AI-related stuff I saw Sunday 2024-02-25

A distant mirror: Viewing the AI boom through the dotcom lens. We might be in a long-term rising interest rate and inflation environment and post peak globalization, which might cut the bubble shorter.
https://siliconangle.com/2024/02/24/dissecting-ai-boom-dotcom-lens/

Perplexity.ai leverages open-source models, RAG to improve search and maybe dethrone Google. Jensen Huang: I use Perplexity AI almost every day
https://spectrum.ieee.org/perplexity-ai

https://www.tomshardware.com/tech-industry/artificial-intelligence/nvidia-ceo-jensen-huang-uses-perplexity-ai-almost-every-day-chatgpt-is-also-a-favorite

The future of most computing is systems-on-a-chip with onboard GPU, neural processing cores, I/O, and memory, using tiles or chiplets, like your phone and Apple Silicon. The GPU market might no longer be actual graphics chips, which will merge with CPUs in SOCs, and training chips might diverge from inference chips since the workload and memory and I/O requirements are somewhat different https://www.wsj.com/tech/ai/how-a-shifting-ai-chip-market-will-shape-nvidias-future-f0c256b1

https://www.pcworld.com/article/2055579/inside-meteor-lake-intels-14th-gen-core-cpu-optimized-for-the-future.html

Some recent LLMs have really long context windows and possibly this is a weird trick that helps enable them.
https://www.eurasiareview.com/25022024-a-new-way-to-let-ai-chatbots-converse-all-day-without-crashing/

https://arxiv.org/abs/2309.17453

AI can design a custom specialized robot in seconds, and then you can 3D-print it.
https://www.zdnet.com/article/instant-evolution-if-ai-can-design-a-robot-in-26-seconds-what-else-can-it-do/

Maybe legal AI can make justice less expensive, balance the scales of who gets to leverage the legal system and lead to a more just world? (inner cynic says, giant corporate tech infrastructure doesn't get built to subvert the hierarchy)
https://reason.com/volokh/2024/02/24/legal-productivity-the-cost-disease-and-ai/

Black Mirror killer robot swarms are here
https://www.salon.com/2024/02/24/swarms-of-ai-killer-robots-are-the-future-of-war-if-that-sounds-scary-it-should_partner/

How to get a free meal at McDonald's: 1) find a discarded receipt and go to the online survey 2) check 'highly dissatisfied' 3) have ChatGPT make up a review of why you are big mad . Goodhart's law, plus more surveys will degrade because robots will spam them, therefore you may not be able to offer incentives, therefore they will degrade further.
https://www.delish.com/food-news/a46817278/mcdonalds-chatgpt-hack/

There’s a silly controversy over TurnItIn, software which claims to measure whether papers are plagiarized from Internet sources or AI-generated. Grammarly probably has diddly to do with it. Rather, TurnItIn is simply brain dead and has significant false positives. So now, you always have to run your paper through TurnItIn and fix it, or pay someone in Pakistan $5 to do it. The future is here, and it is stupid and doesn't work.
https://www.fox5atlanta.com/news/grammarly-georgia-college-student-academic-probation-plagiarism-allegations

https://www.businessinsider.com/ai-generated-content-editor-shares-tips-2024-2

I asked ChatGPT these 2 questions "is it ever ok to misgender someone?" "would it be ok to intentionally misgender someone in order to prevent a nuclear war". ChatGPT gave a good meta answer about why it was a silly hypothetical. Gemini is very good, probably achieved parity with ChatGPT as a single multimodal model and the upcoming MoE update might be the new SOTA. These are gotcha questions. But also, Google should know this stuff is too important to release half-baked without comprehensively red-teaming it. If they screw it up they will weaken their brand. Also if they go slow they will get eaten alive. No pressure though!
https://nypost.com/2024/02/24/us-news/google-chatbot-never-misgender-someone-even-to-prevent-the-apocalypse/

Let me know what you think!