Loading video...

Video Failed to Load

There was a problem loading this video. This could be due to a temporary network issue or the video might be unavailable.

Reasoning LLMs generate very long chains-of-thought, so even small quantization errors add up. With AWQ, Qwen3-4B drops 71.0 → 68.2 on MMLU-Pro (~4% relative loss). 😬 ParoQuant fixes this! It keeps only the critical rotation pairs and fuses everything into a single kernel. Recovers most of the lost reasoning... show more

Zhijian Liu

6,263 subscribers

170,758 views • 3 months ago •via X (Twitter)

Education Science & Technology

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

Reasoning VLAs can think. They just can't think fast. Until now. Introducing FlashDrive⚡ 🚀 716 ms → 159 ms on RTX PRO 6000 (up to 5.7×) ✅ Zero accuracy loss FlashDrive = streaming inference + DFlash speculative reasoning + ParoQuant W4A8 Real-time reasoning for autonomous driving is here!

Reasoning VLAs can think. They just can't think fast. Until now. Introducing FlashDrive⚡ 🚀 716 ms → 159 ms on RTX PRO 6000 (up to 5.7×) ✅ Zero accuracy loss FlashDrive = streaming inference + DFlash speculative reasoning + ParoQuant W4A8 Real-time reasoning for autonomous driving is here!

Zhijian Liu

164,286 views • 1 month ago

Most search models need the cloud. II-Search-4B doesn’t. 4B model tuned for reasoning with search tools, built for local use. Performance of models 10x its size. Search that is small, smart, and open.

Most search models need the cloud. II-Search-4B doesn’t. 4B model tuned for reasoning with search tools, built for local use. Performance of models 10x its size. Search that is small, smart, and open.

Intelligent Internet

499,961 views • 10 months ago

Very Powerful but very simple Prompting technique. Simply ask the LLM to re-read the question - and this significantly boosts LLM reasoning across diverse tasks and model types. 💡 Repeats question input twice in prompt, unlocks latent reasoning potential **Original Problem** 🤔: Decoder-only LLMs with unidirectional attention struggle with nuanced reasoning tasks due to limited global understanding of input questions. **Key Insights from this Paper **💡: • Re-reading (RE2) input enhances reasoning by improving question comprehension • Enables "bidirectional" understanding in unidirectional LLMs • Compatible with existing thought-eliciting prompting methods • Effective across various LLM types and reasoning tasks

Very Powerful but very simple Prompting technique. Simply ask the LLM to re-read the question - and this significantly boosts LLM reasoning across diverse tasks and model types. 💡 Repeats question input twice in prompt, unlocks latent reasoning potential Original Problem 🤔: Decoder-only LLMs with unidirectional attention struggle with nuanced reasoning tasks due to limited global understanding of input questions. Key Insights from this Paper 💡: • Re-reading (RE2) input enhances reasoning by improving question comprehension • Enables "bidirectional" understanding in unidirectional LLMs • Compatible with existing thought-eliciting prompting methods • Effective across various LLM types and reasoning tasks

Rohan Paul

169,175 views • 1 year ago

ParoQuant just got a big upgrade 🚀 ✅ Supports the new Qwen3.5 models ⚡ Now runs on MLX (fast local inference on Apple Silicon) 🧠 Preserves reasoning quality with 4-bit quantization We also built an agent demo running locally on my 4-year-old M2 Max. Can't wait to upgrade to an M5 Max and see what kind of magic we can do. ✨

ParoQuant just got a big upgrade 🚀 ✅ Supports the new Qwen3.5 models ⚡ Now runs on MLX (fast local inference on Apple Silicon) 🧠 Preserves reasoning quality with 4-bit quantization We also built an agent demo running locally on my 4-year-old M2 Max. Can't wait to upgrade to an M5 Max and see what kind of magic we can do. ✨

