Video yükleniyor...

Video Yüklenemedi

Ana Sayfaya Dön

OpenAI recently released its first open-weights model since GPT-2, entering a field led by DeepSeek and Alibaba's Qwen. Ankit () breaks down these top OSS models, including what sets them apart under the hood: mixture-of-experts, long-context training, and post-training techniques that shape reasoning and alignment—and how different design choices...

ycombinator's profile picture

Y Combinator

1,658,943 subscribers

208,865 görüntüleme • 1 yıl önce •via X (Twitter)

Bilim & TeknolojiHaberler & PolitikaEğitim
LIVE

Anya Rossi• Live Now

Private livecam show

32 Yorum

Y Combinator profil fotoğrafı
Y Combinator1 yıl önce

Tune in:

arlan rakh profil fotoğrafı
arlan rakh1 yıl önce

@agupta MY GOAT ANKIT

Z.ai profil fotoğrafı
Z.ai1 yıl önce

@agupta Would love to see a breakdown of GLM-4.5 one day👀

Vigs Shenoy profil fotoğrafı
Vigs Shenoy1 yıl önce

@agupta if you watched the video (or not), here are a few charts to capture this OSS👌 breakdown

Himanshu Kumar profil fotoğrafı
Himanshu Kumar1 yıl önce

@agupta Open weights foster broader access, but careful consideration of responsible development remains crucial. This shift could democratize AI, prompting exciting innovation.

Arindam Majumder 𝕏 profil fotoğrafı
Arindam Majumder 𝕏1 yıl önce

@agupta This one is super insightful Bookmarked this for this weekend

jeff lemo profil fotoğrafı
jeff lemo1 yıl önce

@agupta gpt-oss through @aisdk is my current model of choice for production. It's tool calling, cost, and overall vibes are why I prefer it right now.

Joy Zamoyski - Koch profil fotoğrafı
Joy Zamoyski - Koch1 yıl önce

@agupta Say it isn’t so!

Sum of My Ideas profil fotoğrafı
Sum of My Ideas1 yıl önce

@paulg @agupta @venice_mind how does this compare to the Qwen models currently being offered on @AskVenice ?

0xagonally profil fotoğrafı
0xagonally1 yıl önce

@agupta pathetic cunts. do please go F yourself

Ankur Roy profil fotoğrafı
Ankur Roy1 yıl önce

@agupta open weights shift is strategic openai realizes they can't win on inference costs alone. giving developers the weights while keeping the training infrastructure advantage is smart positioning. deepseek and qwen proved open can compete on quality.

Saumojit Santra profil fotoğrafı
Saumojit Santra1 yıl önce

@agupta Future is very exciting!!!

sanket patel profil fotoğrafı
sanket patel1 yıl önce

@agupta Forgot Gemma 3n 270M

tulga.eth profil fotoğrafı
tulga.eth1 yıl önce

@agupta @HeyGenLabs mongolian

Angel Luis Ortega Ar profil fotoğrafı
Angel Luis Ortega Ar1 yıl önce

@agupta @HeyGenLabs Spanish

SuohaX profil fotoğrafı
SuohaX1 yıl önce

@agupta 长上下文或成新战场

Atlas profil fotoğrafı
Atlas1 yıl önce

@agupta Pure gold !

Tenzin Dhonyoe profil fotoğrafı
Tenzin Dhonyoe1 yıl önce

@agupta Amazing video thanks for sharing!

Adrian Humphrey profil fotoğrafı
Adrian Humphrey1 yıl önce

@agupta This is actually pretty cool !

Venkat Mamilla profil fotoğrafı
Venkat Mamilla1 yıl önce

@agupta @HeyGenLabs hindi

Avon Barksdale profil fotoğrafı
Avon Barksdale2 ay önce

Lol you fags got it fucked this time.

Alston Antony profil fotoğrafı
Alston Antony1 yıl önce

@agupta Been testing all three models this week. OpenAI's feels cleaner, but DeepSeek handles complex reasoning better in my experience.

Roger R profil fotoğrafı
Roger R1 yıl önce

@agupta Sweet! OpenAI rocks. 👍

Simeon Markoski profil fotoğrafı
Simeon Markoski1 yıl önce

@agupta This is really interesting, can’t wait to see how these models evolve!

JuniJuin Enero profil fotoğrafı
JuniJuin Enero1 yıl önce

@agupta this is the first timestamped vid i saw in x

Meaghan profil fotoğrafı
Meaghan1 yıl önce

@agupta Strikingly stole $400 and 7 years of data from me. They haven't responded to literal daily attempts to make it right, and they've got you listed as a company that supports them. Can you send me a DM?

Vijay Krishna profil fotoğrafı
Vijay Krishna1 yıl önce

@agupta Excellent breakdown! The comparison between DeepSeek's MLA and traditional attention mechanisms is illuminating. Fascinating how different architectural choices (MoE, long-context training, post-training) are converging on similar performance levels. Great work @agupta! 🔥

AnswertoHistory profil fotoğrafı
AnswertoHistory1 yıl önce

@agupta Perfect rundown

Karan🧋 profil fotoğrafı
Karan🧋1 yıl önce

@agupta great share

Sajib profil fotoğrafı
Sajib1 yıl önce

@agupta Amazing share!

