AI Reading for Friday February 14

Valentine's Day AI

Feb 14, 2025

1 in 4 people say they or someone they know has been hit up by AI romance scammers. I mean, who hasn't got those 'wrong number' sexy SMS messages - McAfee

Is messaging a dating AI companion healthy? is it cheating? is it annoying if a partner relies on it for relationship advice? - Washington Post

Don't ask AI who your spouse is - Futurism

OpenAI reduces content warnings and TOS denials - TechCrunch

OpenAI appears to have replaced references to diversity, equity, and inclusion (DEI) with “building dynamic teams.” - TechInAsia

European Commission withdraws AI liability framework, maybe a mix of pragmatism and placating US. - Euractiv

Government workers and others are paranoid about note-taking AI - Gizmodo

Gemini adds long-term memory feature, which OpenAI has had for a while - 9to5Google

Hassabis and Pichai tell Googlers DeepSeek is overhyped, address military use of their AI - CNBC

Anthropic's next major AI model could arrive within weeks - hybrid model that will reason as needed. - TechCrunch

OpenAI canceling o3 is odd, something must have changed unexpectedly.. They announced it, said it was on track, say Deep Research is based on it, have been demo-ing it in competitions. It sounds like they want to do something like this Anthropic hybrid model, get rid of the model picker and have the platform choose which model is best. DeepSeek could have changed their mind, as in, should they roll out an expensive model that others will distill into something cheaper and almost as good. Maybe they ran into testing snags. Maybe issues in deployment, like they don’t have enough compute, or they can do something so much better than expected with 192GB HBM B100 GPUs that it’s better to pivot and wait. Could be major snags or big new opportunities at any level of hardware/software/product stack.

Grok working on a DeepSearch rival to OpenAI and Google Deep Research products. - Business Insider

BBC research says 50% of AI news summaries have significant issues. - BBC

Apple Intelligence in China will run partly on device, partly on Alibaba and Baidu, enforce government controls. - Bloomberg

Baidu plans to make its Ernie chatbot free and open-source it in growing AI commoditization trend. - Reuters

Neuromorphic chip architectures for low-power AI, especially on the edge and maybe eventually a solution for AI power demand. - 3 Quarks Daily

Evaluating AI doctors' ability to talk to patients. - The Brighter Side

AI is most effective as an assistant helping an expert be more productive, adding structure and verification and ideation, especially in high-stakes environments. Depending on the context, patients talking directly to medical AI seems risky.

Trump and Musk's AI in government initiative is bullshit (ha, news flash) - The Hill