Zhijian Liu

49,004 views • 3 months ago

🔥 Battle for the top reasoning LLM intensifies! The QwQ-32B-Preview is a very good reasoning LLM. Full video of my tests here: Summary of my findings and thoughts: It was able to solve a couple of hard math problems so it looks very promising for maths. It didn’t do so well on my coding task (generating bash script). By the results reported on the LiveCodeBench it has room for improvement. One thing that’s become very clear to me is that the reasoning capabilities of these LLMs are significantly closing the gap between the open and closed-sourced models. The competition is now going to be on a different level and it's going to be focused on which model produces the most efficient, optimized, accurate, and fastest reasoning steps beyond just accurate responses. That's what developers will care about. Traditional benchmarks are not going to be good enough for this. On that note, it's getting harder to assess these models, especially the consistency, efficiency, and quality of reasoning steps. After experimenting with this model, I realized that the reasoning paths are not fully optimized and there is a lot more optimization that needs to happen before these models are used in production settings. There might be a need to build some type of native and efficient self-assessment or self-reflection capability that prevents these reasoning LLMs to go in loops or produce unnecessary lengthy sequences. I also noticed that this model, at least from the HF demo, doesn’t separate the reasoning from the response. I think that actually hurts the performance of the model. On the other hand, o1 and R1 do that really well. In addition to that, I believe the training on reasoning is hurting the performance of the LLM in other areas such as helpfulness (check the code example in the video). Something that’s necessary at the moment is validating or evaluating the quality of the reasoning chains and figuring out a better strategy to optimize them. Current methods are probably not sufficient to solve this problem but that's where innovation will comes next. I recognize that this is a first effort so kudos to the Qwen team on this release. These issues highlight the importance of transparency with reasoning LLMs. We need to know how it was trained and with exact data or optimization strategy. Understanding that will enable researchers and developers to build better intuition and improve the reasoning capabilities and components at a faster rate. There is an opportunity for someone or a company to build a truly open-reasoning LLM. The race is on! I will continue to track the state-of-the-art in reasoning LLMs and report my takes and observations here. Stay tuned for more.

🔥 Battle for the top reasoning LLM intensifies! The QwQ-32B-Preview is a very good reasoning LLM. Full video of my tests here: Summary of my findings and thoughts: It was able to solve a couple of hard math problems so it looks very promising for maths. It didn’t do so well on my coding task (generating bash script). By the results reported on the LiveCodeBench it has room for improvement. One thing that’s become very clear to me is that the reasoning capabilities of these LLMs are significantly closing the gap between the open and closed-sourced models. The competition is now going to be on a different level and it's going to be focused on which model produces the most efficient, optimized, accurate, and fastest reasoning steps beyond just accurate responses. That's what developers will care about. Traditional benchmarks are not going to be good enough for this. On that note, it's getting harder to assess these models, especially the consistency, efficiency, and quality of reasoning steps. After experimenting with this model, I realized that the reasoning paths are not fully optimized and there is a lot more optimization that needs to happen before these models are used in production settings. There might be a need to build some type of native and efficient self-assessment or self-reflection capability that prevents these reasoning LLMs to go in loops or produce unnecessary lengthy sequences. I also noticed that this model, at least from the HF demo, doesn’t separate the reasoning from the response. I think that actually hurts the performance of the model. On the other hand, o1 and R1 do that really well. In addition to that, I believe the training on reasoning is hurting the performance of the LLM in other areas such as helpfulness (check the code example in the video). Something that’s necessary at the moment is validating or evaluating the quality of the reasoning chains and figuring out a better strategy to optimize them. Current methods are probably not sufficient to solve this problem but that's where innovation will comes next. I recognize that this is a first effort so kudos to the Qwen team on this release. These issues highlight the importance of transparency with reasoning LLMs. We need to know how it was trained and with exact data or optimization strategy. Understanding that will enable researchers and developers to build better intuition and improve the reasoning capabilities and components at a faster rate. There is an opportunity for someone or a company to build a truly open-reasoning LLM. The race is on! I will continue to track the state-of-the-art in reasoning LLMs and report my takes and observations here. Stay tuned for more.

elvis

14,740 views • 1 year ago

We are releasing Star Elastic - turn ONE reasoning LLM into MANY sizes with a single post-training run. 360× cheaper than pretraining a family of models. 7× better than SOTA compression. Split reasoning capability. Plus elastic budget control that beats the accuracy-latency frontier. Paper: HF models: Thread 👇

We are releasing Star Elastic - turn ONE reasoning LLM into MANY sizes with a single post-training run. 360× cheaper than pretraining a family of models. 7× better than SOTA compression. Split reasoning capability. Plus elastic budget control that beats the accuracy-latency frontier. Paper: HF models: Thread 👇

Pavlo Molchanov

21,518 views • 26 days ago

Apple's AI Reasoning Model Paper Explained Apple’s disruptive new AI reasoning model research paper, “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models” explained, and what it means for AGI.

Apple's AI Reasoning Model Paper Explained Apple’s disruptive new AI reasoning model research paper, “The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models” explained, and what it means for AGI.

Harper Carroll

37,956 views • 1 year ago

The new Qwen 3.5 by Qwen running on-device on iPhone 17 Pro. Qwen 3.5 beats models 4 times its size, has strong visual understanding, and can toggle reasoning on or off. The 2B 6-bit model here is running with MLX optimized for Apple Silicon.