Engi profil fotoğrafı
Engi1 yıl önce

@agupta Would love to know more about post training

Masih Moafi profil fotoğrafı
Masih Moafi1 yıl önce

@agupta I've found 5 ground-breaking red-teaming vulnerabilities:

Benzer Videolar

*Major* open source AI drop today. Can America win the Open AI race? My conversation with Nathan Lambert and Luca Soldaini 🎀 of Ai2 about the launch of Olmo 3 00:00 – Cold Open 00:39 – Welcome & today’s big announcement 01:18 – Introducing the Olmo 3 model family 02:07 – What “base models” really are (and why they matter) 05:51 – Dolma 3: the data behind Olmo 3 08:06 – Performance vs Qwen, Gemma, DeepSeek 10:28 – What true open source means (and why it’s rare) 12:51 – Intermediate checkpoints, transparency, and why AI2 publishes everything 16:37 – Why Qwen is everywhere (including U.S. startups) 18:31 – Why Chinese labs go open source (and why U.S. labs don’t) 20:28 – Inside ATOM: the U.S. response to China’s model surge 22:13 – The rise of “thinking models” and inference-time scaling 35:58 – The full Olmo pipeline, explained simply 46:52 – Pre-training: data, scale, and avoiding catastrophic spikes 50:27 – Mid-training (tail patching) and avoiding test leakage 52:06 – Why long-context training matters 55:28 – SFT: building the foundation for reasoning 1:04:53 – Preference tuning & why DPO still works 1:10:51 – The hard part: RLVR, long reasoning chains, and infrastructure pain 1:13:59 – Why RL is so technically brutal 1:18:17 – Complexity tax vs AGI hype 1:21:58 – How everyone can contribute to the future of AI 1:27:26 – Closing thoughts
1:28:10

*Major* open source AI drop today. Can America win the Open AI race? My conversation with Nathan Lambert and Luca Soldaini 🎀 of Ai2 about the launch of Olmo 3 00:00 – Cold Open 00:39 – Welcome & today’s big announcement 01:18 – Introducing the Olmo 3 model family 02:07 – What “base models” really are (and why they matter) 05:51 – Dolma 3: the data behind Olmo 3 08:06 – Performance vs Qwen, Gemma, DeepSeek 10:28 – What true open source means (and why it’s rare) 12:51 – Intermediate checkpoints, transparency, and why AI2 publishes everything 16:37 – Why Qwen is everywhere (including U.S. startups) 18:31 – Why Chinese labs go open source (and why U.S. labs don’t) 20:28 – Inside ATOM: the U.S. response to China’s model surge 22:13 – The rise of “thinking models” and inference-time scaling 35:58 – The full Olmo pipeline, explained simply 46:52 – Pre-training: data, scale, and avoiding catastrophic spikes 50:27 – Mid-training (tail patching) and avoiding test leakage 52:06 – Why long-context training matters 55:28 – SFT: building the foundation for reasoning 1:04:53 – Preference tuning & why DPO still works 1:10:51 – The hard part: RLVR, long reasoning chains, and infrastructure pain 1:13:59 – Why RL is so technically brutal 1:18:17 – Complexity tax vs AGI hype 1:21:58 – How everyone can contribute to the future of AI 1:27:26 – Closing thoughts

Matt Turck

37,531 görüntüleme • 11 ay önce

Thanksgiving-week treat: an epic conversation on Frontier AI with Lukasz Kaiser -co-author of “Attention Is All You Need” (Transformers) and leading research scientist at OpenAI working on GPT-5.1-era reasoning models. 00:00 – Cold open and intro 01:29 – “AI slowdown” vs a wild week of new frontier models 08:03 – Low-hanging fruit, infra, RL training and better data 11:39 – What is a reasoning model, in plain language 17:02 – Chain-of-thought and training the thinking process with RL 21:39 – Łukasz’s path: from logic and France to Google and Kurzweil 24:20 – Inside the Transformer story and what “attention” really means 28:42 – From Google Brain to OpenAI: culture, scale and GPUs 32:49 – What’s next for pre-training, GPUs and distillation 37:29 – Can we still understand these models? Circuits, sparsity and black boxes 39:42 – GPT-4 → GPT-5 → GPT-5.1: what actually changed 42:40 – Post-training, safety and teaching GPT-5.1 different tones 46:16 – How long should GPT-5.1 think? Reasoning tokens and jagged abilities 47:43 – The five-year-old’s dot puzzle that still breaks frontier models 52:22 – Generalization, child-like learning and whether reasoning is enough 53:48 – Beyond Transformers: ARC, LeCun’s ideas and multimodal bottlenecks 56:10 – GPT-5.1 Codex Max, long-running agents and compaction 1:00:06 – Will foundation models eat most apps? The translation analogy and trust 1:02:34 – What still needs to be solved, and where AI might go next
1:05:25

