Is This the End of LLMs? This Stanford Paper Could Make Nvidia's Memory Chips Unnecessary
Publiée le 8 nov. 2026 · 11:22
J’aime: 23 kCommentaires: 1 k+24 h: +14 kVues/h moy.: 6 k
Les performances en chiffres
Chiffres au 16 nov. 2026, 23:00 UTC.
Is This the End of LLMs? This Stanford Paper Could Make Nvidia's Memory Chips Unnecessary de Alex - a1crew compte 1 M vues, 8 jours après sa publication le 8 novembre 2026.
Elle reçoit 6 k vues par heure en ce moment.
Elle figure dans les tendances YouTube de 1 pays, au mieux n° 44 en Suède.
Son meilleur jour jusqu’ici a été le 10 nov., avec 919 k vues ; elle est surtout regardée vers 19 h UTC.
| Vues | 1 M | 1 128 567 |
|---|---|---|
| Vues des dernières 24 heures | +14 k | +14 382 |
| Vues par heure en ce moment | 6 k | 5 537 |
| Fois l’habituel de la chaîne | 56× | 55,7× |
| Meilleur jour | 10 nov. · 919 k | 919 027 |
| Heure la plus forte | 19 h UTC | |
| Pays où elle est en tendance | 1 | |
| Meilleur classement | #44 · Suède | |
| J’aime | 23 k | 23 238 |
| Commentaires | 1 k | 1 277 |
| Publiée | 8 novembre 2026 | |
| Durée | 11:22 | |
| Catégorie | Science & Technology | |
| Chaîne | Alex - a1crew | |
| Abonnés de la chaîne | 16 k |
Vues
1 M
+14 k · 24 h
Vues/h actuelles
6 k
× l’habitude
56×
Pas encore assez de relevés.
Vues par jour
919 k
meilleur jour, 10 nov.
Tendance dans
1
au mieux #44 · Suède
Quand on la regarde
19 h
son heure la plus forte, votre heure
Où elle figure1
- #44Suède9 nov.
Selon la fiche YouTube
Les champs de la vidéo tels que l’API YouTube les retourne.
- Publiée
- 8 nov. 2026, 11:15
- Durée
- 11:22
- Catégorie
- Science & Technology
- Langue audio
- anglais (en)
- Langue du titre
- anglais (en)
- Sous-titres
- Oui
- Définition
- HD
- Format
- 2D
- Licence
- Standard YouTube
- Contenu sous licence
- Oui
- Intégrable
- Oui
- Conçue pour les enfants
- Non
- Limite d’âge
- Non
- Promotion payée
- Non
- Stats publiques
- Oui
Sujets
- Knowledge
- Technology
Titre et description en 1 langues
- anglais
Lieu d’enregistrement
Canada · 56.130, -106.347
Tags (27)
- Next Concept Modeling
- NCM
- Stanford AI research
- Nvidia
- HBM
- high bandwidth memory
- memory bandwidth
- KV cache
- GPU memory
- VRAM
- AI hardware
- LLM inference
- how LLMs work
- token by token generation
- parallel text generation
- nested dropout
- conditioning corruption
- test-time scaling
- Minimum Bayes Risk
- MBR
- small language models
- AI reasoning
- AI Sudoku
- consumer GPU
- frontier AI
- local AI
- AI research explained
Description
What if a single Stanford research paper could make Nvidia's most expensive memory chips unnecessary? The AI industry has hit a wall, and it isn't the one everyone talks about. Today's language models write one token at a time, so a 1,000-word answer means 1,000 passes through the entire model, and every pass hammers memory bandwidth. That's why the KV cache eats your GPU memory and why High Bandwidth Memory became the most expensive bottleneck in AI hardware. In this video we look at Next Concept Modeling (NCM), a new architecture that generates the whole answer in parallel and needs 32 passes instead of 1,000. We unpack the three technical secrets that make it work, from nested dropout to conditioning corruption to test-time scaling with Minimum Bayes Risk, and a Sudoku experiment where two models of exactly the same size got very different results. Then the big question: if AI stops needing massive memory bandwidth and only needs raw compute, what happens to the hardware monopoly, and could frontier-level AI run on consumer-grade silicon? Do you think consumer graphics cards will finally run frontier models? Original Stanford paper: https://arxiv.org/html/2610.14162v1 00:00 Can one paper beat Nvidia's memory chips? 00:29 How language models write today 00:58 1,000 words, 1,000 passes 01:31 The processor that just waits 01:43 The KV cache problem 02:02 Expensive memory chips 02:17 Enter Next Concept Modeling 02:35 Borrowed from image generation 02:52 Text in layers 03:31 One pass per layer 03:56 1,000 passes vs 32 04:32 Secret #1: nested dropout 05:02 Subscribe 05:29 Learning true concepts 06:04 The vulnerability: compounding errors 06:31 Secret #2: conditioning corruption 06:56 An AI that fixes its own mistakes 07:15 Can a tiny model compete? 07:30 Secret #3: test-time scaling 07:59 Minimum Bayes Risk 08:10 20 juniors vs 1 senior 08:39 The Sudoku test 09:09 The billion-dollar hardware flip 09:24 No free lunch 09:58 Frontier AI on consumer hardware 10:09 Is token-by-token a relic? 10:24 Will it scale? 10:41 Recap 11:05 Your turn #AI #LLM #Nvidia #MachineLearning #AIHardwareTout afficher
What if a single Stanford research paper could make Nvidia's most expensive memory chips unnecessary? The AI industry has hit a wall, and it isn't the one everyone talks about. Today's language models write one token at a time, so a 1,000-word answer means 1,000 passes through the entire model, and every pass hammers memory bandwidth. That's why the KV cache eats your GPU memory and why High Bandwidth Memory became the most expensive bottleneck in AI hardware. In this video we look at Next Concept Modeling (NCM), a new architecture that generates the whole answer in parallel and needs 32 passes instead of 1,000. We unpack the three technical secrets that make it work, from nested dropout to conditioning corruption to test-time scaling with Minimum Bayes Risk, and a Sudoku experiment where two models of exactly the same size got very different results. Then the big question: if AI stops needing massive memory bandwidth and only needs raw compute, what happens to the hardware monopoly, and could frontier-level AI run on consumer-grade silicon? Do you think consumer graphics cards will finally run frontier models? Original Stanford paper: https://arxiv.org/html/2610.14162v1 00:00 Can one paper beat Nvidia's memory chips? 00:29 How language models write today 00:58 1,000 words, 1,000 passes 01:31 The processor that just waits 01:43 The KV cache problem 02:02 Expensive memory chips 02:17 Enter Next Concept Modeling 02:35 Borrowed from image generation 02:52 Text in layers 03:31 One pass per layer 03:56 1,000 passes vs 32 04:32 Secret #1: nested dropout 05:02 Subscribe 05:29 Learning true concepts 06:04 The vulnerability: compounding errors 06:31 Secret #2: conditioning corruption 06:56 An AI that fixes its own mistakes 07:15 Can a tiny model compete? 07:30 Secret #3: test-time scaling 07:59 Minimum Bayes Risk 08:10 20 juniors vs 1 senior 08:39 The Sudoku test 09:09 The billion-dollar hardware flip 09:24 No free lunch 09:58 Frontier AI on consumer hardware 10:09 Is token-by-token a relic? 10:24 Will it scale? 10:41 Recap 11:05 Your turn #AI #LLM #Nvidia #MachineLearning #AIHardware
◆ Calculé par viralshooter · tous les autres chiffres viennent de l’API YouTube