https://lnkd.in/euMMXJR4 embedding spaces are a window on knowledge structure (exciting times ! 😃 ) one more step toward modularity ?
Jean-Frédéric Ferté’s Post
More Relevant Posts
-
Want to draft diagrams quickly and easily. There's a tool for that!
DiagramGPT – AI diagram generator
eraser.io
To view or add a comment, sign in
-
A new way to “fine-tune” LLMs via representation! We are all familiar with several philosophies for fine-tuning LLMs. 1️⃣ Full fine-tuning — the old-school method involves training all your weights. 2️⃣ Parameter-efficient fine-tuning (PEFT) methods, such as LoRA, DoRA etc - train only a small subset of weights. 💡And the new one I only recently learned about: Representation Fine-Tuning (ReFT) — the main idea is that you learn to modify input representations instead of model weights! It's 10x-50x smaller than any PEFT while outperforming it! Generic code to get started: https://lnkd.in/gA8jvvbp. It still needs to be battle-tested, but it looks very promising!
GitHub - stanfordnlp/pyreft: ReFT: Representation Finetuning for Language Models
github.com
To view or add a comment, sign in
-
bootcamp.uxdesign.cc: The most recent Midjourney Weekly Office Hours covered upcoming releases and discussed keeping up with rapid changes and developments. - Artificial Intelligence topics! #ai #artificialintelligence #intelligenzaartificiale
Quick Midjourney Update: 5 key highlights to know (8th November 2023)
bootcamp.uxdesign.cc
To view or add a comment, sign in
-
Learned a lot from this article on building embeddings!
How to Build a State-of-the-art Text Embedding Model
medium.com
To view or add a comment, sign in
-
Role Play with Large Language Models This paper was highlighted by Thomas Wolfe (HuggingFace) in his recent videos. https://lnkd.in/ddErJ8f8 Anthropomorphism - is the tendency to attribute human characteristics to non-human things. When a suitably prompted and sampled LLM can mimic human languages convincingly, we tend to fall into the trap of anthropomorphism. When LLM-based dialogue agents use the words "I" or "me"; it may suggest the presence of a self-aware entity with goals and a concern for its own survival. But there is a fundamental dissimilarities between an artificial system and a human. The two important cases of dialogue-agent behaviour is via deception and self-awareness. The authors have suggested the use of "role-plays" and simulation to describe the behaviour of dialog-agents. Many users have managed to jail-break dialogue agents by coaxing them into issuing threats or using toxic and abusive languages. Does apparently "exposing" its true nature. This is a classical case of anthropomorphism; to think of it as revealing the entitys own agenda. In a broader sense, it effectively embodies the role it's given. The agent excels at acting its part, allowing us to distinguish three cases of false information: confabulation (fabricating experiences as memory compensation), unintentional falsehoods during role-playing (incorrect information in weights), and deliberate deception when portraying a deceptive character. It may be reassuring to know that LLM-based dialogue agents are not conscious entities with their own agendas and instinct for self-preservation, but are simply role-playing. But an agent equipped with access to APIs can through its actions have a real-world consequence. There is also a larger question -- are humans "role-playing"? Or are our everyday actions, a result of our inherent conditioning and acceptance of our core beliefs? Arxiv: https://lnkd.in/dJbTWsfr Day 23/30 I will try to cover 30 papers over the next three months. These will be related to Large language models, generative models, reinforcement learning and other similar topics.You can read the curated list of my other posts on ThinkEvolve website: https://lnkd.in/dVgiXAN8 #largelanguagemodels #MPT #generativeai #generativemodels #hallucinations #guardrails #aiforgood #GPT4ALL #Naomic.ai #Openai #llama #RefinedWeb #falconLLM #hallucinations #nlp #languagegenerations #generativemodels #foundationalmodels #foundationmodels #computervision #segmentanything #SAM #FAIR #RAI #ResponsiblityAwareAI #Codex #HumanEval #CoPilot #Github #Nerf #photogrammetry #RealityStudio #Blender #Unity3D #tabulardata #tabulartransformers #huggingface #transformers #scalinglaws
large-language-models-30-papers-that-matter – Think Evolve Consultancy
https://www.thinkevolveconsulting.com
To view or add a comment, sign in
-
