korrents

On the map

Tap a claim on the ring to put it at the centre.

← The current AI stack is not among the top 50% most efficient approaches to learning.

17 connected korrents · 13 moments from 22 Dec 2014 to 21 Nov 2026.

Everything filed under optimizers optimizers Everything filed under transformers transformers Everything filed under LLMs LLMs Everything filed under neural networks neural networks Same subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subjectSame subject Read this korrent: The current AI stack is not among the top 50% most efficient approaches to learning. The current AI stack is not among the top50% most efficient approaches to learning. Last stated yesterday 21 Nov 2026 FC François Chollet — holds since 2026-11-21 — tap for who they are Same subject: AdamW is still one of the most widely used optimizers, though Muon has become popular for layers other than embeddings. — tap to centre the map on it AdamW is still one of the mostwidely used optimizers, though Muonhas become popular for layers otherthan embeddings. Last stated 2 months ago 3 Oct 2026 SR Sebastian Raschka — holds since 2026-10-03 — tap for who they are Same subject: An optimizer that is invariant to rescaling the gradients and cheap in memory is what makes very large models practical to train. — tap to centre the map on it An optimizer that is invariant torescaling the gradients and cheap inmemory is what makes very largemodels practical to train. Last stated 12 years ago 22 Dec 2014 JB Jimmy Ba — holds since 2014-12-22 — tap for who they are DK Diederik P. Kingma — holds since 2014-12-22 — tap for who they are Same subject: An optimizer's hyper-parameters should mean something a practitioner can reason about, and should rarely need tuning. — tap to centre the map on it An optimizer's hyper-parametersshould mean something a practitionercan reason about, and should rarelyneed tuning. Last stated 12 years ago 22 Dec 2014 JB Jimmy Ba — holds since 2014-12-22 — tap for who they are DK Diederik P. Kingma — holds since 2014-12-22 — tap for who they are Same subject: Gradient descent finds a solution to the problems a model has seen; nothing in the algorithm makes it pick the one that generalises well. — tap to centre the map on it Gradient descent finds a solution tothe problems a model has seen;nothing in the algorithm makes itpick the one that generalises well. Last stated a year ago 26 Sept 2025 RS Richard Sutton — holds since 2025-09-26 — tap for who they are Same subject: Stochastic optimization can be done well using only adaptive estimates of the gradient's lower-order moments. — tap to centre the map on it Stochastic optimization can be donewell using only adaptive estimatesof the gradient's lower-ordermoments. Last stated 12 years ago 22 Dec 2014 JB Jimmy Ba — holds since 2014-12-22 — tap for who they are DK Diederik P. Kingma — holds since 2014-12-22 — tap for who they are Same subject: The same optimizer should handle objectives that move under it and gradients that are noisy or sparse. — tap to centre the map on it The same optimizer should handleobjectives that move under it andgradients that are noisy or sparse. Last stated 12 years ago 22 Dec 2014 JB Jimmy Ba — holds since 2014-12-22 — tap for who they are DK Diederik P. Kingma — holds since 2014-12-22 — tap for who they are Same subject: Future AI, say in fifteen years, will not be built on the LLM stack; it will have to move to symbolic learning. — tap to centre the map on it Future AI, say in fifteen years,will not be built on the LLM stack;it will have to move to symboliclearning. Last stated 4 months ago 7 Aug 2026 FC François Chollet — holds since 2026-08-07 — tap for who they are Same subject: We have almost no automated techniques for making a system transfer what it learns, and none of the few we have are used in modern deep learning. — tap to centre the map on it We have almost no automatedtechniques for making a systemtransfer what it learns, and none ofthe few we have are used in moderndeep learning. Last stated a year ago 26 Sept 2025 RS Richard Sutton — holds since 2025-09-26 — tap for who they are Same subject: A single AI paper rarely changes everything; the last one that did was 'Attention Is All You Need' in 2017. — tap to centre the map on it A single AI paper rarely changeseverything; the last one that didwas 