All about AI, Web 3.0, BCI
This channel about AI, Web 3.0, metaverse and brain computer interface(BCI)
owner @Aniaslanyan
owner @Aniaslanyan
Рэйтынг TGlist
0
0
ТыпПублічны
Вертыфікацыя
Не вертыфікаваныНадзейнасць
Не надзейныРазмяшчэнне
МоваІншая
Дата стварэння каналаApr 28, 2022
Дадана ў TGlist
Jun 11, 2024Прыкрепленая група
Апошнія публікацыі ў групе "All about AI, Web 3.0, BCI"
10.04.202517:11
OpenAI has rolled out a significant enhancement to ChatGPT's memory capabilities.
The AI can now reference your entire chat history to deliver more personalized responses based on your preferences and interests.
Unlike before, where ChatGPT only used specifically saved memories, it can now automatically analyze patterns across all your past conversations to provide more relevant assistance for writing, advice, learning, and more.
Memory isn't just another product feature. It signals a shift from episodic interactions (think a call center) to evolving ones (more like a colleague or friend).
Users maintain full control - you can opt out of this feature in settings or use temporary chats for private conversations. Currently available to Plus and Pro users (except in certain European regions).
The AI can now reference your entire chat history to deliver more personalized responses based on your preferences and interests.
Unlike before, where ChatGPT only used specifically saved memories, it can now automatically analyze patterns across all your past conversations to provide more relevant assistance for writing, advice, learning, and more.
Memory isn't just another product feature. It signals a shift from episodic interactions (think a call center) to evolving ones (more like a colleague or friend).
Users maintain full control - you can opt out of this feature in settings or use temporary chats for private conversations. Currently available to Plus and Pro users (except in certain European regions).
09.04.202513:42
Google announced their new TPU 'Ironwood'. Each individual chip has a peak compute of 4,614 TFLOPs. Google is calling it 'a monumental leap in AI capability'.
09.04.202511:00
Amazon launched Nova Sonic speech-to-speech AI for human-like interactions
—Outperforms OpenAI's voice models with ~ 80% less cost
—4.2% word error rate across languages
— 46.7% better accuracy than GPT-4o for noisy environments
—On Amazon Bedrock.
Amazon also dropped an upgraded Nova Reel 1.1 video model
—Delivers improved quality, style consistency
—Extends generations to 2 min via automated and manual, shot-by-shot modes
—Also available on Amazon Bedrock.
—Outperforms OpenAI's voice models with ~ 80% less cost
—4.2% word error rate across languages
— 46.7% better accuracy than GPT-4o for noisy environments
—On Amazon Bedrock.
Amazon also dropped an upgraded Nova Reel 1.1 video model
—Delivers improved quality, style consistency
—Extends generations to 2 min via automated and manual, shot-by-shot modes
—Also available on Amazon Bedrock.
08.04.202517:49
Anthropic're giving out $50k in free API credits for devs to try Claude Code and build more stuff using it.
08.04.202517:21
Multi-agent architectures are the future
Here are 6 different types:
𝟭. 𝗛𝗶𝗲𝗿𝗮𝗿𝗰𝗵𝗶𝗰𝗮𝗹 (𝗩𝗲𝗿𝘁𝗶𝗰𝗮𝗹) 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
A supervisor agent orchestrates multiple specialized agents.
𝘌𝘹𝘢𝘮𝘱𝘭𝘦:
• One agent retrieves information from internal data sources
• Another agent specializes in public information from web searches
• A third agent specializes in retrieving information from personal accounts (email, chat)
𝟮. 𝗛𝘂𝗺𝗮𝗻-𝗶𝗻-𝘁𝗵𝗲-𝗟𝗼𝗼𝗽 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Incorporates human verification before proceeding to next actions, used when handling sensitive information.
𝟯/𝟱. 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 (𝗛𝗼𝗿𝗶𝘇𝗼𝗻𝘁𝗮𝗹) 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Agents communicate directly with one another in a many-to-many fashion. Forms a decentralized network without strict hierarchical structure.
𝟰. 𝗦𝗲𝗾𝘂𝗲𝗻𝘁𝗶𝗮𝗹 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Agents process tasks in sequence, where one agent's output becomes input for the next.
𝘌𝘹𝘢𝘮𝘱𝘭𝘦: Three sequential agents where:
• First query agent retrieves information from vector search
• Second query agent retrieves additional information from web search based on first agent's findings
• Final generation agent creates a response using information from both query agents
𝟱. 𝗗𝗮𝘁𝗮 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Includes agents dedicated to transforming data.
