正在加载视频...

视频加载失败

加载此视频时出现问题。这可能是由于临时网络问题，或视频可能不可用。

Modern DRAM is based on a brilliant design from IBM. But, we're still paying for a latency penalty that's existed since the 60s! In this video, I'm introducing my research project (Tailslayer) that immensely reduces p99.99 latency on traditional RAM! By implementing a hedged read strategy taking advantage of... show more

LaurieWired

160,439 subscribers

864,333 次观看 • 5 个月前 •via X (Twitter)

教育科学技术

Anya Rossi• Live Now

Private livecam show

0 条评论

暂无评论

原始帖子的评论将显示在这里

相关视频

The Reality portal is here, and you can test it yourself by downloading Decart XR from SideQuest 👇 It works in 3 steps: 1- Decart #XR streams the #MetaQuest camera feed to its AI models over WebRTC 2- MirageLSD the world transformation model by Decart, transforms the video in real time based on the prompt, 3- The processed feed is sent back into Unity for in-headset display. There is some latency, and every few seconds the generation kind of resets as if a new seed is generated, but this is still something that blew my mind.

The Reality portal is here, and you can test it yourself by downloading Decart XR from SideQuest 👇 It works in 3 steps: 1- Decart #XR streams the #MetaQuest camera feed to its AI models over WebRTC 2- MirageLSD the world transformation model by Decart, transforms the video in real time based on the prompt, 3- The processed feed is sent back into Unity for in-headset display. There is some latency, and every few seconds the generation kind of resets as if a new seed is generated, but this is still something that blew my mind.

Gabriele Romagnoli

10,993 次观看 • 11 个月前

The value of the work we're doing at Optimum is encapsulated quite well by the phrase "speed is money". In modern markets there are real economic advantages to latency reduction. This is nothing new. Wall Street firms have long been optimizing on latency, primarily through colocation and top of the line hardware. However, when it comes to decentralized systems, expensive hardware and geographic concentration are antithetical to their purpose. Therefore we should optimize decentralized network latency through software, which I'm thrilled about because it's exactly what I've spent the better part of the past 2 decades working on with Random Linear Network Coding. Now let’s talk about networking economics, the relationship between speed and money. First, it's important to note that users will only pay for low latency if it can be consistently guaranteed. Second, you can only make that latency guarantee for a certain number of users. This is a universal law of networking. We can model this relationship on a delay curve, shown below. The delay curve is determined by the utilization rate of the network, meaning how much traffic is flowing through the network divided by the network's throughput. As you approach a level of traffic equal to the available throughput, latency trends infinitely higher. On this delay curve we can impose some utility thresholds. These thresholds are the levels of latency which are important to different groups of users because of how that latency guarantee improves their economic outcomes. Finding the point on the curve where each threshold intersects will tell us what level of traffic we can guarantee that level of latency for. Essentially, there exists a finite supply of speed on a network and the highest utility users of that speed are willing to pay more for it. I like to think of this similarly to expedited shipping options on Amazon. This is why we say speed is money, and why we can create a Latency Marketplace. The only way to increase the supply of speed is to fundamentally increase network throughput. This is what we work on at Optimum by using Random Linear Network Coding. The same relationship between traffic and throughput still applies, but now the delay curve is shifted out further to the right. Now more traffic can be processed at the same latency, or the same traffic can be processed at a lower latency. More speed available to the network. More value unlocked for the network’s users. Crucially, that value is no longer only reserved for those who can afford to sit closest to the machine. Expanding the supply of speed widens who can reach each latency threshold, keeping the network's advantage decentralized rather than concentrated in the hands of a few. When nodes join Optimum and participate, they reap the benefits, but they also add to the capacity. Rather than vying against each other in a zero-sum game, nodes help themselves and others.