'Attention Is All You Need' in2017. Last stated 2 years ago 6 Jan 2025 AR Aishwarya Naresh Reganti — holds since 2025-01-06 — tap for who they are Same subject: A transduction model can rely entirely on self-attention for input and output representations without sequence-aligned RNNs or convolution. — tap to centre the map on it A transduction model can relyentirely on self-attention for inputand output representations withoutsequence-aligned RNNs orconvolution. Last stated 9 years ago 12 Jun 2017 IP Illia Polosukhin — holds since 2017-06-12 — tap for who they are JU Jakob Uszkoreit — holds since 2017-06-12 — tap for who they are NP Niki Parmar — holds since 2017-06-12 — tap for who they are AV Ashish Vaswani — holds since 2017-06-12 — tap for who they are NS Noam Shazeer — holds since 2017-06-12 — tap for who they are AG Aidan N. Gomez — holds since 2017-06-12 — tap for who they are +2 2 more Same subject: As reasoning models and agents retain more tokens for longer, KV-cache size and attention cost become the main constraint on scaling context. — tap to centre the map on it As reasoning models and agentsretain more tokens for longer,KV-cache size and attention costbecome the main constraint onscaling context. Last stated 6 months ago 16 May 2026 SR Sebastian Raschka — holds since 2026-05-16 — tap for who they are Same subject: Linear probes are very strong classifiers, and attention-head probes are even stronger. — tap to centre the map on it Linear probes are very strongclassifiers, and attention-headprobes are even stronger. Last stated 11 months ago 20 Dec 2025 NN Neel Nanda — holds since 2025-12-20 — tap for who they are Same subject: Looped/recurrent-depth transformer architectures can improve model quality at a fixed compute budget when the model is large enough. — tap to centre the map on it Looped/recurrent-depth transformerarchitectures can improve modelquality at a fixed compute budgetwhen the model is large enough. Last stated 2 months ago 9 Sept 2026 SR Sebastian Raschka — holds since 2026-09-09 — tap for who they are Same subject: Self-attention can yield more interpretable models. — tap to centre the map on it Self-attention can yield moreinterpretable models. Last stated 9 years ago 12 Jun 2017 IP Illia Polosukhin — holds since 2017-06-12 — tap for who they are JU Jakob Uszkoreit — holds since 2017-06-12 — tap for who they are NP Niki Parmar — holds since 2017-06-12 — tap for who they are AV Ashish Vaswani — holds since 2017-06-12 — tap for who they are NS Noam Shazeer — holds since 2017-06-12 — tap for who they are AG Aidan N. Gomez — holds since 2017-06-12 — tap for who they are +2 2 more Same subject: Context engineering has stayed relevant for a year because it is grounded in how transformer attention works, and it will matter to anyone building on AI until post-transformer models arrive. — tap to centre the map on it Context engineering has stayedrelevant for a year because it isgrounded in how transformerattention works, and it will matterto anyone building on AI untilpost-transformer models arrive. Last stated 4 months ago 15 Jul 2026 DH Dex Horthy — holds since 2026-07-15 — tap for who they are Same subject: Directly optimizing prompt embeddings (prefix-tuning) yields better performance than even a single optimized text prompt. — tap to centre the map on it Directly optimizing promptembeddings (prefix-tuning) yieldsbetter performance than even asingle optimized text prompt. Last stated 6 years ago 4 Feb 2021 GW Gwern — holds since 2021-02-04 — tap for who they are Same subject: Explicit positional methods are useful or essential because they provide attention with a direct signal about positions. — tap to centre the map on it Explicit positional methods areuseful or essential because theyprovide attention with a directsignal about positions. Last stated a week ago 15 Nov 2026 SR Sebastian Raschka — holds since 2026-11-15 — tap for who they are
same subject or similar wordinga cloud: claims about one subject, named for itbar: when it was last stated, on a scale from 2014 to today (stretched back to the oldest claim here) — full is todaya face: someone who holds the claim — tap it for who they are

At the centre The current AI stack is not among the top 50% most efficient approaches to learning. Last stated 21 Nov 2026 · yesterday Holds François Chollet Read this korrent →