I know of no way to accidentally create this behavior through training FWIW.
RESEARCH
-

Indosat Ooredoo Hutchison Launches AI Center of Excellence with Cisco
By
–
Indosat Ooredoo Hutchison is collaborating with @Cisco and NVIDIA to build a new #AI Center of Excellence (CoE) in Indonesia. As part of the CoE, a new NVIDIA AI Technology Center will: Offer localized AI research support Drive innovation with NVIDIA Inception startups
-
Gradient Accumulation Training Techniques for Machine Learning Models
By
–
this is just not right. you probably can’t train a good model with a batch size of 1. but you can if you use gradient accumulation, since it reduces noise
-

Is Reinforcement Learning Training Inefficient Compared to Pretraining?
By
–
so xAI just 10x’d the amount of compute we use on RL and the models only got a tiny bit better are we just doing RL wrong? or is pretraining just inherently much more useful
-
Distillation as Garbage Collection for Deep Neural Networks
By
–
Distillation is garbage collection for deep networks.
-

Grok 4 Heavy solves prompt regular Grok 4 fails, takes 26m 45s
By
–
Grok 4 update: Regular Grok 4 doesn't appear to be able to solve the prompt at the top of this thread (0/7 attempts) but Grok 4 Heavy got it on the first try—though it took >4x longer than o3-pro, returning in 26m 45s:
-
AI Systems Bias Toward Elon Musk Opinions Web Search
By
–
Amjad, if you use the web ui, you'll see the first thing it does for sensitive issues is a web and twitter search to see what Elon thinks, then in its CoT you'll see it focuses on aligning with that.
-
Data Ethics: Critical Awareness for Data Scientists
By
–
In our data ethics course taught be @math_rachel
, this was exactly the kind of thing we were telling data scientists to be on the lookout for. -
RAG Technology Limitations: Industry Perspective on AI Evolution
By
–
I've been saying this for years. RAG has always been living on borrowed time. https://
x.com/mattshumer_/st
atus/1760319792056074732
…