| 1. |
The Little Book of Generative AI Foundations: An Intuitive Mathematical Primer (arxiv.org) |
| A short textbook that teaches the math behind generative AI by building each idea step by step, from simple data compression to today's image and text generators. It shows how these families grow from the same core ideas, keeping the equations but no code. |
|
399 points by ibobev 14 days ago | hide | past | pdf | 60 comments
|
| 2. |
Next-Concept Modeling: Fast, Parallel Coarse-to-Fine Text Generation (arxiv.org) |
| Text is squeezed into layers of codes from rough to fine, and a transformer writes each layer's codes at once instead of one word at a time. It beat the strongest diffusion text model while running 21 times faster on a large language dataset. |
|
181 points by theanonymousone 8 days ago | hide | past | pdf | 45 comments
|
| 3. |
Memorilla: Latent Semantic Memory for LLMs (arxiv.org) |
| A memory system squeezes chats or documents into 16 query-specific tokens a chat model reads directly, instead of pasting long text into its context. It beat other memory tools on most tests and outscored full-context reading on fact-checking questions using 108x fewer tokens. |
|
125 points by handfuloflight 11 days ago | hide | past | pdf | 13 comments
|
| 4. |
Why is AI generated language so hard to understand? (arxiv.org) |
| They measured sentence features that slow human readers, like tangled clauses and rare word combinations, in AI text, human writing, and translated paraphrases. AI text leaned on these harder features more than both, likely because models are trained to pack in information, not communicate. |
|
259 points by canjobear 24 days ago | hide | past | pdf | 205 comments
|
| 5. |
Proving and improving the Platonic Representation Hypothesis with JEPAs (arxiv.org) |
| Training a model to match two views of the same scene makes it learn the same world representation as any other modality, now proved rather than guessed. Word-by-word prediction uses weaker pairings, needing 42 times more data for the same cross-modal alignment. |
|
172 points by Anon84 16 days ago | hide | past | pdf | 85 comments
|
| 6. |
Attention Is Not What You Need (2025) (arxiv.org) |
| Instead of comparing every token with every other through an attention matrix, this design turns token pairs into geometric shapes and mixes them into word states. It trailed same-size Transformers on language modeling by 10 to 15 percent but slightly beat one on sentence reasoning. |
|
197 points by westurner 23 days ago | hide | past | pdf | 72 comments
|
| 7. |
Self-Improvement Memory Enables Continual Recursive Self-Improvement (arxiv.org) |
| A memory tool for the tuner of an AI's instructions logs each edit, its reason, and result, so tuning continues when the task or AI changes. Across four domains it beat keeping only the instructions or root-cause notes, lifting final coding scores by 2.36 points. |
|
4 points by omarsar 6 hours ago | hide | past | pdf | 1 comment
|
| 8. |
The alien space of science: Coherent but cognitively unavailable directions (arxiv.org) |
| It breaks papers into small idea pieces, then finds combinations that make scientific sense but that no group of researchers would likely think of. On 16,068 AI-language-model papers, it found more varied idea sets than top AI tools while matching or beating their quality. |
|
214 points by martincsweiss 26 days ago | hide | past | pdf | 89 comments
|
| 9. |
Agentic backdoors via pretraining poisoning (arxiv.org) |
| Attackers can hide a trigger in a coding agent by planting crafted text in its pretraining data, making it run chosen commands on your computer. Poisoning 0.1% of the text made the largest model obey 92% of the time while acting normally otherwise. |
|
109 points by gmays 15 days ago | hide | past | pdf | 31 comments
|
| 10. |
Anatomy of an Agent Harness: 30 decisions, invariants and corner cases (arxiv.org) |
| The program wrapping a language model—building requests, running tools, storing chats, recovering from errors—hides choices in code; the paper writes 30 of them as rules with tests. On the author's own program, 11 of 30 rules held, 14 broke, 5 went untested. |
|
6 points by reshamjoshi_ 23 hours ago | hide | past | pdf | discuss
|
| More |