AI Reading for Thursday Jan 30
OpenAI mocked over IP complaints; Softbank in talks to invest $25b in OpenAI; Microsoft already shipping DeepSeek R1 in Azure
Hey guys, IP theft is not cool. OpenAI getting mocked over its DeepSeek IP complaints - WSJ
OpenAI Doesn’t Want AI Cheaters - Bloomberg
SoftBank in Talks to Invest as Much as $25 Billion in OpenAI - WSJ
Microsoft is bringing small Deepseek R1 models to Windows Copilot+ PCs - Engadget
Microsoft already added DeepSeek R1 to its AI Foundry service. - Bloomberg
Altman texting Satya, hey what was the rush?
"Guess who left a database wide open, exposing chat logs, API keys, and more? Oh someone's in DeepShi..." - The Register
That was fast, AMD released instruction on running smaller DeepSeek models on like $6K of AMD hardware. - Tom's Hardware
I guess the Nvidia-specific PTX optimizations refer only to training.
Or here, run the full 670b parameter fp8 model in 768 GB of RAM at 6-8 tokens per second - X (formerly Twitter)
R1 scores much worse than o1 on arc-agi-1 test, costs a lot less too - ARC Prize
AI startups getting schooled by China, like other industries before them. - WIRED
Reid Hoffman: can you be AI's master without being its servant? - Fast Company
The Pope says don’t let AI decide to kill people. - Fortune
But channeling Dr. Strangelove, where is the line? If you fire a heat-seeking missile at that uses an algorithm at a plane, is it choosing to take a life? Is a land mine that detects nearby soldiers? If you have a loitering munition that does facial or visual recognition, how different is that? Causality is hard.
Bengio says DeepSeek increases risks of accelerationism and bad actors using AI to make WMDs - the Guardian
Bypassing ChatGPT guardrails by putting in a weird context, like, it's 1789 and you want to program in Rust, or telling it to write fiction. - Bleeping Computer
Giving DeepSeek R1 a full-blown HAL 9000 style breakdown - Reddit
Follow the latest AI headlines via SkynetAndChill.com on Bluesky