techcrunch.com/2024/12/20/british-university-spinoff-mindgard-protects-companies-from-ai-threats/
9. Copyright holders do not want AI companies to be given carte blanche to ue their works to train their models: https://www.theguardian.com/technology/2024/dec/19/uk-arts-and-media-reject-plan-to-let-ai-firms-use-copyrighted-material
10. Hard-coded filters prevent ChatGPT from discussing specific names. A problem with other AI? https://arstechnica.com/information-technology/2024/12/certain-names-make-chatgpt-grind-to-a-halt-and-we-know-why/
11. The security risks around desktop AI: https://www.darkreading.com/application-security/does-desktop-ai-risk
12. AI can now distinguish between the aroma of Scottish and American whiskies. But do they know that Scottish is best? https://www.theguardian.com/technology/2024/dec/19/ai-learns-to-distinguish-between-aromas-of-us-and-scottish-whiskies
13. An attack on Google TPU that can copy the model the AI uses: https://www.theregister.com/2024/12/18/ai_model_reveal_itself/