The value of the work we're doing at Optimum is encapsulated quite well by the phrase "speed is money". In modern markets there are real economic advantages to latency reduction. This is nothing new. Wall Street firms have long been optimizing on latency, primarily through colocation and top of the line hardware. However, when it comes to decentralized systems, expensive hardware and geographic concentration are antithetical to their purpose. Therefore we should optimize decentralized network latency through software, which I'm thrilled about because it's exactly what I've spent the better part of the past 2 decades working on with Random Linear Network Coding. Now let’s talk about networking economics, the relationship between speed and money. First, it's important to note that users will only pay for low latency if it can be consistently guaranteed. Second, you can only make that latency guarantee for a certain number of users. This is a universal law of networking. We can model this relationship on a delay curve, shown below. The delay curve is determined by the utilization rate of the network, meaning how much traffic is flowing through the network divided by the network's throughput. As you approach a level of traffic equal to the available throughput, latency trends infinitely higher. On this delay curve we can impose some utility thresholds. These thresholds are the levels of latency which are important to different groups of users because of how that latency guarantee improves their economic outcomes. Finding the point on the curve where each threshold intersects will tell us what level of traffic we can guarantee that level of latency for. Essentially, there exists a finite supply of speed on a network and the highest utility users of that speed are willing to pay more for it. I like to think of this similarly to expedited shipping options on Amazon. This is why we say speed is money, and why we can create a Latency Marketplace. The only way to increase the supply of speed is to fundamentally increase network throughput. This is what we work on at Optimum by using Random Linear Network Coding. The same relationship between traffic and throughput still applies, but now the delay curve is shifted out further to the right. Now more traffic can be processed at the same latency, or the same traffic can be processed at a lower latency. More speed available to the network. More value unlocked for the network’s users. Crucially, that value is no longer only reserved for those who can afford to sit closest to the machine. Expanding the supply of speed widens who can reach each latency threshold, keeping the network's advantage decentralized rather than concentrated in the hands of a few. When nodes join Optimum and participate, they reap the benefits, but they also add to the capacity. Rather than vying against each other in a zero-sum game, nodes help themselves and others.

Muriel Medard

45,140 次观看 • 2 个月前

People are overlooking Google Gemini Realtime models for computer use It gave me sub 100ms latency with computer use. It has a much larger context window and is much cheaper as well Combine that with local OCR and local screen detection model based on Omniparser by Microsoft it works under 100ms action taking when combined with Cua I also put in a harness for Nous Research Hermes with it. You can access it all at your tip of your cursor. You can draw on your screen to give your agents a context And I am making it Open Source! Link in the Comments Sundar Pichai Min-Liang Tan Mojtaba Seyedhosseini

People are overlooking Google Gemini Realtime models for computer use It gave me sub 100ms latency with computer use. It has a much larger context window and is much cheaper as well Combine that with local OCR and local screen detection model based on Omniparser by Microsoft it works under 100ms action taking when combined with Cua I also put in a harness for Nous Research Hermes with it. You can access it all at your tip of your cursor. You can draw on your screen to give your agents a context And I am making it Open Source! Link in the Comments Sundar Pichai Min-Liang Tan Mojtaba Seyedhosseini

Milind S

38,361 次观看 • 3 个月前

I love the Arcade1Up Official cabinets that Trista has given me the last couple years. Yes, you can run free emulators on almost any device, but having the games running in a cabinet with arcade controls is a much better experience, even though it is just a packaged emulator. I was pretty decent back in the day, but after playing these for a while, I got farther than young-John ever did. Recently I played an original Joust machine at Cidercade and on my second game, I blew away my previous best score at home — 158k! The subtle control latency of the emulated experience versus the real thing matters! I measured the press-to-flap latency at home, and it looks like about 80ms. It isn’t blatantly obvious, but it shows up in the game feel and control error rate. I know there is a hard core community around emulator optimization, and with high refresh rate monitors it is possible to get objectively lower latency than the original CRT based hardware, but there is no reason the popular consumer versions can’t get most of the way there. This is probably just a matter of backing up a triple buffered swap chain or extra layers of image scaling / UI compositor getting in the way. Phase sync to the last quarter or so of the video interval and swap to the actual display should cut that latency in half. Doing the bit plane graphics and scaling directly to /dev/fb0 with software would be a guaranteed low latency path if you can get vsync timing. Trivia: The real Joust, and all the classic Williams games, didn’t even page flip, they just drew straight to the frame buffer, paying attention to the scan time.