The new Qwen 3.5 by Qwen running on-device on iPhone 17 Pro. Qwen 3.5 beats models 4 times its size, has strong visual understanding, and can toggle reasoning on or off. The 2B 6-bit model here is running with MLX optimized for Apple Silicon.

Adrien Grondin

3,551,792 views • 3 months ago

Multimodal reasoning with Phi-4-reasoning-vision, new work on scaling LLM inference, benchmarking AI agents in network operations, cinematic video generation, adaptive evaluation for LLMs, and using AI to improve individual and population health.

Multimodal reasoning with Phi-4-reasoning-vision, new work on scaling LLM inference, benchmarking AI agents in network operations, cinematic video generation, adaptive evaluation for LLMs, and using AI to improve individual and population health.

Microsoft Research

11,672 views • 3 months ago

Sentient took AI Startup of the Year at the Minsky Awards 🙌 ROMA Search achieved an industry-leading 81.7% accuracy, significantly surpassing competitors on the FRAMES benchmark for multi-step reasoning. On the SEAL-0 benchmark for complex, multi-source reasoning, ROMA with 45.6% accuracy more than doubled the score of Gemini 2.5 Pro (19.8%). ROMA is available on GitHub, try it out today👇

Sentient took AI Startup of the Year at the Minsky Awards 🙌 ROMA Search achieved an industry-leading 81.7% accuracy, significantly surpassing competitors on the FRAMES benchmark for multi-step reasoning. On the SEAL-0 benchmark for complex, multi-source reasoning, ROMA with 45.6% accuracy more than doubled the score of Gemini 2.5 Pro (19.8%). ROMA is available on GitHub, try it out today👇

Sentient

482,307 views • 8 months ago

Elon Musk said Tesla will add “reasoning” to FSD, likely by the end of the year. “With reasoning, it will literally think about which parking spot to pick. It’ll drop you off at the store entrance, then go find a spot. It will spot empty spaces much better than a human and use reasoning to solve problems.”

Elon Musk said Tesla will add “reasoning” to FSD, likely by the end of the year. “With reasoning, it will literally think about which parking spot to pick. It’ll drop you off at the store entrance, then go find a spot. It will spot empty spaces much better than a human and use reasoning to solve problems.”

Nic Cruz Patane

242,927 views • 6 months ago

A whole set of generation has lost the power of critical thinking and logical reasoning. What is the reason for this collective failure of natural selection?

A whole set of generation has lost the power of critical thinking and logical reasoning. What is the reason for this collective failure of natural selection?

Devi

427,885 views • 2 years ago

Which LLM reasons best when it doesn't have all the information? Enter LLM Poker Arena to find out. It's a Poker Playing benchmark where top reasoning models play Texas Hold'em poker against each other. Claude Opus 4.5, GPT-5.2, Gemini 2.5 Pro, and Grok 4 all sit at the same table and play full tournaments to see who finishes with the chips. Poker is very different when it comes to reasoning. It has to balance probabilistic reasoning, opponent modeling and make decisions under uncertainty. Poker is an interesting evaluation because it tests reasoning under incomplete information, something most coding benchmarks do not capture. In this tournaments the rules are: - Each LLM starts with $1,000 chips - Small and big blinds start at $25 / $50 - Blinds double every 3 minutes - All models run in their reasoning or thinking modes After the first 5 tournaments: - Claude Opus 4.5 with Thinking has 3 wins - GPT-5.2 has 2 wins - Grok 4 and Gemini 2.5 Pro have 0 wins Early results suggest Claude performs quite well at poker as well. Also five is a very small sample size. Planning to run many more tournaments, publish the full benchmark data and add a prediction market on top of it. Thanks for the suggestion clipz. Much more coming as part of Poker Cities !! This was built on Replit ⠕ using their AI integrations, which made it straightforward to connect Claude, GPT, and Gemini. What model do you think wins after 100 tournaments?

Which LLM reasons best when it doesn't have all the information? Enter LLM Poker Arena to find out. It's a Poker Playing benchmark where top reasoning models play Texas Hold'em poker against each other. Claude Opus 4.5, GPT-5.2, Gemini 2.5 Pro, and Grok 4 all sit at the same table and play full tournaments to see who finishes with the chips. Poker is very different when it comes to reasoning. It has to balance probabilistic reasoning, opponent modeling and make decisions under uncertainty. Poker is an interesting evaluation because it tests reasoning under incomplete information, something most coding benchmarks do not capture. In this tournaments the rules are: - Each LLM starts with $1,000 chips - Small and big blinds start at $25 / $50 - Blinds double every 3 minutes - All models run in their reasoning or thinking modes After the first 5 tournaments: - Claude Opus 4.5 with Thinking has 3 wins - GPT-5.2 has 2 wins - Grok 4 and Gemini 2.5 Pro have 0 wins Early results suggest Claude performs quite well at poker as well. Also five is a very small sample size. Planning to run many more tournaments, publish the full benchmark data and add a prediction market on top of it. Thanks for the suggestion clipz. Much more coming as part of Poker Cities !! This was built on Replit ⠕ using their AI integrations, which made it straightforward to connect Claude, GPT, and Gemini. What model do you think wins after 100 tournaments?

