iiinova — where your digital self keeps evolving

Ksenia Se / Turing Post @turingpost.bsky.social · 11/12/2024

"Densing Law of LLMs" paper: arxiv.org/abs/2412.04315

Ksenia Se / Turing Post @turingpost.bsky.social · 11/12/2024

Here are the key findings from the study: • Costs to run models are dropping as they are becoming more efficient. • The release of ChatGPT sped up the growth of efficiency of new models up to 50%! • Techniques like pruning and distillation don’t necessarily make models more efficient.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 11/12/2024

Reading about scaling laws recently I came by the interesting point: Focus on a balance between models' size and performance is more important that aiming for larger models Tsinghua University and ModelBest Inc propose the idea of “capacity density” to measure how efficiently a model uses its size

124

Ksenia Se / Turing Post @turingpost.bsky.social · 10/12/2024

1. GoogleDeepMind's Genie 2 Generates 3D environments with object interactions, animations, and physical effects from one image or text prompt. You can interact with them in real-time using a keyboard and mouse. Paper: deepmind.google/discover/blo... Our example: www.youtube.com/watch?v=YjO6...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 10/12/2024

An incredible shift is happening in spatial intelligence! Here are 2 latest revolutional World Models, which create interactive 3D environments: 1. GoogleDeepMind's Genie 2 2. AI system from World Labs, co-founded by Fei-Fei Li Explore more below 👇

120

Ksenia Se / Turing Post @turingpost.bsky.social · 05/12/2024

What is Flow Matching? Flow Matching (FM) is used in top generative models, like Flux, F5-TTS, E2-TTS, and MovieGen with state-pf-the-art results. Some experts even say that FM might surpass diffusion models👇

100

Ksenia Se / Turing Post @turingpost.bsky.social · 05/12/2024

INTELLECT-1 by Prime Intellect INTELLECT-1 is a 10B open-source LLM trained over 42 days on 1T tokens across 14 global nodes, leverages the PRIME framework for exceptional efficiency (400× bandwidth reduction). github.com/PrimeIntelle...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 05/12/2024

MultiFoley by Adobe Research MultiFoley is an AI model generating high-quality sound effects from text, audio, and video inputs. Cool demos highlight its creative potential. arxiv.org/abs/2411.17698

100

Ksenia Se / Turing Post @turingpost.bsky.social · 05/12/2024

ShowUI by Show Lab, NUS, Microsoft ShowUI is a 2B vision-language-action model tailored for GUI tasks: - features UI-guided token selection (33% fewer tokens) - interleaved streaming for multi-turn tasks - 256K dataset - achieves 75.1% zero-shot grounding accuracy arxiv.org/abs/2411.17465

100

Ksenia Se / Turing Post @turingpost.bsky.social · 05/12/2024

OLMo 2 by Allen AI OLMo 2, a family of fully open LMs with 7B and 13B parameter, is trained on 5 trillion tokens. allenai.org/blog/olmo2

110

Ksenia Se / Turing Post @turingpost.bsky.social · 05/12/2024

Amazing models of the week: • Alibaba’s QwQ-32B • OLMo 2 by Allen AI • ShowUI by Show Lab, NUS, Microsoft • Adobe's MultiFoley • INTELLECT-1 by Prime Intellect 🧵

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

Boundless Socratic Learning with Language Games, Google DeepMind This framework leverages recursive language-based "games" for self-improvement, focusing of feedback, coverage, and scalability. It suggests a roadmap for scalable AI via autonomous data gen and feedback loops arxiv.org/abs/2411.16905

110

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

MH-MoE: Multi-Head Mixture-of-Experts @msftresearch.bsky.social’s MH-MoE improves sparse MoE by adding multi-head attention, reducing perplexity without increasing FLOPs, and demonstrating robust performance under quantization. arxiv.org/abs/2411.16205

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

LLM-as-a-Judge: Presents a taxonomy of methodologies and applications of LLMs for judgment tasks, highlighting bias, vulnerabilities, and self-judgment, with future directions in human-LLM collaboration and bias mitigation arxiv.org/abs/2411.16594

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

Star Attention: NVIDIA introduced a block-sparse attention mechanism for Transformer-based LLMs. It uses local/global attention phases to achieve up to 11x inference speedup on sequences up to 1M tokens, retaining 95-100% accuracy. arxiv.org/abs/2411.17116 Code: github.com/NVIDIA/Star-...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