I love the Arcade1Up Official cabinets that Trista has given me the last couple years. Yes, you can run free emulators on almost any device, but having the games running in a cabinet with arcade controls is a much better experience, even though it is just a packaged emulator. I was pretty decent back in the day, but after playing these for a while, I got farther than young-John ever did. Recently I played an original Joust machine at Cidercade and on my second game, I blew away my previous best score at home — 158k! The subtle control latency of the emulated experience versus the real thing matters! I measured the press-to-flap latency at home, and it looks like about 80ms. It isn’t blatantly obvious, but it shows up in the game feel and control error rate. I know there is a hard core community around emulator optimization, and with high refresh rate monitors it is possible to get objectively lower latency than the original CRT based hardware, but there is no reason the popular consumer versions can’t get most of the way there. This is probably just a matter of backing up a triple buffered swap chain or extra layers of image scaling / UI compositor getting in the way. Phase sync to the last quarter or so of the video interval and swap to the actual display should cut that latency in half. Doing the bit plane graphics and scaling directly to /dev/fb0 with software would be a guaranteed low latency path if you can get vsync timing. Trivia: The real Joust, and all the classic Williams games, didn’t even page flip, they just drew straight to the frame buffer, paying attention to the scan time.

John Carmack

61,855 次观看 • 1 年前

The lies recently told by scientists and bureaucrats about the likelihood of a LAB ORIGIN for Covid really got under my skin. They were publicly saying "not a chance!" while privately discussing that it was, in fact, the most likely probability based on the data they had. While not a surprise, and I always suspected as much, the (1) brazenness of it and (2) the fact that Congress isn't doing anything about it from a penalty standpoint explains why I lose it a bit in the video below (language warning). This is a piece I did for my subscribers a couple of days ago, but enough of them said "release it!" that I am doing so here. As always, I bring the receipts, this time in my farm wear...

The lies recently told by scientists and bureaucrats about the likelihood of a LAB ORIGIN for Covid really got under my skin. They were publicly saying "not a chance!" while privately discussing that it was, in fact, the most likely probability based on the data they had. While not a surprise, and I always suspected as much, the (1) brazenness of it and (2) the fact that Congress isn't doing anything about it from a penalty standpoint explains why I lose it a bit in the video below (language warning). This is a piece I did for my subscribers a couple of days ago, but enough of them said "release it!" that I am doing so here. As always, I bring the receipts, this time in my farm wear...

Chris Martenson

51,940 次观看 • 3 年前

If you haven't already, make sure to check out the latest full moon video for 'A Hard Lesson' over on YouTube. You can watch the full video here: "This is the oldest track on the project. It probably started in the late 80s or early 90s when I was in Senegal and it was known as Dakar 12 then." - pg

If you haven't already, make sure to check out the latest full moon video for 'A Hard Lesson' over on YouTube. You can watch the full video here: "This is the oldest track on the project. It probably started in the late 80s or early 90s when I was in Senegal and it was known as Dakar 12 then." - pg

Peter Gabriel

13,506 次观看 • 2 个月前

The great Joe Rogan heard about Hueyatlaco from Michael Button’s video! Here is my reaction as well as some footage I got when I visited the site in 2022. I wrote about it in chapter 12 of my book. I have a video on YouTube titled “Exposing the Biggest Cover Up in Archeology.” Hope you all get a chance to check out both Michael’s video and mine! Great job Michael!

The great Joe Rogan heard about Hueyatlaco from Michael Button’s video! Here is my reaction as well as some footage I got when I visited the site in 2022. I wrote about it in chapter 12 of my book. I have a video on YouTube titled “Exposing the Biggest Cover Up in Archeology.” Hope you all get a chance to check out both Michael’s video and mine! Great job Michael!