Anshul Dhawan

31,651 views • 4 months ago

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Introducing Reinforcement-Learned Teachers (RLTs): Transforming how we teach LLMs to reason with reinforcement learning (RL). Blog: Paper: Traditional RL focuses on “learning to solve” challenging problems with expensive LLMs and constitutes a key step in making student AI systems ultimately acquire reasoning capabilities via distillation and cold-starting. Enter our RLTs—a new class of models prompted with not only a problem’s question but also its solution, and directly trained to generate clear, step-by-step “explanations” to teach their students. Remarkably, an RLT with only 7B parameters produces superior results when distilling and cold-starting students in competitive and graduate-level reasoning tasks than orders-of-magnitude larger LLMs. RLTs are as effective even when distilling 32B students, much larger than the teacher itself—unlocking a new standard for efficiency in developing reasoning language models with RL. Code:

Sakana AI

179,210 views • 11 months ago

4. A search agent with multi-step trip planning and reasoning

4. A search agent with multi-step trip planning and reasoning

Chief AI Officer

56,086 views • 2 years ago

Alibaba’s Qwen 3.5 is now running fully on-device on the iPhone 17 Pro. It outperforms models 4x its size, delivers strong visual understanding, and lets you switch reasoning on or off. This demo uses the 2B 6-bit version, optimized with MLX for Apple Silicon.

Alibaba’s Qwen 3.5 is now running fully on-device on the iPhone 17 Pro. It outperforms models 4x its size, delivers strong visual understanding, and lets you switch reasoning on or off. This demo uses the 2B 6-bit version, optimized with MLX for Apple Silicon.

Hugging Models

202,648 views • 2 months ago

timelapse #85 (27.5 hrs): - currently cant rely on any other coding models except grok code fast 1 + grok 4 fast (for complex reasoning grok 4 fast is 20 cents for 1M tokens) - wrote qwen3-next trainer entirely from scratch to make it more managable - each piece completely done by grok-code-fast-1 in cursor as it seems to handle this task pretty well without the grok 4 fast reasoning - take on smaller problems and complete them quickly (makes it easier with 400 toks/sec over the api) - got distributed fp8 qwen3-next trainer running at 0.8 seconds per step on 8xH100s (still need to finish checkpoint loading logic) - perfect timing as the fp8 version of qwen3-next drops as im writing this - ill be in LA in 2 days (will visit SF mid way through as well) - 12.5% margarita - steak dinner with family - gained intuition on FlashAttention in very long context settings - caught up w/ Kearm h/eng and Arnie Ramesh

timelapse #85 (27.5 hrs): - currently cant rely on any other coding models except grok code fast 1 + grok 4 fast (for complex reasoning grok 4 fast is 20 cents for 1M tokens) - wrote qwen3-next trainer entirely from scratch to make it more managable - each piece completely done by grok-code-fast-1 in cursor as it seems to handle this task pretty well without the grok 4 fast reasoning - take on smaller problems and complete them quickly (makes it easier with 400 toks/sec over the api) - got distributed fp8 qwen3-next trainer running at 0.8 seconds per step on 8xH100s (still need to finish checkpoint loading logic) - perfect timing as the fp8 version of qwen3-next drops as im writing this - ill be in LA in 2 days (will visit SF mid way through as well) - 12.5% margarita - steak dinner with family - gained intuition on FlashAttention in very long context settings - caught up w/ Kearm h/eng and Arnie Ramesh

Elliot Arledge

283,820 views • 8 months ago

Turn any open-source LLM into reasoning powerhouse! Using reinforcement finetuning you can add reasoning abilities to any LLM, even without a labelled dataset. Step-by-step explanation with code:

Turn any open-source LLM into reasoning powerhouse! Using reinforcement finetuning you can add reasoning abilities to any LLM, even without a labelled dataset. Step-by-step explanation with code:

Akshay 🚀

50,423 views • 1 year ago

Luminal fuses entire models into a single GPU kernel, automatically. Let's talk about why this matters for inference at the speed of light:

Luminal fuses entire models into a single GPU kernel, automatically. Let's talk about why this matters for inference at the speed of light:

Joe Fioti

24,304 views • 4 months ago