Yannic Kilcher Videos (Audio Only‪)‬ Yannic Kilcher

#llm #ai #chatgpt

How does one run inference for a generative autoregressive language model that has been trained with a fixed context size? Streaming LLMs combine the performance of windowed attention, but avoid the drop in performance by using attention sinks - an interesting phenomenon where the token at position 0 acts as an absorber of "extra" attention.

OUTLINE:
0:00 - Introduction
1:20 - What is the problem?
10:30 - The hypothesis: Attention Sinks
15:10 - Experimental evidence
18:45 - Streaming LLMs
20:45 - Semantics or position?
22:30 - Can attention sinks be learned?
27:45 - More experiments
30:10 - Comparison to Big Bird


Paper: https://arxiv.org/abs/2309.17453

Abstract:
Deploying Large Language Models (LLMs) in streaming applications such as multi-round dialogue, where long interactions are expected, is urgently needed but poses two major challenges. Firstly, during the decoding stage, caching previous tokens' Key and Value states (KV) consumes extensive memory. Secondly, popular LLMs cannot generalize to longer texts than the training sequence length. Window attention, where only the most recent KVs are cached, is a natural approach -- but we show that it fails when the text length surpasses the cache size. We observe an interesting phenomenon, namely attention sink, that keeping the KV of initial tokens will largely recover the performance of window attention. In this paper, we first demonstrate that the emergence of attention sink is due to the strong attention scores towards initial tokens as a ``sink'' even if they are not semantically important. Based on the above analysis, we introduce StreamingLLM, an efficient framework that enables LLMs trained with a finite length attention window to generalize to infinite sequence lengths without any fine-tuning. We show that StreamingLLM can enable Llama-2, MPT, Falcon, and Pythia to perform stable and efficient language modeling with up to 4 million tokens and more. In addition, we discover that adding a placeholder token as a dedicated attention sink during pre-training can further improve streaming deployment. In streaming settings, StreamingLLM outperforms the sliding window recomputation baseline by up to 22.2x speedup. Code and datasets are provided at this https URL.

Authors: Guangxuan Xiao, Yuandong Tian, Beidi Chen, Song Han, Mike Lewis

Links:
Homepage: https://ykilcher.com
Merch: https://ykilcher.com/merch
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://ykilcher.com/discord
LinkedIn: https://www.linkedin.com/in/ykilcher

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

#ai #promptengineering #evolution

Promptbreeder is a self-improving self-referential system for automated prompt engineering. Give it a task description and a dataset, and it will automatically come up with appropriate prompts for the task. This is achieved by an evolutionary algorithm where not only the prompts, but also the mutation-prompts are improved over time in a population-based, diversity-focused approach.

OUTLINE:
0:00 - Introduction
2:10 - From manual to automated prompt engineering
10:40 - How does Promptbreeder work?
21:30 - Mutation operators
36:00 - Experimental Results
38:05 - A walk through the appendix

Paper: https://arxiv.org/abs/2309.16797

Abstract:
Popular prompt strategies like Chain-of-Thought Prompting can dramatically improve the reasoning abilities of Large Language Models (LLMs) in various domains. However, such hand-crafted prompt-strategies are often sub-optimal. In this paper, we present Promptbreeder, a general-purpose self-referential self-improvement mechanism that evolves and adapts prompts for a given domain. Driven by an LLM, Promptbreeder mutates a population of task-prompts, and subsequently evaluates them for fitness on a training set. Crucially, the mutation of these task-prompts is governed by mutation-prompts that the LLM generates and improves throughout evolution in a self-referential way. That is, Promptbreeder is not just improving task-prompts, but it is also improving the mutationprompts that improve these task-prompts. Promptbreeder outperforms state-of-the-art prompt strategies such as Chain-of-Thought and Plan-and-Solve Prompting on commonly used arithmetic and commonsense reasoning benchmarks. Furthermore, Promptbreeder is able to evolve intricate task-prompts for the challenging problem of hate speech classification.

Authors: Chrisantha Fernando, Dylan Banarse, Henryk Michalewski, Simon Osindero, Tim Rocktäschel

Links:
Homepage: https://ykilcher.com
Merch: https://ykilcher.com/merch
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://ykilcher.com/discord
LinkedIn: https://www.linkedin.com/in/ykilcher

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

#ai #retnet #transformers

Retention is an alternative to Attention in Transformers that can both be written in a parallel and in a recurrent fashion. This means the architecture achieves training parallelism while maintaining low-cost inference. Experiments in the paper look very promising.