𝘌𝘹𝘢𝘮𝘱𝘭𝘦:
• A transformation agent that enriches data at insert-time or transforms existing collections
There are also some other patterns that can be combined with these architectures:
• 𝗟𝗼𝗼𝗽 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: Iterative cycles for continuous improvement
• 𝗣𝗮𝗿𝗮𝗹𝗹𝗲𝗹 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: Multiple agents working simultaneously on different parts of a task
• 𝗥𝗼𝘂𝘁𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: A central router determining which agents to invoke
• 𝗔𝗴𝗴𝗿𝗲𝗴𝗮𝘁𝗼𝗿/𝘀𝘆𝗻𝘁𝗵𝗲𝘀𝗶𝘇𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: Collecting and synthesizing outputs from multiple agents
Here are 6 different types:
𝟭. 𝗛𝗶𝗲𝗿𝗮𝗿𝗰𝗵𝗶𝗰𝗮𝗹 (𝗩𝗲𝗿𝘁𝗶𝗰𝗮𝗹) 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
A supervisor agent orchestrates multiple specialized agents.
𝘌𝘹𝘢𝘮𝘱𝘭𝘦:
• One agent retrieves information from internal data sources
• Another agent specializes in public information from web searches
• A third agent specializes in retrieving information from personal accounts (email, chat)
𝟮. 𝗛𝘂𝗺𝗮𝗻-𝗶𝗻-𝘁𝗵𝗲-𝗟𝗼𝗼𝗽 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Incorporates human verification before proceeding to next actions, used when handling sensitive information.
𝟯/𝟱. 𝗡𝗲𝘁𝘄𝗼𝗿𝗸 (𝗛𝗼𝗿𝗶𝘇𝗼𝗻𝘁𝗮𝗹) 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Agents communicate directly with one another in a many-to-many fashion. Forms a decentralized network without strict hierarchical structure.
𝟰. 𝗦𝗲𝗾𝘂𝗲𝗻𝘁𝗶𝗮𝗹 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Agents process tasks in sequence, where one agent's output becomes input for the next.
𝘌𝘹𝘢𝘮𝘱𝘭𝘦: Three sequential agents where:
• First query agent retrieves information from vector search
• Second query agent retrieves additional information from web search based on first agent's findings
• Final generation agent creates a response using information from both query agents
𝟱. 𝗗𝗮𝘁𝗮 𝗧𝗿𝗮𝗻𝘀𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲
Includes agents dedicated to transforming data.
𝘌𝘹𝘢𝘮𝘱𝘭𝘦:
• A transformation agent that enriches data at insert-time or transforms existing collections
There are also some other patterns that can be combined with these architectures:
• 𝗟𝗼𝗼𝗽 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: Iterative cycles for continuous improvement
• 𝗣𝗮𝗿𝗮𝗹𝗹𝗲𝗹 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: Multiple agents working simultaneously on different parts of a task
• 𝗥𝗼𝘂𝘁𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: A central router determining which agents to invoke
• 𝗔𝗴𝗴𝗿𝗲𝗴𝗮𝘁𝗼𝗿/𝘀𝘆𝗻𝘁𝗵𝗲𝘀𝗶𝘇𝗲𝗿 𝗽𝗮𝘁𝘁𝗲𝗿𝗻: Collecting and synthesizing outputs from multiple agents
08.04.202514:47
Nvidia has shrunk 405B, post-trained for reasoning, and is claiming advantage over R1.
Nvidia released Llama-Nemotron-Ultra. This is a reasoning ON/OFF, dense 253B model. Open weights and post-training data.
Nvidia released Llama-Nemotron-Ultra. This is a reasoning ON/OFF, dense 253B model. Open weights and post-training data.
08.04.202513:58
Test-Time Training (TTT) is now on Video and not just a 5-second video. U can generate a full 1-min video!
TTT module is an RNN module that provides an explicit and efficient memory mechanism. It models the hidden state of an RNN with a machine learning model, which is updated via gradient descent.
TTT module is an RNN module that provides an explicit and efficient memory mechanism. It models the hidden state of an RNN with a machine learning model, which is updated via gradient descent.
08.04.202512:04
A new brain-computer interface so small it slips between your hair follicles just solved the last barrier to neural control: movement
96.4% accuracy while running.
this means AR glasses you control with thought, anywhere, anytime.
The team demonstrated an AR video calling system entirely controlled by brain signals. Users focused on visual stimuli in AR glasses to answer/reject/end calls while walking freely. They got 99.2% accuracy while standing, 97.5% while walking, and 92.5% while running.
