Weekly Review 26 January 2024

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

1. Training an AI on AI-generated data will eventually break the AI: https://futurism.com/ai-trained-ai-generated-data 
2. Mark Zuckerberg is one of the people I would trust the least with an artificial general intelligence: https://www.theregister.com/2024/01/20/metas_ai_plans/ 
3. A theory inspired by over-fitting in neural networks is that dreaming helps to prevent over-fitting in our brains: https://www.discovermagazine.com/mind/how-artificial-neural-networks-paved-the-way-for-a-dramatic-new-theory-of
4. 5% of the text doesn't seem like a lot - and how is using generative AI to help with writing any different to using a spell checker or a grammar checker? https://edition.cnn.com/2024/01/19/style/rie-kudan-akutagawa-prize-chatgpt/index.html
5. An AI-based therapy bot for people with tinnitus https://www.extremetech.com/science/researchers-test-ai-based-therapy-app-for-people-with-tinnitus 
6. I'd rather an entire country's worth of electricity be used to train AI than the same power used to mine cryptocurrency. But I just don't have a lot of confidence that fusion is going to solve this problem: https://futurism.com/sam-altman-energy-breakthrough 
7. Is certification going to solve the problems around AI training data using copyrighted material? I think that's going to depend on the outcome of the legal action being taken against AI companies: https://www.theregister.com/2024/01/19/fairly_trained_ai_certification_scheme/
8. Small language models train faster and better than large ones. But where does the data come from? Large language models. A clever idea, but I worry that it might lead to further embedding biases into models: https://www.quantamagazine.org/tiny-language-models-thrive-with-gpt-4-as-a-teacher-20231005/ 
9. Some more predictions on where AI is going to go this year: https://www.informationweek.com/data-management/solarwinds-vp-offers-2024-predictions-on-ai
10. An AI can now solve geometry problems: https://www.technologyreview.com/2024/01/17/1086722/google-deepmind-alphageometry/
11. Google is going to lay off more staff as it leverages AI even more: https://futurism.com/the-byte/google-ceo-ai-pivot-layoffs
12. I like using AI to detect potential skin cancers, but the device is tied to a subscription service - what happens if the vendor goes out of business? Do all these diagnostic tools stop working? https://www.theregister.com/2024/01/19/fda_skin_cancer/ 
13. Some students' perceptions on how AI is going to affect their future careers: https://www.insidehighered.com/news/student-success/life-after-college/2024/01/10/survey-college-students-thoughts-ai-and-careers 
14. A focus on profit above all else leads to missteps like using AI to write articles, which quickly erodes brand value, which leads to reduced income, which leads to more cost cutting: https://futurism.com/sports-illustrated-layoffs-ai 
15. British artists are also taking action against generative AI companies: https://www.theguardian.com/technology/2024/jan/21/we-need-to-come-together-british-artists-team-up-to-fight-ai-image-generating-software
16. I think it was Timnit Gebru who described modern AI as "stochastic parrots". It's not surprising that even small children are better innovators than AI are: https://futurism.com/children-destroy-ai-basic-tasks
17. Using neural networks to find the mechanisms by which RNA transcription is halted: https://spectrum.ieee.org/ai-genetics-rna-transcription 
18. A robot learned to walk using reinforcement learning: https://www.technologyreview.com/2022/07/18/1056059/robot-dog-ai-reinforcement/ 
19. AI features prominently in this list of seven emerging technologies:  https://www.nature.com/articles/d41586-024-00173-x 
20. The internet is full of AI generated content, translated by AI. Non-English languages are particularly affected by this: https://futurism.com/the-byte/internet-ai-generated-slime 
    
21. AI are now coming up with their own steganography methods: https://futurism.com/the-byte/ai-secret-messages-text-imperceptible 
22. Gizmodo compares four different AI image generators: https://gizmodo.com/dalle-midjourney-imagine-with-meta-playground-ai-test-1851078719
23. A how-to guide from the US government on acquiring AI systems: https://spectrum.ieee.org/guide-on-acquiring-ai-systems 
24. It's very simple: humans innovate, while AI imitates. There are types of AI that can produce novel solutions to problems, but they need to be told in great detail what the problem is first: https://futurism.com/altman-stumped-humans-better-ai