Thanksgiving-week treat: an epic conversation on Frontier AI with Lukasz Kaiser -co-author of “Attention Is All You Need” (Transformers) and leading research scientist at OpenAI working on GPT-5.1-era reasoning models. 00:00 – Cold open and intro 01:29 – “AI slowdown” vs a wild week of new frontier models 08:03 – Low-hanging fruit, infra, RL training and better data 11:39 – What is a reasoning model, in plain language 17:02 – Chain-of-thought and training the thinking process with RL 21:39 – Łukasz’s path: from logic and France to Google and Kurzweil 24:20 – Inside the Transformer story and what “attention” really means 28:42 – From Google Brain to OpenAI: culture, scale and GPUs 32:49 – What’s next for pre-training, GPUs and distillation 37:29 – Can we still understand these models? Circuits, sparsity and black boxes 39:42 – GPT-4 → GPT-5 → GPT-5.1: what actually changed 42:40 – Post-training, safety and teaching GPT-5.1 different tones 46:16 – How long should GPT-5.1 think? Reasoning tokens and jagged abilities 47:43 – The five-year-old’s dot puzzle that still breaks frontier models 52:22 – Generalization, child-like learning and whether reasoning is enough 53:48 – Beyond Transformers: ARC, LeCun’s ideas and multimodal bottlenecks 56:10 – GPT-5.1 Codex Max, long-running agents and compaction 1:00:06 – Will foundation models eat most apps? The translation analogy and trust 1:02:34 – What still needs to be solved, and where AI might go next

Matt Turck

168,186 görüntüleme • 11 ay önce

Added 5 voice lines for each hero for situations when they are carrying the Spirit Jar and start stalling, standing still for too long, or delaying the delivery. (Timecodes for all heroes are below). #deadlock Abrams 00:00 - 00:32 Apollo 00:32 - 01:05 Bebop 01:05 - 01:42 Billy 01:42 - 02:07 Calico 02:07 - 02:39 Celeste 02:39 - 03:11 Paradox 03:11 - 03:50 Doorman 03:50 - 04:16 Drifter 04:16 - 04:49 Dynamo 04:49 - 05:22 Graves 05:22 - 05:55 Grey Talon 05:55 - 06:32 Haze 06:32 - 07:03 Holliday 07:03 - 07:40 Infernus 07:40 - 08:08 Ivy 08:08 - 08:34 Kelvin 08:34 - 09:09 Lady Geist 09:09 - 09:39 Lash 09:39 - 10:07 Sinclair 10:07 - 10:35 McGinnis 10:35 - 11:12 Mina 11:12 - 11:40 Mirage 11:40 - 12:08 Mo&Krill 12:08 - 12:32 Paige 12:32 - 12:58 Pocket 12:58 - 13:25 Rem 13:25 - 13:56 Seven 13:56 - 14:27 Shiv 14:27 - 15:01 Silver 15:01 - 15:39 Venator 15:39 - 16:10 Victor 16:10 - 16:42 Vindicta 16:42 - 17:21 Viscous 17:21 - 17:46 Vyper 17:46 - 18:18 Warden 18:18 - 18:47 Wraith 18:47 - 19:13 Yamato 19:13 - 19:47
19:59

Added 5 voice lines for each hero for situations when they are carrying the Spirit Jar and start stalling, standing still for too long, or delaying the delivery. (Timecodes for all heroes are below). #deadlock Abrams 00:00 - 00:32 Apollo 00:32 - 01:05 Bebop 01:05 - 01:42 Billy 01:42 - 02:07 Calico 02:07 - 02:39 Celeste 02:39 - 03:11 Paradox 03:11 - 03:50 Doorman 03:50 - 04:16 Drifter 04:16 - 04:49 Dynamo 04:49 - 05:22 Graves 05:22 - 05:55 Grey Talon 05:55 - 06:32 Haze 06:32 - 07:03 Holliday 07:03 - 07:40 Infernus 07:40 - 08:08 Ivy 08:08 - 08:34 Kelvin 08:34 - 09:09 Lady Geist 09:09 - 09:39 Lash 09:39 - 10:07 Sinclair 10:07 - 10:35 McGinnis 10:35 - 11:12 Mina 11:12 - 11:40 Mirage 11:40 - 12:08 Mo&Krill 12:08 - 12:32 Paige 12:32 - 12:58 Pocket 12:58 - 13:25 Rem 13:25 - 13:56 Seven 13:56 - 14:27 Shiv 14:27 - 15:01 Silver 15:01 - 15:39 Venator 15:39 - 16:10 Victor 16:10 - 16:42 Vindicta 16:42 - 17:21 Viscous 17:21 - 17:46 Vyper 17:46 - 18:18 Warden 18:18 - 18:47 Wraith 18:47 - 19:13 Yamato 19:13 - 19:47

Deadlock UA

179,352 görüntüleme • 3 ay önce

How GPT-5 thinks, with OpenAI VP of Research Jerry Tworek 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models
1:16:04

How GPT-5 thinks, with OpenAI VP of Research Jerry Tworek 00:00 - Intro 01:01 - What Reasoning Actually Means in AI 02:32 - Chain of Thought: Models Thinking in Words 05:25 - How Models Decide How Long to Think 07:24 - Evolution from o1 to o3 to GPT-5 11:00 - The Road to OpenAI: Growing up in Poland, Dropping out of School, Trading 20:32 - Working on Robotics and Rubik's Cube Solving 23:02 - A Day in the Life: Talking to Researchers 24:06 - How Research Priorities Are Determined 26:53 - OpenAI's Culture of Transparency 29:32 - Balancing Research with Shipping Fast 31:52 - Using OpenAI's Own Tools Daily 32:43 - Pre-Training Plus RL: The Modern AI Stack 35:10 - Reinforcement Learning 101: Training Dogs 40:17 - The Evolution of Deep Reinforcement Learning 42:09 - When GPT-4 Seemed Underwhelming at First 45:39 - How RLHF Made GPT-4 Actually Useful 48:02 - Unsupervised vs Supervised Learning 49:59 - GRPO and How DeepSeek Accelerated US Research 53:05 - What It Takes to Scale Reinforcement Learning 55:36 - Agentic AI and Long-Horizon Thinking 59:19 - Alignment as an RL Problem 1:01:11 - Winning ICPC World Finals Without Specific Training 1:05:53 - Applying RL Beyond Math and Coding 1:09:15 - The Path from Here to AGI 1:12:23 - Pure RL vs Language Models