Micro-brain sensors are tiny cross-shaped needles with conductive polymer coatings that target gaps between hair follicles. They penetrate just the outer skin layers, bypassing the insulating stratum corneum to establish direct contact with brain signals.
The core innovation combines three technologies:
1) dental resin microneedles for biocompatibility
2) PEDOT conductive polymer coating with 5-8x higher conductivity than standard materials
3) serpentine interconnectors that isolate the sensors from motion vibrations.
The breakthrough is ultralow impedance density (0.03 kΩ·cm−2) - the lowest reported, combined with their form factor, these sensors record high-fidelity signals even as subjects run, walk, climb stairs, or navigate hallways. No other neural electrode can match this.
The system maintained signal quality over a 12-hour test period with great stability. While conventional electrodes' signals degraded within hours, these sensors showed consistent performance while standing, walking, and running.
Gold cup electrodes: impedance density 10-100x higher, signals corrupted by movement, need conductive gel, hair interferes with signals. Micro-sensors: maintain clear signals while running, insert between hair follicles, can be worn 12+ hours, no skin preparation required .
96.4% accuracy while running.
this means AR glasses you control with thought, anywhere, anytime.
The team demonstrated an AR video calling system entirely controlled by brain signals. Users focused on visual stimuli in AR glasses to answer/reject/end calls while walking freely. They got 99.2% accuracy while standing, 97.5% while walking, and 92.5% while running.
Micro-brain sensors are tiny cross-shaped needles with conductive polymer coatings that target gaps between hair follicles. They penetrate just the outer skin layers, bypassing the insulating stratum corneum to establish direct contact with brain signals.
The core innovation combines three technologies:
1) dental resin microneedles for biocompatibility
2) PEDOT conductive polymer coating with 5-8x higher conductivity than standard materials
3) serpentine interconnectors that isolate the sensors from motion vibrations.
The breakthrough is ultralow impedance density (0.03 kΩ·cm−2) - the lowest reported, combined with their form factor, these sensors record high-fidelity signals even as subjects run, walk, climb stairs, or navigate hallways. No other neural electrode can match this.
The system maintained signal quality over a 12-hour test period with great stability. While conventional electrodes' signals degraded within hours, these sensors showed consistent performance while standing, walking, and running.
Gold cup electrodes: impedance density 10-100x higher, signals corrupted by movement, need conductive gel, hair interferes with signals. Micro-sensors: maintain clear signals while running, insert between hair follicles, can be worn 12+ hours, no skin preparation required .
08.04.202510:06
Salesforce presents APIGen-MT a framework for generating high-quality, verifiable multi-turn training data for AI agents.
APIGen-MT employs a sophisticated two-phase methodology:
Phase 1: Task Configuration and Groundtruth Generation
The first phase systematically creates detailed "blueprints" for each task, complete with user instructions, verifiable groundtruth actions, and expected outputs.
Phase 2: Human-Agent-Environment Interaction
The second phase transforms blueprints into natural conversations by simulating interactions between an LLM-based "human" and a test agent operating within an executable environment. The simulation captures dialogue turns, API calls, and environment responses, resulting in complete interaction trajectories.
Only trajectories that verifiably accomplish the task are accepted into the final dataset, ensuring both conversational realism and solution correctness.
The APIGen-MT approach offers several significant advantages:
Verifiability: All interaction data is grounded in pre-validated task configurations
Realism: The simulation focuses on natural dialogue without sacrificing task correctness
Efficiency: Smaller models trained on this data can match or exceed the performance of much larger alternatives
Consistency: Models demonstrate reliable performance across multiple trials
This work demonstrates that carefully designed synthetic data generation can dramatically improve AI assistant capabilities, particularly for complex multi-turn interactions requiring tool use—potentially transforming how we train the next generation of AI agents.
APIGen-MT employs a sophisticated two-phase methodology:
Phase 1: Task Configuration and Groundtruth Generation
The first phase systematically creates detailed "blueprints" for each task, complete with user instructions, verifiable groundtruth actions, and expected outputs.
Phase 2: Human-Agent-Environment Interaction
The second phase transforms blueprints into natural conversations by simulating interactions between an LLM-based "human" and a test agent operating within an executable environment. The simulation captures dialogue turns, API calls, and environment responses, resulting in complete interaction trajectories.
Only trajectories that verifiably accomplish the task are accepted into the final dataset, ensuring both conversational realism and solution correctness.
The APIGen-MT approach offers several significant advantages:
Verifiability: All interaction data is grounded in pre-validated task configurations
Realism: The simulation focuses on natural dialogue without sacrificing task correctness
Efficiency: Smaller models trained on this data can match or exceed the performance of much larger alternatives
Consistency: Models demonstrate reliable performance across multiple trials
This work demonstrates that carefully designed synthetic data generation can dramatically improve AI assistant capabilities, particularly for complex multi-turn interactions requiring tool use—potentially transforming how we train the next generation of AI agents.