Top 5 researches of the week: • Natural Language Reinforcement Learning • Star Attention, NVIDIA • Opportunities and Challenges of LLM-as-a-judge • MH-MoE: Multi-Head Mixture-of-Experts, @msftresearch.bsky.social • Boundless Socratic Learning with Language Games, Google DeepMind 🧵

120

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

Paper: Large Language Model-Brained GUI Agents: A Survey arxiv.org/abs/2411.18279

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

2. Prompt engineering or creating a plan: After "looking" at the screen, the agent prepares a prompt for the AI model, which includes the user’s instructions, the visual data (like screenshots or buttons layout), and other context the agent needs to understand the task.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

1. Understanding the environment: Firstly, the agent needs to "see" the software it’s working with. This is done through methods that capture the layout of the app or website, such as screenshots, lists of buttons and menus called widget trees.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 02/12/2024

LLM-brained GUI agents are a way to interact with GUIs in a much more flexible and human-like way. They blend LLMs' capabilities with software interaction to work with websites, mobile apps, and desktop software, simplifying complex tasks. A new survey on LLM-brained GUI agents was published👇

111

Ksenia Se / Turing Post @turingpost.bsky.social · 01/12/2024

Our TuringPost Twitter Library provides useful resources, such as lists of tools and models, papers and courses on various popular aspects of AI and machine learning for you every week. Check it out. www.turingpost.com/t/Twitter-Li...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 01/12/2024

Top 10 GitHub Repositories to master ML, AI and Data Science: • 100 Days of ML Code • Data Science For Beginners • Awesome Data Science • Data Science Masters • Homemade Machine Learning • 500+ AI Projects List with Code • Awesome Artificial Intelligence ... Check out for more👇

121

Ksenia Se / Turing Post @turingpost.bsky.social · 30/11/2024

Benefits of HiAR-ICL: - Thanks to thought cards, HiAR-ICL requires less human input to guide it and adapts better to different types of problems -Faster by focusing only on actions from thought card - Accuracy: It achieved 79.6% accuracy on a math benchmark, compared to GPT-4o’s 76.6%.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 30/11/2024

Thought Cards with Monte Carlo Tree Search (MCTS): Thought cards are like reusable problem-solving templates. After MCTS generates multiple possible solution paths for each problem, HiAR-ICL picks the best one by weighing the paths using Value of Computation (VOC).

110

Ksenia Se / Turing Post @turingpost.bsky.social · 30/11/2024

This new approach to In-Context Learning (ICL) is very interesting: The idea of HiAR-ICL (High-level Automated Reasoning in ICL) is to teaches the model abstract thinking patterns, using atomic reasoning actions and thought cards. It doesn't rely only on examples and prompts like in ICL Details👇

120

Ksenia Se / Turing Post @turingpost.bsky.social · 29/11/2024

What is Self-Supervised Learning? Whether you are a professional or a beginner in Machine Learning, we made our flashcards to be easy to digest and help everyone refresh key ML concepts. Here's Self-Supervised Learning, a technique used to train ML models👇

100

Ksenia Se / Turing Post @turingpost.bsky.social · 29/11/2024

In NLRL, LLMs take on different roles like: • They critique and plan using prompts (for example, explain their choices). • Learn to evaluate actions using language (act as a validator). • They are "actors" (decision-makers) and "critics" (evaluators) in a feedback loop

110

Ksenia Se / Turing Post @turingpost.bsky.social · 29/11/2024

2. Language Temporal-Difference Estimate: Uses the language Bellman equation to breaks down the value of a state into two parts: the immediate reward and the value of the next state.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 29/11/2024

1. Language Monte-Carlo Estimate: Uses a language aggregator tool to combine the information into a summary value.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 29/11/2024

Natural Language Reinforcement Learning (NLRL) redefines Reinforcement Learning (RL). NLRL's main idea: The core parts of RL like goals, strategies, and evaluation methods are reimagined using natural language instead of rigid math. Let's explore this approach more precisely🧵

121

Ksenia Se / Turing Post @turingpost.bsky.social · 29/11/2024

A quick reminder of what AI agents are built of: 1. Profiling: Keeps the agent aligned with its purpose. 2. Knowledge: Provides domain-specific expertise. Find other components below👇 P.S.: As an example, here's LangChain’s Harrison Chase's agent framework.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