INCREDIBLE HISTORY

25,211 次观看 • 9 个月前

Neil thinks 90% of AI workloads will run in the background versus real time. In that future, latency matters less and cost matters more, and he's configured his company around a unique way of serving tokens at the cheapest possible price. "There's no bad chips. There's only bad pricing, and I will make any chip work at the right price. Let's talk about AMD, great chips overall. People don't understand how to program them very well. That's music to my ears. I'm happy for them to sleep on this chip and for me to buy as much as I can." "One of the ways I describe what we do is, we will buy any chip anywhere in the world for any duration of time. That is a level of flexibility and liquidity that no one else has right now. You'd have basically zero buyers for a data center that is ninety-five percent uptime. I'm that first buyer. First we scavenge chips, and then we scavenge power for those chips. The idea is, in both cases, I do not want to be bidding against Anthropic or OpenAI for compute capacity. I'm not gonna win against them, and I don't want to. I want to be more creative and use the supply that they don't find legible today. Over time I amass enough aggregate supply...and build my aggregate factory that is unbeatable in economics." "My whole goal is to so dramatically expand the supply of power across the United States that I have a home for a lot of chips that otherwise would not have earned their place in a data center." "We've really pushed AI to be an interactive chatbot tool, and everyone has chosen latency optimization because the shape of usage was chatbot oriented. That's the most profound change we're going to see in the next year, we're gonna move away from chatbots to more proactive or background agents, and in that world, it makes a lot more sense to build a stack around throughput." "I love this market because it's unbounded. There's no human in the loop, so you can consume as many tokens as you like in the background versus human attention span. So long term, we're going to end this year at maybe fifty-fifty background and real time workloads, but I see this going to ninety ten in favor of background." "The best latency is no latency at all. When you wake up in the morning, the work's already been done overnight, you didn't even have to ask for it. What we'd like is the agent to operate on more human time scales. You don't manage your colleagues every five minutes. You come back and check in maybe once a week." "My job is to make the tokens as cheap as humanly possible, and I will do it through every layer in the stack available to me."

Neil thinks 90% of AI workloads will run in the background versus real time. In that future, latency matters less and cost matters more, and he's configured his company around a unique way of serving tokens at the cheapest possible price. "There's no bad chips. There's only bad pricing, and I will make any chip work at the right price. Let's talk about AMD, great chips overall. People don't understand how to program them very well. That's music to my ears. I'm happy for them to sleep on this chip and for me to buy as much as I can." "One of the ways I describe what we do is, we will buy any chip anywhere in the world for any duration of time. That is a level of flexibility and liquidity that no one else has right now. You'd have basically zero buyers for a data center that is ninety-five percent uptime. I'm that first buyer. First we scavenge chips, and then we scavenge power for those chips. The idea is, in both cases, I do not want to be bidding against Anthropic or OpenAI for compute capacity. I'm not gonna win against them, and I don't want to. I want to be more creative and use the supply that they don't find legible today. Over time I amass enough aggregate supply...and build my aggregate factory that is unbeatable in economics." "My whole goal is to so dramatically expand the supply of power across the United States that I have a home for a lot of chips that otherwise would not have earned their place in a data center." "We've really pushed AI to be an interactive chatbot tool, and everyone has chosen latency optimization because the shape of usage was chatbot oriented. That's the most profound change we're going to see in the next year, we're gonna move away from chatbots to more proactive or background agents, and in that world, it makes a lot more sense to build a stack around throughput." "I love this market because it's unbounded. There's no human in the loop, so you can consume as many tokens as you like in the background versus human attention span. So long term, we're going to end this year at maybe fifty-fifty background and real time workloads, but I see this going to ninety ten in favor of background." "The best latency is no latency at all. When you wake up in the morning, the work's already been done overnight, you didn't even have to ask for it. What we'd like is the agent to operate on more human time scales. You don't manage your colleagues every five minutes. You come back and check in maybe once a week." "My job is to make the tokens as cheap as humanly possible, and I will do it through every layer in the stack available to me."

Patrick OShaughnessy

82,061 次观看 • 10 天前

Although you can't see much for the most part of this video, it is among the most intense combat videos I've seen so far. It shows a group of Ukrainians entering Andriivka, in the Bakhmut direction, firing at a Russian position, taking over it, and hiding in a basement. From the second half, you can mostly see a black screen but you can hear everything very well. The whole engagement takes mere minutes. I've done the best I could with subtitles. It is the 1st rifle battalion of the 93rd Brigade in action. Video published on MotoLife's YouTube channel:

Although you can't see much for the most part of this video, it is among the most intense combat videos I've seen so far. It shows a group of Ukrainians entering Andriivka, in the Bakhmut direction, firing at a Russian position, taking over it, and hiding in a basement. From the second half, you can mostly see a black screen but you can hear everything very well. The whole engagement takes mere minutes. I've done the best I could with subtitles. It is the 1st rifle battalion of the 93rd Brigade in action. Video published on MotoLife's YouTube channel:

WarTranslated

743,664 次观看 • 2 年前

The trailer for Project Ringfall is here! This rework is both a mod... and a journey I'd like you to join me on. My goal for Project Ringfall is to highlight what makes Halo Infinite unique and fun, while also making changes in areas often discussed here in the Halo community. This mod is for both Social and Competitive players! Along the way, I'll be talking about my own design philosophy, breaking down my approach to tuning weapons and player abilities, as well as sharing my design documents as I go! I appreciate every single one of you that have stuck with me as I've learned and grown as a designer. I really hope you'll enjoy this mod on it's release. I don't ever ask this, but liking and retweeting this post helps a lot. I'm putting my soul into this project, and if you like what I'm doing, the best way to support me is by sending a like and telling your friends.

The trailer for Project Ringfall is here! This rework is both a mod... and a journey I'd like you to join me on. My goal for Project Ringfall is to highlight what makes Halo Infinite unique and fun, while also making changes in areas often discussed here in the Halo community. This mod is for both Social and Competitive players! Along the way, I'll be talking about my own design philosophy, breaking down my approach to tuning weapons and player abilities, as well as sharing my design documents as I go! I appreciate every single one of you that have stuck with me as I've learned and grown as a designer. I really hope you'll enjoy this mod on it's release. I don't ever ask this, but liking and retweeting this post helps a lot. I'm putting my soul into this project, and if you like what I'm doing, the best way to support me is by sending a like and telling your friends.

ANCHr

59,749 次观看 • 2 年前

This is probably the most complex workflow I’ve ever built, only with open-source tools. It took my 4 days. It takes four inputs: author, title, and style; and generates a full visual animated story in one click in ComfyUI . I worked on it for four days. There are still some bugs, but here’s the first preview. Here’s a quick breakdown: - The four inputs are sent to LLMs with precise instructions to generate: first, prompts for images and image modifications; second, prompts for animations; third, prompts for generating music. - All voices are generated from the text and timed precisely, as they determine the length of each animation segment. - The first image and video are generated to serve as the title, but also as the guide for all other images created for the video. - Titles and subtitles are also added automatically in Comfy. - I also developed a lot of custom nodes for minor frame calculations, mostly to match audio and video. - The full system is a large loop that, for each line of text, generates an image and then a video from that image. The loop was the hardest part to build in this workflow, so it can process either a 20-second video or a 2-minute video with the same input. - There are multiple combinations of LLMs that try to understand the text in the best way to provide the best prompts for images and video. - The final video is assembled entirely within ComfyUI. - The music is generated based on the LLM output and matches the exact timing of the full animation. - Done! For reference, this workflow uses a lot of models and only works on an RTX 6000 Pro with plenty of RAM. My goal is not to replace humans, as I’ll try to explain later, this workflow is highly controlled and can be adapted or reworked at any point by real artists! My aim was to create a tool that can animate text in one go, allowing the AI some freedom while keeping a strict flow. I don’t know yet how I’ll share this workflow with people, I still need to polish it properly, but maybe through Patreon. Anyway, I hope you enjoy my research, and let’s always keep pushing further! :)

This is probably the most complex workflow I’ve ever built, only with open-source tools. It took my 4 days. It takes four inputs: author, title, and style; and generates a full visual animated story in one click in ComfyUI . I worked on it for four days. There are still some bugs, but here’s the first preview. Here’s a quick breakdown: - The four inputs are sent to LLMs with precise instructions to generate: first, prompts for images and image modifications; second, prompts for animations; third, prompts for generating music. - All voices are generated from the text and timed precisely, as they determine the length of each animation segment. - The first image and video are generated to serve as the title, but also as the guide for all other images created for the video. - Titles and subtitles are also added automatically in Comfy. - I also developed a lot of custom nodes for minor frame calculations, mostly to match audio and video. - The full system is a large loop that, for each line of text, generates an image and then a video from that image. The loop was the hardest part to build in this workflow, so it can process either a 20-second video or a 2-minute video with the same input. - There are multiple combinations of LLMs that try to understand the text in the best way to provide the best prompts for images and video. - The final video is assembled entirely within ComfyUI. - The music is generated based on the LLM output and matches the exact timing of the full animation. - Done! For reference, this workflow uses a lot of models and only works on an RTX 6000 Pro with plenty of RAM. My goal is not to replace humans, as I’ll try to explain later, this workflow is highly controlled and can be adapted or reworked at any point by real artists! My aim was to create a tool that can animate text in one go, allowing the AI some freedom while keeping a strict flow. I don’t know yet how I’ll share this workflow with people, I still need to polish it properly, but maybe through Patreon. Anyway, I hope you enjoy my research, and let’s always keep pushing further! :)

