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.

LLMs have a limited context window. When conversations grow too long this affects output quality, performance and cost. New blog post from Earendil engineer Vegard Stikbakke on how compaction addresses this and how we’ve implemented it in Pi. Read the full post below

Pi

55,703 subscribers

262,644 views • 1 month ago •via X (Twitter)

Science & Technology Education

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

A harness turns a model into an agent. At it’s core it provides 4 things: - a system prompt - tools - an agentic loop - a translation layer across models New blog post from Earendil co-founder colin hanna on what a harness is, and how you can own yours. Full post below

A harness turns a model into an agent. At it’s core it provides 4 things: - a system prompt - tools - an agentic loop - a translation layer across models New blog post from Earendil co-founder colin hanna on what a harness is, and how you can own yours. Full post below

Pi

424,548 views • 1 month ago

This is how easy you add tenting to your keyboard (Go60), while being also sleek, small and portable. A have a few more photos and videos about the tenting in my soon to published blog post.

This is how easy you add tenting to your keyboard (Go60), while being also sleek, small and portable. A have a few more photos and videos about the tenting in my soon to published blog post.

Fatih Arslan

39,267 views • 4 months ago

LongWriter Unleashing 10,000+ Word Generation from Long Context LLMs discuss: Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability.

LongWriter Unleashing 10,000+ Word Generation from Long Context LLMs discuss: Current long context large language models (LLMs) can process inputs up to 100,000 tokens, yet struggle to generate outputs exceeding even a modest length of 2,000 words. Through controlled experiments, we find that the model's effective generation length is inherently bounded by the sample it has seen during supervised fine-tuning (SFT). In other words, their output limitation is due to the scarcity of long-output examples in existing SFT datasets. To address this, we introduce AgentWrite, an agent-based pipeline that decomposes ultra-long generation tasks into subtasks, enabling off-the-shelf LLMs to generate coherent outputs exceeding 20,000 words. Leveraging AgentWrite, we construct LongWriter-6k, a dataset containing 6,000 SFT data with output lengths ranging from 2k to 32k words. By incorporating this dataset into model training, we successfully scale the output length of existing models to over 10,000 words while maintaining output quality. We also develop LongBench-Write, a comprehensive benchmark for evaluating ultra-long generation capabilities. Our 9B parameter model, further improved through DPO, achieves state-of-the-art performance on this benchmark, surpassing even much larger proprietary models. In general, our work demonstrates that existing long context LLM already possesses the potential for a larger output window--all you need is data with extended output during model alignment to unlock this capability.

AK

50,995 views • 2 years ago

As promised! New blog with Josh Newton combining our talks from AI Engineer's first design track. In addition, we wanted to actually show a demonstration of how you can iterate and improve upon that first design output from AI. We had a ton of fun building the interactions for this post (especially the horse animator hahaha). Would love to hear your thoughts!

As promised! New blog with Josh Newton combining our talks from AI Engineer's first design track. In addition, we wanted to actually show a demonstration of how you can iterate and improve upon that first design output from AI. We had a ton of fun building the interactions for this post (especially the horse animator hahaha). Would love to hear your thoughts!

Sabina

13,570 views • 2 months ago

🔮 This limited-edition Nyanners case could be yours! 🔮 Its one of our most popular designs to date, and this one needs a brand-new home. Want to take it home? Here's how to enter: -Follow us (Ironside Computers ) -Like this post -Tag a Friend in the comments below Best of luck to everyone entering! ✨️

Ironside Computers

43,980 views • 1 month ago

In the files of “don’t be this guy” Perfect example of the problem with so many lifters… Ego lifting and caught up in how much weight they have on the bar or machine These are all shitty half ass reps. How about dropping the weight in half and actually performing a full rep below parallel? This needs to be addressed. See this way too often. And don’t comment that I shouldn’t post this. I really don’t care. Same guy who screams and grunts super load on every rep.

In the files of “don’t be this guy” Perfect example of the problem with so many lifters… Ego lifting and caught up in how much weight they have on the bar or machine These are all shitty half ass reps. How about dropping the weight in half and actually performing a full rep below parallel? This needs to be addressed. See this way too often. And don’t comment that I shouldn’t post this. I really don’t care. Same guy who screams and grunts super load on every rep.

BowTiedMeatHead 🥩💪

28,271 views • 6 months ago

Aave Glass, our take on Apple's Liquid Glass, now works across mobile and web. We've published a post that breaks down how we built Aave Glass and the family of components it now powers. Read it below.

Aave Glass, our take on Apple's Liquid Glass, now works across mobile and web. We've published a post that breaks down how we built Aave Glass and the family of components it now powers. Read it below.

Aave

92,620 views • 4 months ago

🍇 This limited-edition Grape Juicebox case could be yours!🍇 It’s hands-down one of our absolute favorite Juicebox designs to date, and this particular one needs a brand-new home.💜 Want to take it home? Here’s how to enter: - Follow us - Like this post - Tag a friend in the comments below Best of luck to everyone entering!✨️

Ironside Computers

121,721 views • 2 months ago

The bold part of this variable font was generated with my custom AI harness, img2bez, and a small neural network I trained from scratch. I will have a blog post out Monday explaining the full workflow.