AI should be for everyone, right? :) Here are 6 FREE AI courses for beginners: • Introduction to Artificial Intelligence • Artificial Intelligence for Beginners • AI For Everyone • Machine Learning for Beginners • Introduction to Data Science Specialization • Data Science for Beginners Links👇

140

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Does LLaVA-o1 VLM challenge OpenAI's o1 model? Here's what enhances LLaVA-o1's complex multimodal reasoning: - Reasoning in 4 stages - Effective inference-time scaling - Stage-level beam search - Special dataset with step-by-step reasoning examples 👇

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Hymba's performance: Hymba-1.5B achieves 1.32% higher accuracy than Llama-3.2-3B on commonsense reasoning tasks. It reduces memory cache size by 11.67x and achieves 3.49x faster throughput compared to Llama-3.2-3B.

110

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Hymba's attention map combines 3 key elements: - Meta tokens: Offloads attention from the "beginning-of-sequence" (BOS) token, and the model focuses more on real input tokens. - Sliding window attention - SSMs: Summarizes the overall context while focusing on current tokens.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Meta tokens: They are special, learnable tokens added at the input's start to help Hymba focus on key information. They create a modified sequence where the tokens act as a guide for the attention mechanism. During inference, the meta tokens are fixed, and their computations can be preprocessed.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Hybrid-head architecture: Hymba combines attention and SSM layers in parallel within the same layer, unlike existing hybrid models that use separate layers for them. This makes Hymba flexible and effective across different tasks.

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

NVIDIA's Hymba small model is a great combo of 2 concepts: - Transformer attention to help the model remember details. - State Space Models (SSMs) to efficiently summarize context. This model is also a treasure trove of interesting features. Here are the details:

200

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

A Flexible Large Language Models Guardrail Development Methodology Applied to Off-Topic Prompt Detection Proposes a guardrail methodology for detecting off-topic prompts, utilizing synthetic datasets and fine-tuning for robust performance. arxiv.org/abs/2411.12946 huggingface.co/collections/...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models Identifies directions in model representations to mitigate hallucinations and refine entity recognition. arxiv.org/abs/2411.14257

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Building Trust: Foundations of Security, Safety, and Transparency in AI Establishes safety and security frameworks for responsible AI development and standardized risk management. arxiv.org/abs/2411.12275

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Search, Verify and Feedback: Towards Next Generation Post-training Paradigm of Foundation Models via Verifier Engineering Proposes a post-training paradigm for foundation models, enhancing scalability and alignment with verification techniques. arxiv.org/abs/2411.11504 github.com/icip-cas/Ver...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

One to Rule Them All: Natural Language to Bind Communication, Perception, and Action Introduces a robotic architecture integrating LLMs for task execution and dynamic environmental adaptation. arxiv.org/abs/2411.15033

110

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Loss-to-Loss Prediction: Scaling Laws for All Datasets Develops predictive scaling laws for model performance across tasks and datasets, enhancing efficiency and planning. arxiv.org/abs/2411.12925 Models: huggingface.co/KempnerInsti...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

When Precision Meets Position: BFloat16 Breaks Down RoPE in Long-Context Training Optimizes long-context processing by resolving precision issues, accelerating training while preserving performance. arxiv.org/abs/2411.13476 Code: github.com/haonan3/Anch...

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Ultra-Sparse Memory Network Introduces ultra-sparse architectures that reduce latency and improve memory efficiency, rivaling larger dense models. arxiv.org/abs/2411.12364

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Generative World Explorer Develops Genex for imaginative exploration in 3D environments, enabling decision-making and belief revision without physical movement. arxiv.org/abs/2411.11844

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Drowning in Documents: Consequences of Scaling Reranker Inference Analyzes challenges of scaling rerankers for large datasets, showing limitations and proposing robust listwise reranking alternatives. arxiv.org/abs/2411.11767

100

Ksenia Se / Turing Post @turingpost.bsky.social · 28/11/2024

Openscholar: Synthesizing Scientific Literature With Retrieval-Augmented LMs Builds a retrieval-augmented LLM, outperforming GPT-4o in synthesizing scientific queries from a large literature dataset. arxiv.org/abs/2411.14199 Code: github.com/AkariAsai/Op...

100