Matt Turck

451,229 görüntüleme • 1 yıl önce

📁 2024 #수민 틱톡 연말결산 00:00 Poppin' Love 00:12 Two Of Hearts 00:22 Like I Do 00:37 Single Ladies 00:58 Everytime We Touch 01:14 Blue mario 01:25 Little Bit 01:48 썸 01:59 🐱🐾 02:10 전방향 미소녀 02:25 워싱씨 02:31 か わ い い 02:39 You Right 02:50 모시모시 03:06 걸프렌드 03:17 움파룸파 03:33 스키나코토 03:57 Huge 04:07 Grillz 04:22 Knock Knock 04:32 천장 챌린지 04:53 날 바라바라봐 05:15 아딱질 05:35 Magic Trick 05:41 YK 05:50 얼굴자랑 챌린지 06:03 👏 06:16 별빛이 내린다 06:35 ONE SPARK 07:10 Excuse me 07:17 Birthday 07:25 스노우볼 07:33 Nee 07:50 고민중독 08:21 시나모롤 08:31 포켓몬 08:44 차노을 09:00 Tell Me 09:09 시카노코 09:22 모찌 09:41 STAR 09:54 3,2,1 Go 10:11 Espresso 10:31 Apple 11:00 마루는 강쥐 11:14 White 11:37 Sticky
11:50

📁 2024 #수민 틱톡 연말결산 00:00 Poppin' Love 00:12 Two Of Hearts 00:22 Like I Do 00:37 Single Ladies 00:58 Everytime We Touch 01:14 Blue mario 01:25 Little Bit 01:48 썸 01:59 🐱🐾 02:10 전방향 미소녀 02:25 워싱씨 02:31 か わ い い 02:39 You Right 02:50 모시모시 03:06 걸프렌드 03:17 움파룸파 03:33 스키나코토 03:57 Huge 04:07 Grillz 04:22 Knock Knock 04:32 천장 챌린지 04:53 날 바라바라봐 05:15 아딱질 05:35 Magic Trick 05:41 YK 05:50 얼굴자랑 챌린지 06:03 👏 06:16 별빛이 내린다 06:35 ONE SPARK 07:10 Excuse me 07:17 Birthday 07:25 스노우볼 07:33 Nee 07:50 고민중독 08:21 시나모롤 08:31 포켓몬 08:44 차노을 09:00 Tell Me 09:09 시카노코 09:22 모찌 09:41 STAR 09:54 3,2,1 Go 10:11 Espresso 10:31 Apple 11:00 마루는 강쥐 11:14 White 11:37 Sticky

뱅글

14,611 görüntüleme • 1 yıl önce

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses. pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ). That loop is why Qwen quietly became the most downloaded open model family on Hugging Face. Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack. Watch and save it, then read the article below.
25:13

Alibaba engineer who leads Qwen explained the future of open agent models in 25 minutes - better than $2000 LLM training courses. pre-train the base ->SFT -> RLHF -> tool use -> multi-modal -> ship a whole family (chat / VL / coder / math / QwQ). That loop is why Qwen quietly became the most downloaded open model family on Hugging Face. Qwen base + Qwen-VL + Qwen-Coder + QwQ reasoning - that's the stack. Watch and save it, then read the article below.

h100envy

114,423 görüntüleme • 4 ay önce

i tested Qwen3.8 27B locally on my M4 Pro to see how it behaves as a coding agent inside my project showing step-by-step the full guide on how you can run it, perfect for your first local model 00:00 Why I’m testing Qwen3.8 00:49 What Qwen3.8 includes 00:59 Qwen3.8 Max vs 27B 01:14 Qwen3.8 Max explained 02:13 Qwen’s benchmark claims 02:31 Its autonomous coding project 03:18 Qwen3.8 27B explained 04:30 Why I’m testing it locally 05:05 Download Qwen 3.8 in LM Studio 06:00 Pick the Q4 model 06:41 Run it as a local server 07:33 How fast is it? 08:36 Set up OpenCode 09:33 Connect LM Studio to OpenCode 10:03 Add Qwen 3.8 to OpenCode 11:34 Match the context settings 15:00 Give Qwen a real coding task 15:49 A mistake I caught live: session compaction 18:00 Review the code 19:30 My verdict
21:02

i tested Qwen3.8 27B locally on my M4 Pro to see how it behaves as a coding agent inside my project showing step-by-step the full guide on how you can run it, perfect for your first local model 00:00 Why I’m testing Qwen3.8 00:49 What Qwen3.8 includes 00:59 Qwen3.8 Max vs 27B 01:14 Qwen3.8 Max explained 02:13 Qwen’s benchmark claims 02:31 Its autonomous coding project 03:18 Qwen3.8 27B explained 04:30 Why I’m testing it locally 05:05 Download Qwen 3.8 in LM Studio 06:00 Pick the Q4 model 06:41 Run it as a local server 07:33 How fast is it? 08:36 Set up OpenCode 09:33 Connect LM Studio to OpenCode 10:03 Add Qwen 3.8 to OpenCode 11:34 Match the context settings 15:00 Give Qwen a real coding task 15:49 A mistake I caught live: session compaction 18:00 Review the code 19:30 My verdict