OUTLINE:
0:00 - Intro
2:40 - The impossible triangle
6:55 - Parallel vs sequential
15:35 - Retention mechanism
21:00 - Chunkwise and multi-scale retention
24:10 - Comparison to other architectures
26:30 - Experimental evaluation

Paper: https://arxiv.org/abs/2307.08621

Abstract:
In this work, we propose Retentive Network (RetNet) as a foundation architecture for large language models, simultaneously achieving training parallelism, low-cost inference, and good performance. We theoretically derive the connection between recurrence and attention. Then we propose the retention mechanism for sequence modeling, which supports three computation paradigms, i.e., parallel, recurrent, and chunkwise recurrent. Specifically, the parallel representation allows for training parallelism. The recurrent representation enables low-cost O(1) inference, which improves decoding throughput, latency, and GPU memory without sacrificing performance. The chunkwise recurrent representation facilitates efficient long-sequence modeling with linear complexity, where each chunk is encoded parallelly while recurrently summarizing the chunks. Experimental results on language modeling show that RetNet achieves favorable scaling results, parallel training, low-cost deployment, and efficient inference. The intriguing properties make RetNet a strong successor to Transformer for large language models. Code will be available at this https URL.

Authors: Yutao Sun, Li Dong, Shaohan Huang, Shuming Ma, Yuqing Xia, Jilong Xue, Jianyong Wang, Furu Wei


Links:
Homepage: https://ykilcher.com
Merch: https://ykilcher.com/merch
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://ykilcher.com/discord
LinkedIn: https://www.linkedin.com/in/ykilcher

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

#ai #rlhf #llm

ReST uses a bootsrap-like method to produce its own extended dataset and trains on ever higher-quality subsets of it to improve its own reward. The method allows for re-using the same generated data multiple times and thus has an efficiency advantage with respect to Online RL techniques like PPO.

Paper: https://arxiv.org/abs/2308.08998

Abstract:
Reinforcement learning from human feedback (RLHF) can improve the quality of large language model's (LLM) outputs by aligning them with human preferences. We propose a simple algorithm for aligning LLMs with human preferences inspired by growing batch reinforcement learning (RL), which we call Reinforced Self-Training (ReST). Given an initial LLM policy, ReST produces a dataset by generating samples from the policy, which are then used to improve the LLM policy using offline RL algorithms. ReST is more efficient than typical online RLHF methods because the training dataset is produced offline, which allows data reuse. While ReST is a general approach applicable to all generative learning settings, we focus on its application to machine translation. Our results show that ReST can substantially improve translation quality, as measured by automated metrics and human evaluation on machine translation benchmarks in a compute and sample-efficient manner.

Authors: Caglar Gulcehre, Tom Le Paine, Srivatsan Srinivasan, Ksenia Konyushkova, Lotte Weerts, Abhishek Sharma, Aditya Siddhant, Alex Ahern, Miaosen Wang, Chenjie Gu, Wolfgang Macherey, Arnaud Doucet, Orhan Firat, Nando de Freitas

Links:
Homepage: https://ykilcher.com
Merch: https://ykilcher.com/merch
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://ykilcher.com/discord
LinkedIn: https://www.linkedin.com/in/ykilcher

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n

#mlnews #llama2 #openai

Your regular irregular update on the world of Machine Learning.

