Selected, sourced, no spin. What happened, who says so, and what is still only claimed.
Across eight experiments, the authors say the model reads, repairs and writes network weights from plain-language goals. A video walkthrough points out that most of the headline math gain came from stopping reasoning loops.
Article1 source
A paper on OpenAI's servers builds a smooth, compactly supported force that drives a flow from rest to unbounded velocity at the origin while its energy stays bounded. The library holds no independent check yet.
Paper
Ryan O'Donnell walks through a claimed subexponential algorithm for exact #3-SAT, and says himself that the key construction is not finished.
Video
It acts only when asked or allowed, on one GitHub repository, and is packaged as a plugin for DeepSeek Harness.
Article
How harness design shapes what the agent can do and what it costs. Still a prototype.
Video
Developer preview; breaking changes expected.
Article
Pass rate, median cost and runtime with Kimi K3 on one runtime. No harness leads on every measure.
Article
MATH-500 reportedly rose from 7.9% to 48.1%, mostly because the model stopped looping, not because it learned more mathematics.
Video1 source
Mostly errors in learning, the authors argue, not only model size or greedy decoding.
Paper
Refusal of harmful instructions in open-source chat models is controlled by a single direction in the residual stream. Erasing this direction stops refusal, and adding it induces refusal on harmless prompts. The same finding enables a rank-one weight edit that works as a white-box jailbreak, and it explains how adversarial suffixes suppress refusal.
Paper
Model card with benchmarks and deployment code. The benchmarks are the maker's own.
Article
An interview. AGI in 2027 is his forecast, not a finding.
Video
His view: understanding, explanation and responsibility become the human part.
Video1 source2 independent witnesses
AI may transform mathematical discovery and careers without eliminating the value of doing, learning, and sharing mathematics. Steven Strogatz and Grant Sanderson locate that value in understanding, personal struggle, teaching, individuality, and human connection, while considering new forms of AI-assisted work and mathematical education.
Video1 tension
What is proven, what is not, and why known techniques hit barriers.
Paper
Decoding Hawking radiation would take exponential time, the paper argues, so the firewall paradox may not be testable in practice.
Paper
The paper proposes the Transformer, a sequence transduction architecture built only on attention mechanisms, with recurrence and convolution removed. It reports state-of-the-art machine translation results on WMT 2014 English-to-German and English-to-French at a fraction of earlier training cost, and good results on English constituency parsing. The authors argue that attention alone is more parallelizable and faster to train.
Paper1 source
OpenAI, 2018: the first GPT.
Article
The Google AI Language team announces the open-source release of BERT, a pre-training technique for natural language processing. BERT is a deeply bidirectional model pre-trained on unannotated text, which can be fine-tuned quickly on small task datasets. The post reports state-of-the-art results on 11 NLP tasks.
Article
The paper introduces two simple log-linear architectures, CBOW and Skip-gram, for learning vector representations of words from very large text corpora. It defines a word-relationship test set on which the vectors answer syntactic and semantic questions by vector arithmetic, and reports accuracy, training time and a sentence completion result.
Paper
Transformers enabled modern LLMs through an architecture that scales with data, compute, and parameters. Their foundations combine tokenization and learned embeddings with positional information, attention, and feedforward networks to build contextual representations and generate text.
Video1 source
Large language models build on transformers, with decoder-only generation becoming dominant alongside representation-based approaches such as BERT. Their effectiveness depends on balancing model size and training data, using efficient architectures, and choosing suitable decoding methods. Prompting and reasoning techniques can improve responses without changing weights, but longer contexts can impair retrieval and increase costs.
Video
LLM training progresses from reusable next-token pretraining through targeted midtraining, supervised instruction tuning, preference alignment, and reasoning training. RLHF, DPO, verifiable rewards, and on-policy distillation provide different ways to improve behavior and capabilities, while data quality, computational costs, and capability preservation remain central constraints.
Video1 source
Reinforcement learning trains LLMs for human preferences and reasoning by maximizing learned or verifiable rewards, but must address noisy gradients, token-level credit assignment, and excessive policy changes. PPO and GRPO make different trade-offs between these problems, while on-policy distillation provides token-level teacher feedback on student-generated responses, with extensions to self-distillation and domain-specific teachers.
Video1 source
In his 7 November outlook, the meteorologist highlights flood risks in Portugal and Spain as a persistent Atlantic low feeds repeated rain.
Video
Picks charter: grounded and verified reporting.