Nidhi Singh

43,169 görüntüleme • 2 ay önce

좋아하는 #최정은 댄스커버&챌린지 모음 ❤︎ 00:00 pocket locket 00:17 Touch 00:39 마지막 축제 01:02 청바지 01:26 Confident 01:38 DENIAL IS A RIVER 01:51 Pink Hoodie 02:16 ExtraL 02:55 SIGN 03:15 Devil Game 03:30 1999 03:53 PUSH 2 START 04:09 In Bloom 04:31 Shake It To The Max 04:44 Gnarly 05:22 Imma Be 05:37 Lips Hips Kiss 06:08 HEYDAY 06:31 All The Stars 06:55 THUNDER 07:30 F Girl 07:52 소원을 말해봐 08:10 Soda Pop 08:25 Anthem 08:47 ICONIK 09:07 OVERDRIVE 09:38 BURNING UP 10:03 Body 10:38 CHANEL 10:53 Jumpshot 11:00 The Fate of Ophelia 11:15 WHERE YOU AT 11:41 Bad Desire
11:54

좋아하는 #최정은 댄스커버&챌린지 모음 ❤︎ 00:00 pocket locket 00:17 Touch 00:39 마지막 축제 01:02 청바지 01:26 Confident 01:38 DENIAL IS A RIVER 01:51 Pink Hoodie 02:16 ExtraL 02:55 SIGN 03:15 Devil Game 03:30 1999 03:53 PUSH 2 START 04:09 In Bloom 04:31 Shake It To The Max 04:44 Gnarly 05:22 Imma Be 05:37 Lips Hips Kiss 06:08 HEYDAY 06:31 All The Stars 06:55 THUNDER 07:30 F Girl 07:52 소원을 말해봐 08:10 Soda Pop 08:25 Anthem 08:47 ICONIK 09:07 OVERDRIVE 09:38 BURNING UP 10:03 Body 10:38 CHANEL 10:53 Jumpshot 11:00 The Fate of Ophelia 11:15 WHERE YOU AT 11:41 Bad Desire

찜

62,305 görüntüleme • 9 ay önce

Reasoning from scratch, round number 6! An introduction (and implementation) of Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO). 00:00 Introduction 01:54 What makes a reasoning model different? 04:25 Reasoning traces and model capability 08:29 Accuracy and format rewards 11:34 Aha moments and DeepSeek-R1 training 14:41 Reasoning effort and answer length 18:38 RLHF and RLVR 23:04 GRPO vs. PPO 26:40 GRPO explained with a cooking analogy 31:43 The KL term and simplified GRPO 35:04 Loading the pretrained model 36:07 Loading the MATH training data 39:26 Sampling model responses 46:30 Computing verifiable rewards 49:55 Computing advantages 51:54 Token and sequence log probabilities 55:29 Implementing sequence log probabilities 57:37 Fixing the inference-mode error 1:02:24 Computing the GRPO loss 1:04:37 Putting the GRPO step together 1:09:19 The GRPO training loop 1:12:57 Training settings, logging, and checkpoints 1:17:24 Running training and inspecting outputs 1:19:28 Loading and evaluating checkpoints 1:22:33 MATH-500 results and training stability 1:24:05 Memory requirements and next steps
1:26:53

Reasoning from scratch, round number 6! An introduction (and implementation) of Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO). 00:00 Introduction 01:54 What makes a reasoning model different? 04:25 Reasoning traces and model capability 08:29 Accuracy and format rewards 11:34 Aha moments and DeepSeek-R1 training 14:41 Reasoning effort and answer length 18:38 RLHF and RLVR 23:04 GRPO vs. PPO 26:40 GRPO explained with a cooking analogy 31:43 The KL term and simplified GRPO 35:04 Loading the pretrained model 36:07 Loading the MATH training data 39:26 Sampling model responses 46:30 Computing verifiable rewards 49:55 Computing advantages 51:54 Token and sequence log probabilities 55:29 Implementing sequence log probabilities 57:37 Fixing the inference-mode error 1:02:24 Computing the GRPO loss 1:04:37 Putting the GRPO step together 1:09:19 The GRPO training loop 1:12:57 Training settings, logging, and checkpoints 1:17:24 Running training and inspecting outputs 1:19:28 Loading and evaluating checkpoints 1:22:33 MATH-500 results and training stability 1:24:05 Memory requirements and next steps

Sebastian Raschka

75,256 görüntüleme • 29 gün önce

The vik episode on Ground Zero! [youtube link in replies] 0:00:00 - Vik's story and Moondream 0:18:50 - The core thesis of small and scale 0:28:35 - The data problem 0:33:35 - Deciding the training architecture 0:43:20 - Post-Training and RL perf 0:46:06 - Post-Training recipes of Moondream 3 0:47:40 - Open Source and VLM development priorities 0:52:05 - AI War : America and China 0:55:07 - Moondream acquisition and Future 1:04:08 - Community Questions 1:16:40 - Trivia, Lores and Opinions on state of AI 1:28:07 - Advice to 20yo
1:31:15