Gaussian Splatting SLAM Simultaneous Localisation and Mapping (SLAM) method recreates 3D scenes from live video images! Using a single monocular camera, the authors have reconstructed a 3D scene using Gaussian Splatting. To achieve accurate tracking, the authors have used atleast 50 iterations of gradient descent per frame. This requirement emphasises the necessity of presentation with computationally efficient view synthesis and gradient computation, making the choice of 3D representation a crucial part of designing a SLAM system. I personally think that hardware costs (of sensors) will reduce substantially over the next decades, probably making these techniques redundant. Paper: https://lnkd.in/dQwNFPvK Code: <To be released in Feb 2024> Day 29/30 I will try to cover 30 papers over the next three months...(its taken more than 6 months! But plan to complete them in this year!) These will be related to Large language models, generative models, reinforcement learning and other similar topics. You can read the curated list of my other posts on ThinkEvolve website: https://lnkd.in/df53QBwc #largelanguagemodels #MPT #generativeai #generativemodels #hallucinations #guardrails #aiforgood #GPT4ALL #Naomic.ai #Openai #llama #RefinedWeb #falconLLM #hallucinations #nlp #languagegenerations #generativemodels #foundationalmodels #foundationmodels #computervision #segmentanything #SAM #FAIR #RAI #ResponsiblityAwareAI #Codex #HumanEval #CoPilot #Github #Nerf #photogrammetry #RealityStudio #Blender #Unity3D #tabulardata #tabulartransformers #huggingface #transformers #scalinglaws #localgpt #localllama #privacy
LinkedIn
https://www.thinkevolveconsulting.com
To view or add a comment, sign in
-
Google Developer Expert (GDE), AWS Community Builder, Senior Manager Data Science, Consultant, Trainer, Podcaster, Founder Malaysia R User Group, AI & ML Malaysia User Group
What We Learned from a Year of Building with LLMs https://lnkd.in/gsnVRPjE
What We Learned from a Year of Building with LLMs (Part I)
oreilly.com
To view or add a comment, sign in
-
Video Generation Models as World Simulators {SORA} "World Simulators" -- there is a certain grand standing to it! 😉 The model is now available to a select few, so a lot of the discussion will be speculation based on the technical paper that was shared. Sora can generate videos and images spanning diverse aspect ratios, resolutions and duration (upto a full minute of HD video) Each video is first compressed into a lower dimension space and then further decomposed into spacetime patches. These are equivalent to "text tokens" in large language models. The model has been trained on videos at their native aspect ratios -- which apparently improves the framing and composition. They have also used high descriptive captions to train the model, this ensures textual fidelity and Sora is able accurately follow user prompts. Other superpowers 🦹 include: animating images, extending generated videos, so you can have an infinite looping video. Ability to edit videos -- like changing the background from a snowy to a lush jungle. Interpolating between videos, 3D consistency, long-range coherence, object permanence and object interactions (bite marks in an apple) Fluid dynamics is well captured, while other physics like glass shattering is not good. I have captured other critiques in my twitter thread below Link to Technical Note: https://ow.ly/TjUS50QEEqE My twitter critque on the shared videos: https://ow.ly/Vsjx50QEEqF Day 31/30 I will try to cover 30 papers over the next three months...(its taken more than 6 months! But plan to complete them in this year!) These will be related to Large language models, generative models, reinforcement learning and other similar topics. You can read the curated list of my other posts on ThinkEvolve website: https://ow.ly/zPRj50QEEqG #sora #technicalpaper #largelanguagemodels #MPT #generativeai #generativemodels #hallucinations #guardrails #aiforgood #GPT4ALL #Naomic.ai #Openai #llama #RefinedWeb #falconLLM #hallucinations #nlp #languagegenerations #generativemodels #foundationalmodels #foundationmodels #computervision #segmentanything #SAM #FAIR #RAI #ResponsiblityAwareAI #Codex #HumanEval #CoPilot #Github #Nerf #photogrammetry #RealityStudio #Blender #Unity3D #tabulardata #tabulartransformers #huggingface #transformers #scalinglaws #localgpt #localllama #privacy
large-language-models-30-papers-that-matter – Think Evolve Consultancy
https://www.thinkevolveconsulting.com
To view or add a comment, sign in
-
Visual Walkthrough for Vectorized BERTScore to Evaluate Text Generation https://bit.ly/49gSXL3
Visual Walkthrough for Vectorized BERTScore to Evaluate Text Generation
https://towardsai.net
To view or add a comment, sign in
-
Visual Walkthrough for Vectorized BERTScore to Evaluate Text Generation https://bit.ly/49gSXL3
Visual Walkthrough for Vectorized BERTScore to Evaluate Text Generation
https://towardsai.net
To view or add a comment, sign in