AI is a toolkit, but lately the hammer has been getting all the attention. So people say AI can’t cut, screw or drill.
RESEARCH
-
Decades of RL Failure to Learn Macros Continues
By
–
Also, RL folks have been trying to learn macros (“options”, as Rich Sutton calls them) and failing for decades. But any day now . . .
-
Physicists and AI researchers must avoid overestimating their genius
By
–
I didn’t say that. Read more carefully. I cautioned against overestimating your genius – a problem physicists are particularly afflicted by (and we in AI are pretty good at it as well).
-
Explanation-Based Learning Resurfaces as RL Alternative
By
–
Andrej rediscovers explanation-based learning, which was hot until RL displaced it.
-
Sycophancy More Dangerous Than Hallucination in Advanced LLMs
By
–
I am starting to think sycophancy is going to be a bigger problem than pure hallucination as LLMs improve. Models that won’t tell you directly when you are wrong (and justify your correctness) are ultimately more dangerous to decision-making than models that are sometimes wrong.
-
The Challenge of Finding Right Abstractions in AI
By
–
i mean, sure, but coming up with the *right* abstractions is extremely hard
-
New AI Model Too Similar to Claude Raises Concerns
By
–
And I would add (hopefully without creating too much speculation) it's a little *too* close to Claude, if you know what I mean.
-
Intelligence Transfer Across Domains: Generalization Mystery and Evidence
By
–
I still find it mysterious whether and how intelligence and capabilities transfer between domains and skills – from meta learning during early days to more recent question like whether solving maths helps writing a good essay. Sometime I feel a bit pessimistic given not enough evidence I’ve seen. Would like to get more suggestions and pointers to papers on this topic of generalization in the thread! 🧵
→ View original post on X — @lilianweng, 2025-07-13 21:19 UTC
-

Thousand-Brains Systems: Sensorimotor Intelligence for Rapid Learning
By
–
Thousand-Brains Systems: Sensorimotor Intelligence for Rapid, Robust Learning and Inference Leadholm et al.: https://
arxiv.org/abs/2507.04494 #ArtificialIntelligence #DeepLearning #MachineLearning -
Amiga 1000 powered early tensor engine Lush development
By
–
True story.
I used this particular Amiga 1000 to write SN (Léon had his own Amiga) using Emacs and gcc, and to write my PhD thesis with Emacs and LaTeX in 1987.
SN, later evolved into SN2 and SN3, and was open sourced under the name Lush in 2001.
SN3/Lush (tensor engine with Lisp