07.04.202517:44
Pakistan appoints Binance founder Changpeng Zhao as official strategic advisor to Pakistan Cryptocurrency Council.
CZ said: Pakistan is a country with a population of 240 million, of which more than 60% are under the age of 30.
CZ said: Pakistan is a country with a population of 240 million, of which more than 60% are under the age of 30.
04.04.202511:48
Midjourney released V7 Alpha
V7 Alpha is much smarter with text prompts, has higher-quality images with improved textures and coherence, default-on personalization requiring a ~5-minute unlock, new fraft Mode rendering images 10x faster at half cost with conversational and voice input, Turbo (2x cost) and Relax modes, temporary fallback to V6 for upscaling/editing/retexture, ongoing Moodboard and SREF improvements, and planned updates every 1-2 weeks for the next 60 days, starting with new V7 character and object reference features.
V7 Alpha is much smarter with text prompts, has higher-quality images with improved textures and coherence, default-on personalization requiring a ~5-minute unlock, new fraft Mode rendering images 10x faster at half cost with conversational and voice input, Turbo (2x cost) and Relax modes, temporary fallback to V6 for upscaling/editing/retexture, ongoing Moodboard and SREF improvements, and planned updates every 1-2 weeks for the next 60 days, starting with new V7 character and object reference features.
04.04.202509:20
New #DeepSeek Paper+Model
DeepSeek-GRM models automatically generate judging principles and critiques without needing a human in the loop to achieve better reward scaling with inference-time compute.
Open-source model coming.
They're worried automated principles/critiques might amplify biases from toxic training data without a human in the loop.
DeepSeek-GRM models automatically generate judging principles and critiques without needing a human in the loop to achieve better reward scaling with inference-time compute.
Open-source model coming.
They're worried automated principles/critiques might amplify biases from toxic training data without a human in the loop.
03.04.202512:56
CAMEL-AI's Trifecta: Loong, OWL, and CRAB - The Future of AI Agent Systems
Loong: Self-Improving AI in Specialized Domains
Project Loong tackles the fundamental challenge of training LLMs to reason effectively in specialized domains without expensive labeled data. Instead of manually creating massive datasets, Loong generates and verifies synthetic data automatically.
GitHub. HF.
Key features:
Generates domain-specific synthetic Q&A pairs from small seed datasets
Verifies correctness through dual validation: code execution and Chain-of-Thought reasoning
Trains models through reinforcement learning on verified synthetic data
Currently supports 8 specialized domains including advanced mathematics, physics, computational biology, and finance
Loong enables LLMs to develop expertise in domains where curated data is scarce, unlocking new potential for specialized AI.
OWL: The Digital Task Automator
OWL (Optimized Workforce Learning) addresses real-world task automation with an impressive track record - ranking #1 among open-source submissions on the GAIA benchmark.
Key capabilities:
Browser automation via Playwright
Multi-search engine support
Python code execution
Document parsing across formats
Multimodal processing (video, images, audio)
Integration with numerous specialized toolkits
OWL's multi-agent architecture uses a UserAgent to break down tasks and an AssistantAgent to execute them using various tools, making it effective for complex workflow automation.
CRAB: Breaking the Environment Barrier
CRAB (CRoss-environment Agent Benchmark) is the first framework enabling agents to perform tasks across multiple environments - from smartphones to desktops and beyond.
The Integrated Ecosystem: MCP as the Universal Connector
All 3 projects integrate with the Model Context Protocol (MCP), introduced by Anthropic and rapidly becoming the "USB interface" of the LLM world. MCP provides standardized connections between AI assistants and data systems, enabling seamless operation across tools and environments.
Together, these projects represent a comprehensive approach to autonomous AI:
Loong trains specialized domain knowledge
OWL executes complex tasks
CRAB enables operation across diverse environments.
This trifecta addresses the "last mile" challenge in agent automation: long-term decision-making and adaptation. Where traditional agents follow instructions but don't truly evolve, CAMEL-AI's ecosystem creates environments where agents can perceive, act, and learn from experience.
Loong: Self-Improving AI in Specialized Domains
Project Loong tackles the fundamental challenge of training LLMs to reason effectively in specialized domains without expensive labeled data. Instead of manually creating massive datasets, Loong generates and verifies synthetic data automatically.
GitHub. HF.