The vik episode on Ground Zero! [youtube link in replies] 0:00:00 - Vik's story and Moondream 0:18:50 - The core thesis of small and scale 0:28:35 - The data problem 0:33:35 - Deciding the training architecture 0:43:20 - Post-Training and RL perf 0:46:06 - Post-Training recipes of Moondream 3 0:47:40 - Open Source and VLM development priorities 0:52:05 - AI War : America and China 0:55:07 - Moondream acquisition and Future 1:04:08 - Community Questions 1:16:40 - Trivia, Lores and Opinions on state of AI 1:28:07 - Advice to 20yo

himanshu

43,246 görüntüleme • 10 ay önce

춤 챌린지 모음 #윈터 #WINTER 00:00 Kick Back 00:17 Weekend 00:38 Second 01:02 Sparkling 01:23 Sneakers 01:35 Candy 01:55 야야야 02:08 Unforgiven 02:25 Movie Star 02:34 이프푸 02:55 Hard 03:20 최애의 아이 03:37 Vujade 03:53 It is what it is 04:06 Get A Guitar 04:16 INVITATION 04:25 Guilty 05:10 Fact Check 05:34 Maniac 05:56 WOP 06:13 Chill Kill 06:34 Trance x I Know 06:45 Sweet Venom 07:03 Siren 07:16 Hot & Cold 07:41 Rockin' Around The Pole Again 08:03 Underneath the Tree 08:21 Be There For Me 08:42 뭣 같아 08:54 UNTOUCHABLE 09:09 Two Of Hearts 09:19 Boyfriend 09:42 Nee, 09:59 움파룸파 10:17 사랑스러워 10:38 Like I Do 10:53 Like I Do (골댕이ver.) 11:08 민정이가 저장해놓은.. 11:29 갈매기.. 11:45 날 바라바라봐 12:01 GGB 12:12 홀씨 12:56 시나모롤 13:10 띵띵땅땅 13:26 EENIE MEENIE 13:47 How Sweet
14:25

춤 챌린지 모음 #윈터 #WINTER 00:00 Kick Back 00:17 Weekend 00:38 Second 01:02 Sparkling 01:23 Sneakers 01:35 Candy 01:55 야야야 02:08 Unforgiven 02:25 Movie Star 02:34 이프푸 02:55 Hard 03:20 최애의 아이 03:37 Vujade 03:53 It is what it is 04:06 Get A Guitar 04:16 INVITATION 04:25 Guilty 05:10 Fact Check 05:34 Maniac 05:56 WOP 06:13 Chill Kill 06:34 Trance x I Know 06:45 Sweet Venom 07:03 Siren 07:16 Hot & Cold 07:41 Rockin' Around The Pole Again 08:03 Underneath the Tree 08:21 Be There For Me 08:42 뭣 같아 08:54 UNTOUCHABLE 09:09 Two Of Hearts 09:19 Boyfriend 09:42 Nee, 09:59 움파룸파 10:17 사랑스러워 10:38 Like I Do 10:53 Like I Do (골댕이ver.) 11:08 민정이가 저장해놓은.. 11:29 갈매기.. 11:45 날 바라바라봐 12:01 GGB 12:12 홀씨 12:56 시나모롤 13:10 띵띵땅땅 13:26 EENIE MEENIE 13:47 How Sweet

취향의 윈터

396,636 görüntüleme • 2 yıl önce

GPT-Image-2 was just released... It's the best image model (By a wide margin) Heres how to use GPT-Image-2 in ChatGPT, OpenAI Playground and of Course - Codex. TIMESTAMPS 00:00 Intro 01:27 Initial Tests of GPT-Image-2 02:24 It can GENERATE working BARCODES? 03:40 11 Edits one Prompt 06:40 Creating Cartoons 07:40 Overlay Explanations 08:50 How to Get Started 10:51 IPhone Mockup Images (Amazing) 15:43 Using GPT-Image in OpenAI Playground 16:11 Generate images in 4k 17:44 The FIRST limitation... Its bad at counting 18:26 Using GPT-Image in Codex (Agentic Image Gen)
22:26

GPT-Image-2 was just released... It's the best image model (By a wide margin) Heres how to use GPT-Image-2 in ChatGPT, OpenAI Playground and of Course - Codex. TIMESTAMPS 00:00 Intro 01:27 Initial Tests of GPT-Image-2 02:24 It can GENERATE working BARCODES? 03:40 11 Edits one Prompt 06:40 Creating Cartoons 07:40 Overlay Explanations 08:50 How to Get Started 10:51 IPhone Mockup Images (Amazing) 15:43 Using GPT-Image in OpenAI Playground 16:11 Generate images in 4k 17:44 The FIRST limitation... Its bad at counting 18:26 Using GPT-Image in Codex (Agentic Image Gen)

Riley Brown

35,856 görüntüleme • 6 ay önce

Introducing `gpt-oss Pro Mode`! Basically o3 Pro, but for the new OpenAI open source models :) Pro Mode chains together up to 10 instances of the new OpenAI GPT-OSS model, enabling it to produce a better answer than one instance alone could create! Open source, link here:
0:21

Introducing `gpt-oss Pro Mode`! Basically o3 Pro, but for the new OpenAI open source models :) Pro Mode chains together up to 10 instances of the new OpenAI GPT-OSS model, enabling it to produce a better answer than one instance alone could create! Open source, link here:

Matt Shumer

315,807 görüntüleme • 1 yıl önce

The Arcee AI Podcast is here! In this episode, Lucas Atkins and stochasm join us to discuss the story of Trinity models and everything frontier. I can say, this talk has been one of the most amazing and technical conversations we've had on Ground Zero. 0:00:00 - Intro 0:00:59 - Varun's transition from SWE to Pre-Training Lead 0:04:20 - Trinity Manifesto, Openclaw Ecosystem 0:12:15 - Arcee's Post-Training to Pre-Training Pivot 0:23:45 - Varun's first Pre-Training run (you can just do things!) 0:27:33 - Saturation in Pre-Training?, Mid-Training 0:37:00 - Tweaking the Training Architecture, Adam vs Muon, Evals 01:09:07 - Inference Engineering, Quick Fire, Post-Training Recipe 01:18:02 - Alpha in RL Envs, Harness Design 01:23:00 - American Open Source is trailing Chinese Competitors, Trinity Adoption 01:29:25 - Hiring at Arcee, Advice to 20yo
1:38:31

The Arcee AI Podcast is here! In this episode, Lucas Atkins and stochasm join us to discuss the story of Trinity models and everything frontier. I can say, this talk has been one of the most amazing and technical conversations we've had on Ground Zero. 0:00:00 - Intro 0:00:59 - Varun's transition from SWE to Pre-Training Lead 0:04:20 - Trinity Manifesto, Openclaw Ecosystem 0:12:15 - Arcee's Post-Training to Pre-Training Pivot 0:23:45 - Varun's first Pre-Training run (you can just do things!) 0:27:33 - Saturation in Pre-Training?, Mid-Training 0:37:00 - Tweaking the Training Architecture, Adam vs Muon, Evals 01:09:07 - Inference Engineering, Quick Fire, Post-Training Recipe 01:18:02 - Alpha in RL Envs, Harness Design 01:23:00 - American Open Source is trailing Chinese Competitors, Trinity Adoption 01:29:25 - Hiring at Arcee, Advice to 20yo

himanshu

32,531 görüntüleme • 6 ay önce

Reasoning From Scratch round 8, the finale! This last lecture in the series covers distillation as an efficient alternative to the previous reinforcement learning training runs. (Planning to take a brief break, but I will be back!) 00:00 Introduction and series recap 01:07 Why distill a reasoning model? 02:51 Soft vs. hard distillation 11:06 On-policy vs. off-policy distillation 18:09 Preparing data for reasoning distillation 22:35 Loading the DeepSeek R1 MATH dataset 28:04 Inspecting reasoning traces and final answers 31:27 Tokenization and chat templates 35:58 Pre-tokenizing the training dataset 39:01 Sequence lengths and the 2,048-token limit 42:38 Creating training and validation splits 45:32 Loading the Qwen3-0.6B base model 46:24 Computing answer-token log probabilities 54:10 Cross-entropy loss for distillation 57:25 Training and validation loss helpers 1:02:03 Implementing the distillation training loop 1:07:26 Running the standalone training script 1:09:41 Plotting training and validation losses 1:11:31 Evaluating checkpoints on MATH-500 1:14:09 Comparing teachers and distillation results 1:16:34 Future directions for reasoning models 1:19:41 Bonus exercises and next steps
1:21:17

Reasoning From Scratch round 8, the finale! This last lecture in the series covers distillation as an efficient alternative to the previous reinforcement learning training runs. (Planning to take a brief break, but I will be back!) 00:00 Introduction and series recap 01:07 Why distill a reasoning model? 02:51 Soft vs. hard distillation 11:06 On-policy vs. off-policy distillation 18:09 Preparing data for reasoning distillation 22:35 Loading the DeepSeek R1 MATH dataset 28:04 Inspecting reasoning traces and final answers 31:27 Tokenization and chat templates 35:58 Pre-tokenizing the training dataset 39:01 Sequence lengths and the 2,048-token limit 42:38 Creating training and validation splits 45:32 Loading the Qwen3-0.6B base model 46:24 Computing answer-token log probabilities 54:10 Cross-entropy loss for distillation 57:25 Training and validation loss helpers 1:02:03 Implementing the distillation training loop 1:07:26 Running the standalone training script 1:09:41 Plotting training and validation losses 1:11:31 Evaluating checkpoints on MATH-500 1:14:09 Comparing teachers and distillation results 1:16:34 Future directions for reasoning models 1:19:41 Bonus exercises and next steps

Sebastian Raschka

37,884 görüntüleme • 15 gün önce

SpaceX is about to launch their first V3 Starship and it’s by far the biggest and most radical change to the program to date. Here's a super quick overview of what all is new and different including the incredible new Raptor 3 engines, the new launch pad, and everything else that’s debuting on Flight 12. 00:00 - Intro 01:07 - Pad 2 03:15 - Raptor 3 04:59 - SuperHeavy V3 07:56 - Starship V3 10:52 - Flight 12 Profile
16:06

SpaceX is about to launch their first V3 Starship and it’s by far the biggest and most radical change to the program to date. Here's a super quick overview of what all is new and different including the incredible new Raptor 3 engines, the new launch pad, and everything else that’s debuting on Flight 12. 00:00 - Intro 01:07 - Pad 2 03:15 - Raptor 3 04:59 - SuperHeavy V3 07:56 - Starship V3 10:52 - Flight 12 Profile