References:
https://twitter.com/ylecun/status/1681336284453781505
https://ai.meta.com/llama/
https://about.fb.com/news/2023/07/llama-2-statement-of-support/
https://247wallst.com/special-report/2023/08/12/this-is-the-biggest-social-media-platform-ranking-the-worlds-largest-networking-sites/4/
https://github.com/Alpha-VLLM/LLaMA2-Accessory
https://together.ai/blog/llama-2-7b-32k?s=09&utm_source=pocket_saves
https://github.com/imoneoi/openchat
https://twitter.com/lmsysorg/status/1686794639469371393?s=09&t=sS3awkbavmSMSmwp64Ef4A&utm_source=pocket_saves
https://huggingface.co/lmsys/vicuna-13b-v1.5-16k
https://blog.google/outreach-initiatives/public-policy/google-microsoft-openai-anthropic-frontier-model-forum/
https://www.earthdata.nasa.gov/news/impact-ibm-hls-foundation-model?utm_source=pocket_reader
https://huggingface.co/ibm-nasa-geospatial/Prithvi-100M
https://ai.meta.com/blog/generative-ai-text-images-cm3leon/
https://www.deepmind.com/blog/rt-2-new-model-translates-vision-and-language-into-action?utm_source=twitter&utm_medium=social&utm_campaign=rt2
https://arxiv.org/abs/2307.14334
https://sites.research.google/med-palm/
https://open-catalyst.metademolab.com/?utm_source=twitter&utm_medium=organic_social&utm_campaign=opencatalyst&utm_content=card
https://open-catalyst.metademolab.com/demo
https://www.anthropic.com/index/claude-2?utm_source=pocket_reader
https://claude.ai/login
https://audiocraft.metademolab.com/?utm_source=pocket_saves
https://venturebeat.com/programming-development/stability-ai-launches-stablecode-an-llm-for-code-generation/
https://stability.ai/blog/stablecode-llm-generative-ai-coding
https://twitter.com/JeffDean/status/1686806525862608896?s=09&t=LG2z9ok9QExHbSy0fvBsxA&utm_source=pocket_saves
https://sites.research.google/open-buildings/
https://twitter.com/deliprao/status/1687283117873106946?s=09&t=1NmC-B55Z8IuF_HTuGOo7w&utm_source=pocket_saves
https://arxiv.org/pdf/2308.01320.pdf
https://twitter.com/javilopen/status/1687795349719547905?utm_source=pocket_saves
https://research.nvidia.com/labs/par/Perfusion/
https://ar5iv.labs.arxiv.org/html/2307.14936
https://www.linkedin.com/feed/update/urn:li:activity:7093463974750371840/?utm_source=pocket_saves
https://huggingface.co/syzymon/long_llama_3b_instruct
https://arxiv.org/abs/2307.03170
https://dynalang.github.io/
https://github.com/mlfoundations/open_flamingo
https://twitter.com/akshay_pachaar/status/1687079353937698816?s=09&t=fos8QSCsGEEM6dMflhq0Mg&utm_source=pocket_saves
https://github.com/OpenBMB/ToolBench
https://llm-attacks.org/
https://arstechnica.com/information-technology/2023/07/openai-discontinues-its-ai-writing-detector-due-to-low-rate-of-accuracy/
https://sites.google.com/view/steve-1
https://github.com/Shalev-Lifshitz/STEVE-1
https://erichartford.com/dolphin
https://huggingface.co/ehartford/dolphin-llama-13b
https://www.mosaicml.com/blog/long-context-mpt-7b-8k
https://twitter.com/camenduru/status/1688045780244848640?s=09&t=ubJ2Qtz-TG6Xo3_GMtt2Cw&utm_source=pocket_saves
https://github.com/IDEA-Research/DWPose
https://twitter.com/tri_dao/status/1680987577913065472?s=09&t=Q181vFmM6d3nDq-5BwfDeg&utm_source=pocket_saves
https://tridao.me/publications/flash2/flash2.pdf
https://thehackernews.com/2023/07/wormgpt-new-ai-tool-allows.html
https://www.tomshardware.com/news/ai-steals-data-with-keystroke-audio
https://arxiv.org/pdf/2308.01074.pdf
https://www.foxnews.com/politics/ai-test-flight-air-force-unmanned-wingman-aircraft
https://www.theverge.com/2023/8/2/23817406/white-castle-soundhound-ai-sliders
https://www.google.com/search?sca_esv=556495916&q=food+delivery+bot+kicked&tbm=vid&source=lnms&sa=X&ved=2ahUKEwjZ6PDPrdmAAxUThf0HHWzrBGgQ0pQJegQIChAB&cshid=1691920142432720&biw=2327&bih=1180&dpr=2.2
https://www.youtube.com/watch?v=--n_NhmXnfc
https://www.thesun.co.uk/tech/20793591/coop-delivery-robots-cambridge-kicked-by-workers-tiktok/

#cybercrime #chatgpt #security

An interview with Sergey Shykevich, Threat Intelligence Group Manager at Check Point, about how models like ChatGPT have impacted the realm of cyber crime.

https://threatmap.checkpoint.com/

Links:
Homepage: https://ykilcher.com
Merch: https://ykilcher.com/merch
YouTube: https://www.youtube.com/c/yannickilcher
Twitter: https://twitter.com/ykilcher
Discord: https://ykilcher.com/discord
LinkedIn: https://www.linkedin.com/in/ykilcher

If you want to support me, the best thing to do is to share out the content :)

If you want to support me financially (completely optional and voluntary, but a lot of people have asked for this):
SubscribeStar: https://www.subscribestar.com/yannickilcher
Patreon: https://www.patreon.com/yannickilcher
Bitcoin (BTC): bc1q49lsw3q325tr58ygf8sudx2dqfguclvngvy2cq
Ethereum (ETH): 0x7ad3513E3B8f66799f507Aa7874b1B0eBC7F85e2
Litecoin (LTC): LQW2TRyKYetVC8WjFkhpPhtpbDM4Vw7r9m
Monero (XMR): 4ACL8AGrEo5hAir8A9CeVrW8pEauWvnp1WnSDZxW7tziCDLhZAGsgzhRQABDnFy8yuM9fWJDviJPHKRjV4FWt19CJZN9D4n