Lovis Odin

58,841 次观看 • 11 个月前

The scariest moment in algo trading is finding out your strategy works on random data too. neurotrader video shows the kill switch: "How I Develop Trading Strategies | Permutation Tests and Trading Strategy Development with Python." 21 minutes of pure framework. I wrote a companion deep-dive that takes this Monte Carlo foundation Extending it across every simulation method quant firms actually use From particle filters updating in real time on election night To copula models capturing tail dependence across correlated markets. Start with his video. Go deeper with the article.

The scariest moment in algo trading is finding out your strategy works on random data too. neurotrader video shows the kill switch: "How I Develop Trading Strategies | Permutation Tests and Trading Strategy Development with Python." 21 minutes of pure framework. I wrote a companion deep-dive that takes this Monte Carlo foundation Extending it across every simulation method quant firms actually use From particle filters updating in real time on election night To copula models capturing tail dependence across correlated markets. Start with his video. Go deeper with the article.

gemchanger

479,134 次观看 • 6 个月前

250330 Yuju's Ending Ment 🐶 when i'm on stage, i usually try to capture at least one more Buddy in my eyes. But this time, i think i tried to capture the members in my eyes a bit more. When i was dancing, i saw the members' backs, and their backs were still the same as the ones i saw 10 years ago. And i think it was the greatest luck in my life that i was so lucky to become a singer and got to sing GFriend's song. When we started this 10th anniversary project, i thought it would be a time to look back on the past, but as i did it, i think my thoughts changed a lil bit. I think while looking back on the past, it will also be a picture that continues to move forward. Don't remember it as a memory, but think of it nicely that we will continue to move forward together. And now, as i grow older, i've become more careful about promises, so i'm a bit careful saying things like this, but still, let's meet again someday. Thank you so much Buddy

250330 Yuju's Ending Ment 🐶 when i'm on stage, i usually try to capture at least one more Buddy in my eyes. But this time, i think i tried to capture the members in my eyes a bit more. When i was dancing, i saw the members' backs, and their backs were still the same as the ones i saw 10 years ago. And i think it was the greatest luck in my life that i was so lucky to become a singer and got to sing GFriend's song. When we started this 10th anniversary project, i thought it would be a time to look back on the past, but as i did it, i think my thoughts changed a lil bit. I think while looking back on the past, it will also be a picture that continues to move forward. Don't remember it as a memory, but think of it nicely that we will continue to move forward together. And now, as i grow older, i've become more careful about promises, so i'm a bit careful saying things like this, but still, let's meet again someday. Thank you so much Buddy

뷰뷰

15,248 次观看 • 1 年前

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

Introducing PhoneLLM, an open model for voice agents. GPT 5.6 Terra performance on typical voice agent tasks at 1/3 the latency and 1/18 the cost. For voice agents, we need models that are both very low latency and very good at tool calling and instruction following. There's a trade-off here, and we often have to compromise on either latency or capability when building voice agents. With PhoneLLM (and the training and data stack that made this model possible) we're fixing this problem. For the last couple of years, most of the effort in frontier model development has gone towards leveraging test-time compute. Which is awesome! Models of all shapes and sizes are available that perform really, really well ... if you have "thinking" turned on for your model. But if you need your agent to respond at voice conversation speed, you can't use thinking models. PhoneLLM is a full-weights fine-tune of NVIDIA Nemotron Nano 30B. We trained on a wide range of real-world telephone and customer support use cases. The training focused on taking the excellent Nano 30B base capabilities and teaching the model to do typical voice agent tasks with thinking disabled. The results are really good: accurate tool calling and concise, on-topic responses in long conversations. And fast: TTFAT measured server-side is <100ms if you run PhoneLLM on a lightly loaded B200. :-) But seriously, when we characterize model latency, we do it with full, end-to-end, batched request simulations using real Pipecat voice agent pipelines. You can serve more than 80 concurrent agents on a single B200 with P95 end-to-end TTFAT <600ms. Including network overhead. That's an LLM cost-per-minute around $0.0025. (1/4 of a cent.) At a latency lower than any third-party API offers today. More details about this model, including weights on Hugging Face, how to spin it up with one click on Modal, and a starter project repo you can clone, are in the thread ...

