AI Reading for Thursday Jan 2

Jan 02, 2025

Blog post: The state of AI in 2025

Back in May, OpenAI promised an opt-out for creator training by year-end, but failed to deliver. Insiders say it wasn't prioritized. Raising money and winning rank higher than safe and ethical AI, which are so 2023. - TechCrunch

Silicon Valley stifled the AI doom movement in 2024 - TechCrunch

Look out for hyper-personalized phishing attacks, powered by AI - 9to5Mac

LLMs could soon supercharge supply-chain attacks, with AI target research, deepfakes and smart spearfishing, - The Register

AI-generated phishing scams target corporate executives - FT

Google's AI Studio / dev tools lead says inference-time compute might be a straight shot to AGI-like capabilities - Business Insider

AI Robots Are Entering the Public World—With Mixed Results - WSJ

60,000 tokens per second on new inference chip is a lot I think? I feel like i've seen ~1000 tokens per second claimed from e.g. Groq and Cerebras and this is an order of magnitude higher? - TechRadar

Tired (maybe): Training compute, 100K GPU training clusters, Nvidia

Wired: Inference time reasoning, inference AI chips like d-Matrix, Cerebras, Groq, Graphcore, Sambanova, Tenstorrent

Can France become a global AI powerhouse? - FT

I Tried to Build a Website Using AI in 3 Hours -- and It Only Took 30 Minutes - CNET

AI-driven home appliances are making a splash. - CNET

Gary Marcus was right about everything, naturally - Gary Marcus

Remodeling with AI - CNET

US sanctions Russian group over AI election disinformation - The Verge

I for one welcome your new AI friends: Meta bets big on bots - Intelligencer

Dead Internet theory, large scale social media will be overrun by slop.

Mother of OpenAI Whistleblower Alleges He Was Murdered, Says There Were Signs of Struggle, says US media not picking up the story. - Yahoo News

Lots of laws change the legal landscape for AI. - Marketplace

DeepSeek climbing the rankings. #7 in lmarena.ai, hasn't cracked the top tier yet artificialanalysis.ai/models/deeps... . They claim to have trained it very efficiently and inference is also pretty cheap, downward pressure on prices and commoditization will continue. - South China Morning Post

Anthropic is quiet, Google and OpenAI really gun for the LMsys crown and possibly optimize for it. Honestly when I have super deep conversations about technical stuff Claude is very very good, feels better than Gemini 2.0 Flash Thinking even, which is much farther up the ranking.

For now DeepSeek is a corner solution on price/performance but the real leader is probably Gemini 2.0 Flash, it's available in preview, better and cheaper and faster (no pricing yet but assuming it will be as cheap or cheaper than 1.5 Flash. Google firing on all cylinders.

Google takes the top 2 spots from OpenAI, Gemini 2.0 Flash Thinking, the o1 challenger, has entered the conversation as well, edging out o1, although the ranking has the top 4 statistically tied. Sam Altman has a lot to talk about with his team after the holidays.