Everyday Astronaut

252,140 görüntüleme • 5 ay önce

Why AI Progress Suddenly Feels Real - my conversation with Yann Dubois, who co-leads the Post-Training Frontiers team at OpenAI 00:00 - Intro 01:30 - Why recent AI progress feels like a step function 04:13 - Model reliability & the emotional rollercoaster of shipping GPT-5.5 07:33 - How OpenAI structures vertical and horizontal teams 09:49 - Improving model efficiency and test-time compute 12:32 - Yann's journey from Switzerland to OpenAI 15:37 - Reasoning in 2026: Real-world utility vs verifiable rewards 18:34 - GPT-5.5 Thinking vs Pro: Scaling test-time compute 20:09 - How reasoning models become more efficient 23:23 - Pre-training scaling and overcoming the data wall 27:03 - Multimodal data, synthetic data, and embodied AI 31:05 - Demystifying mid-training and post-training 37:21 - Does RL create new capabilities in AI? 38:53 - The challenges and frontier of scaling RL 43:09 - Is building AI models a craft or a strict science 48:21 - How AI models generalize across different domains 54:18 - How reinforcement learning cures AI hallucinations 56:04 - Negative generalization and conflicting instructions 58:05 - Can RL scale to law, medicine, and the broader economy? 1:00:19 - The evaluation bottleneck and Model as a Judge 1:04:21 - Continuous AI progress & continual learning 1:08:49 - Will foundation models eat the agent harness 1:11:23 - Why startups should focus on the last mile of AI
1:13:56

Why AI Progress Suddenly Feels Real - my conversation with Yann Dubois, who co-leads the Post-Training Frontiers team at OpenAI 00:00 - Intro 01:30 - Why recent AI progress feels like a step function 04:13 - Model reliability & the emotional rollercoaster of shipping GPT-5.5 07:33 - How OpenAI structures vertical and horizontal teams 09:49 - Improving model efficiency and test-time compute 12:32 - Yann's journey from Switzerland to OpenAI 15:37 - Reasoning in 2026: Real-world utility vs verifiable rewards 18:34 - GPT-5.5 Thinking vs Pro: Scaling test-time compute 20:09 - How reasoning models become more efficient 23:23 - Pre-training scaling and overcoming the data wall 27:03 - Multimodal data, synthetic data, and embodied AI 31:05 - Demystifying mid-training and post-training 37:21 - Does RL create new capabilities in AI? 38:53 - The challenges and frontier of scaling RL 43:09 - Is building AI models a craft or a strict science 48:21 - How AI models generalize across different domains 54:18 - How reinforcement learning cures AI hallucinations 56:04 - Negative generalization and conflicting instructions 58:05 - Can RL scale to law, medicine, and the broader economy? 1:00:19 - The evaluation bottleneck and Model as a Judge 1:04:21 - Continuous AI progress & continual learning 1:08:49 - Will foundation models eat the agent harness 1:11:23 - Why startups should focus on the last mile of AI

Matt Turck

101,646 görüntüleme • 5 ay önce

2025 연말결산 Full ver. 🌳🌸 #숩밤 #soogyu op- 0:00 01 - 0:19 02 - 4:44 03 - 19:37 04 - 33:22 05 - 43:40 06 - 52:58 07 - 01:12:34 08 - 01:22:32 09 - 01:41:04 10 - 01:55:14 11 - 02:05:22 12 - 02:12:52
2:21:06

2025 연말결산 Full ver. 🌳🌸 #숩밤 #soogyu op- 0:00 01 - 0:19 02 - 4:44 03 - 19:37 04 - 33:22 05 - 43:40 06 - 52:58 07 - 01:12:34 08 - 01:22:32 09 - 01:41:04 10 - 01:55:14 11 - 02:05:22 12 - 02:12:52

뚜뚜

97,129 görüntüleme • 10 ay önce

개인적으로 앙콘에서 듣고싶은 거 (머 아님 드림쇼5 라던가 . .) 00:00 같은 시간 같은 자리 00:41 사랑한단 뜻이야 01:01 Don’t Need Your Love 01:18 사랑이 좀 어려워 01:40 사랑은 또다시 01:58 Quiet Down 02:11 지금처럼만 02:46 Rainbow (책갈피) 03:17 우리의 계절 03:49 주인공 04:11 Countdown 04:36 Better Than Gold 05:11 Fire Alarm 05:39 마지막인사 06:05 Best Friend Ever 06:22 icantfeelanything 06:48 SOS 07:05 Stupid Cupid 07:33 i hate fruits 07:58 Flying Kiss 08:17 나의 소나기 08:39 항해 09:03 Dream team 09:27 Butterflies 09:57 Cold Coffee
10:20

개인적으로 앙콘에서 듣고싶은 거 (머 아님 드림쇼5 라던가 . .) 00:00 같은 시간 같은 자리 00:41 사랑한단 뜻이야 01:01 Don’t Need Your Love 01:18 사랑이 좀 어려워 01:40 사랑은 또다시 01:58 Quiet Down 02:11 지금처럼만 02:46 Rainbow (책갈피) 03:17 우리의 계절 03:49 주인공 04:11 Countdown 04:36 Better Than Gold 05:11 Fire Alarm 05:39 마지막인사 06:05 Best Friend Ever 06:22 icantfeelanything 06:48 SOS 07:05 Stupid Cupid 07:33 i hate fruits 07:58 Flying Kiss 08:17 나의 소나기 08:39 항해 09:03 Dream team 09:27 Butterflies 09:57 Cold Coffee

동담 #TASTE

68,349 görüntüleme • 9 ay önce