kwindla

324,083 次观看 • 8 天前

My Tesla model S was the victim of an attempted deering this morning. Had just left the grocery store and it was still dark ~0615hrs. I saw a slight flash out of the corner of my right eye and hit the brakes right as the collision alert started blaring. It doesn't look as close as I think it was but it felt like the left front of my car could have touched the tail had I been even a second slower in reacting. As an aside editing video on a desktop browser doesn't seem to let you clip things. Go to 0:50s or so in the below video if you want to see the attempted Tesla killer

My Tesla model S was the victim of an attempted deering this morning. Had just left the grocery store and it was still dark ~0615hrs. I saw a slight flash out of the corner of my right eye and hit the brakes right as the collision alert started blaring. It doesn't look as close as I think it was but it felt like the left front of my car could have touched the tail had I been even a second slower in reacting. As an aside editing video on a desktop browser doesn't seem to let you clip things. Go to 0:50s or so in the below video if you want to see the attempted Tesla killer

Ryan Scanlan 👥

120,099 次观看 • 2 年前

day 2 of building a self-driving power wheel today i officially trained a self driving model from scratch and deploy it on the car by just simply brute forcing everything, I: > made a remote tele-op and remote data collection app built on LiveKit infra > feat: 60ms e2e latency between the car and inference compute (car and compute in vietnam with singapore sfu) > feat: data is collected on operator side, baking latency into observation space itself (I expect this made the model more robust against latency) > recorded 30 min of data at 30fps and converted the dataset to lerobot (you can check a sample here) > trained a simple ACT model (3 epoch, batch size 8) to drive the car around my house > deployed the model on the car with remote inference the video explains everything shortly reflection: > the model is ofc bad, idt behavior cloning would work at all for such complex task on such small sample size > it did work in some cases where the observation is well within distribution, even generalizes to back the car when it gets stuck up next: > will hack alpamayo (NVIDIA) or comma ’s e2e to somehow fit this > or train with a llm backbone or a locomotion prior to see if it generalizes

day 2 of building a self-driving power wheel today i officially trained a self driving model from scratch and deploy it on the car by just simply brute forcing everything, I: > made a remote tele-op and remote data collection app built on LiveKit infra > feat: 60ms e2e latency between the car and inference compute (car and compute in vietnam with singapore sfu) > feat: data is collected on operator side, baking latency into observation space itself (I expect this made the model more robust against latency) > recorded 30 min of data at 30fps and converted the dataset to lerobot (you can check a sample here) > trained a simple ACT model (3 epoch, batch size 8) to drive the car around my house > deployed the model on the car with remote inference the video explains everything shortly reflection: > the model is ofc bad, idt behavior cloning would work at all for such complex task on such small sample size > it did work in some cases where the observation is well within distribution, even generalizes to back the car when it gets stuck up next: > will hack alpamayo (NVIDIA) or comma ’s e2e to somehow fit this > or train with a llm backbone or a locomotion prior to see if it generalizes

Binh Pham

19,404 次观看 • 5 个月前

This video is in response to the reactions I've copped from the post about NYE in Melbourne I put up on Facebook, I haven't copped as much over here. Apparently a few of you missed the point I was trying to make, if you still don't get where I'm coming from after watching this then whatever, I can't help you. Hopefully you'll get it one day. It is not wrong to stand up for your country and its people, the sooner people realise that, the better.

This video is in response to the reactions I've copped from the post about NYE in Melbourne I put up on Facebook, I haven't copped as much over here. Apparently a few of you missed the point I was trying to make, if you still don't get where I'm coming from after watching this then whatever, I can't help you. Hopefully you'll get it one day. It is not wrong to stand up for your country and its people, the sooner people realise that, the better.

Bec Freedom

44,331 次观看 • 1 年前