Key features:
Generates domain-specific synthetic Q&A pairs from small seed datasets
Verifies correctness through dual validation: code execution and Chain-of-Thought reasoning
Trains models through reinforcement learning on verified synthetic data
Currently supports 8 specialized domains including advanced mathematics, physics, computational biology, and finance
Loong enables LLMs to develop expertise in domains where curated data is scarce, unlocking new potential for specialized AI.
OWL: The Digital Task Automator
OWL (Optimized Workforce Learning) addresses real-world task automation with an impressive track record - ranking #1 among open-source submissions on the GAIA benchmark.
Key capabilities:
Browser automation via Playwright
Multi-search engine support
Python code execution
Document parsing across formats
Multimodal processing (video, images, audio)
Integration with numerous specialized toolkits
OWL's multi-agent architecture uses a UserAgent to break down tasks and an AssistantAgent to execute them using various tools, making it effective for complex workflow automation.
CRAB: Breaking the Environment Barrier
CRAB (CRoss-environment Agent Benchmark) is the first framework enabling agents to perform tasks across multiple environments - from smartphones to desktops and beyond.
The Integrated Ecosystem: MCP as the Universal Connector
All 3 projects integrate with the Model Context Protocol (MCP), introduced by Anthropic and rapidly becoming the "USB interface" of the LLM world. MCP provides standardized connections between AI assistants and data systems, enabling seamless operation across tools and environments.
Together, these projects represent a comprehensive approach to autonomous AI:
Loong trains specialized domain knowledge
OWL executes complex tasks
CRAB enables operation across diverse environments.
This trifecta addresses the "last mile" challenge in agent automation: long-term decision-making and adaptation. Where traditional agents follow instructions but don't truly evolve, CAMEL-AI's ecosystem creates environments where agents can perceive, act, and learn from experience.
03.04.202512:17
Google DeepMind have achieved a notable milestone in AI with the Dreamer V3 algorithm.
Published in Nature, "Mastering diverse control tasks through world models" introduces a reinforcement learning algorithm that performs well across more than 150 diverse tasks with a single configuration.
Key Achievements
Versatility: Dreamer works effectively across continuous and discrete action spaces, visual and vector inputs, dense and sparse rewards, and various application domains.
Minecraft Diamond Challenge: Applied with its default parameters, Dreamer successfully collects diamonds in Minecraft without human data or curricula - a challenging task requiring farsighted strategies and learning from sparse rewards in an open world.
Consistent Learning: Through techniques based on normalization, balancing, and transformations, Dreamer enables more stable learning across domains that traditionally required extensive hyperparameter tuning.
Scaling Properties: The research shows that larger models achieve higher performance while requiring less interaction with environments, offering a predictable relationship between computational resources and efficiency.
While Dreamer represents an evolutionary advancement rather than a revolutionary breakthrough, it addresses a significant challenge in reinforcement learning: the brittleness of algorithms when applied to new domains. Traditional approaches require substantial human expertise and experimentation for each new application, limiting practical utility.
Published in Nature, "Mastering diverse control tasks through world models" introduces a reinforcement learning algorithm that performs well across more than 150 diverse tasks with a single configuration.
Key Achievements
Versatility: Dreamer works effectively across continuous and discrete action spaces, visual and vector inputs, dense and sparse rewards, and various application domains.
Minecraft Diamond Challenge: Applied with its default parameters, Dreamer successfully collects diamonds in Minecraft without human data or curricula - a challenging task requiring farsighted strategies and learning from sparse rewards in an open world.
Consistent Learning: Through techniques based on normalization, balancing, and transformations, Dreamer enables more stable learning across domains that traditionally required extensive hyperparameter tuning.
Scaling Properties: The research shows that larger models achieve higher performance while requiring less interaction with environments, offering a predictable relationship between computational resources and efficiency.
While Dreamer represents an evolutionary advancement rather than a revolutionary breakthrough, it addresses a significant challenge in reinforcement learning: the brittleness of algorithms when applied to new domains. Traditional approaches require substantial human expertise and experimentation for each new application, limiting practical utility.
03.04.202509:59
General Agents introduced Ace: The First Realtime Computer Autopilot
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds.
Ace is not a chatbot. Ace performs tasks for you.
On your computer. Using your mouse and keyboard.
At superhuman speeds.
Рэкорды
10.04.202523:59
2.8KПадпісчыкаў28.02.202518:46
1200Індэкс цытавання03.04.202515:22
344Ахоп 1 паста23.04.202521:15
0Ахоп рэкламнага паста05.04.202523:59
9.84%ER03.04.202515:22
12.53%ERRУвайдзіце, каб разблакаваць больш функцый.