The bold part of this variable font was generated with my custom AI harness, img2bez, and a small neural network I trained from scratch. I will have a blog post out Monday explaining the full workflow.

Eli Heuer

28,910 views • 2 months ago

Wow. This is one of the best interactive sites I’ve seen for learning how LLMs work! 🔥 It starts w/ a clear intro and guides you through every core component: from Embedding, Layer Norm, and Self-Attention to MLPs, Transformer blocks, Softmax, and Output layers. link in 🧵↓

Wow. This is one of the best interactive sites I’ve seen for learning how LLMs work! 🔥 It starts w/ a clear intro and guides you through every core component: from Embedding, Layer Norm, and Self-Attention to MLPs, Transformer blocks, Softmax, and Output layers. link in 🧵↓

Charly Wargnier

52,092 views • 1 year ago

You have to be hurting on a whole new level to make a post like this and actually post it 🤣🤣🤣🤣 Big man’s teams in turmoil and he’s on TikTok moaning about what we call DEL

You have to be hurting on a whole new level to make a post like this and actually post it 🤣🤣🤣🤣 Big man’s teams in turmoil and he’s on TikTok moaning about what we call DEL

IbroxRoar

46,171 views • 14 days ago

I like how post Astartes and especially post Secret Level that the collective idea of how Space Marines fight has changed. I just really like this style of it. Precise calculated sudden movements, almost like a machine in how efficient every action is.

I like how post Astartes and especially post Secret Level that the collective idea of how Space Marines fight has changed. I just really like this style of it. Precise calculated sudden movements, almost like a machine in how efficient every action is.

Rylanor, The Ancient of Rites

84,324 views • 9 months ago

GEMMA 4 26B ON AN RTX 4060 WITH A 248K TOKEN CONTEXT WINDOW 20 tokens per second and a context window so large you can feed it entire codebases, books and research papers in a single prompt this is not a cloud api and not a server rack, this is a regular consumer gpu running locally with llama.cpp and q4_k_xl quantization 248k context on an 8gb vram card was not supposed to be possible and here it is just running on someone’s desk the article below covers exactly which tools and configs make this kind of setup work in 2026 ↓

GEMMA 4 26B ON AN RTX 4060 WITH A 248K TOKEN CONTEXT WINDOW 20 tokens per second and a context window so large you can feed it entire codebases, books and research papers in a single prompt this is not a cloud api and not a server rack, this is a regular consumer gpu running locally with llama.cpp and q4_k_xl quantization 248k context on an 8gb vram card was not supposed to be possible and here it is just running on someone’s desk the article below covers exactly which tools and configs make this kind of setup work in 2026 ↓

leopardracer

56,865 views • 3 months ago

Insane Viper ultimate for A site post plant on the new map Summit Do it this way to get a pocket to outplay the enemy team 😎 Crazy how it covers the whole site and even more! #VALORANT

Insane Viper ultimate for A site post plant on the new map Summit Do it this way to get a pocket to outplay the enemy team 😎 Crazy how it covers the whole site and even more! #VALORANT

Tseeky - Pro Valorant Lineups

26,985 views • 3 months ago

So many of you asking me what your first post should be ???? … how about we start the “ tongue 👅 challenge” .. everyone just go post this .. say your name and what you do and then let your tongue 😜 goooooo” everyone is overthinking shit 💩 .. listen to the second video on this post to understand and understand the 80/20 rule of content …too many fear a post not doing well, or the judgement in the comments .. we have to stop that ! #firstpost #inspiration #garyvee

So many of you asking me what your first post should be ???? … how about we start the “ tongue 👅 challenge” .. everyone just go post this .. say your name and what you do and then let your tongue 😜 goooooo” everyone is overthinking shit 💩 .. listen to the second video on this post to understand and understand the 80/20 rule of content …too many fear a post not doing well, or the judgement in the comments .. we have to stop that ! #firstpost #inspiration #garyvee

Gary Vaynerchuk

15,775 views • 1 year ago

this is the kind of video modern day robin buckley would post on social media. nancy comments on how cool it looks, how smart she thinks robin is and then has a gay meltdown in the privacy of her room.

this is the kind of video modern day robin buckley would post on social media. nancy comments on how cool it looks, how smart she thinks robin is and then has a gay meltdown in the privacy of her room.

joey ★ is seeing djo

89,431 views • 7 months ago

Today, two of us who have built billion dollar companies are launching the ULTIMATE guide for founders on how to start a lean, AI-native startup in 2025. This covers every single tool Founders should use when building their startup from the ground up, from legal, finance, payroll, sales, tech and more. My coauthor Henry has built Super and I have built Glean into both $100M+ businesses and we've invested in 100+ startups between us in the last few years, and this is a constant question we get. And it's completely FREE. The article drops midnight tonight. Bookmark this post. Link below:

Today, two of us who have built billion dollar companies are launching the ULTIMATE guide for founders on how to start a lean, AI-native startup in 2025. This covers every single tool Founders should use when building their startup from the ground up, from legal, finance, payroll, sales, tech and more. My coauthor Henry has built Super and I have built Glean into both $100M+ businesses and we've invested in 100+ startups between us in the last few years, and this is a constant question we get. And it's completely FREE. The article drops midnight tonight. Bookmark this post. Link below:

Deedy

323,554 views • 1 year ago