A new tutorial on RL by Kevin Patrick Murphy, a Research Scientist at Google DeepMind who also wrote several comprehensive, well-regarded textbooks on ML/DL.
arXiv.org
Reinforcement Learning: An Overview
This manuscript gives a big-picture, up-to-date overview of the field of (deep) reinforcement learning and sequential decision making, covering value-based methods, policy-based methods,...
🦄5
Multi_agent_systems_III_beyond_single_agents_1733748286.pdf
876.1 KB
What Comes After AI Agents?
By Dr. Egor Kraev, Head of AI at Wise
Over the past year, we've witnessed a boom in AI agents. Many frameworks have emerged - Langchain, Autogen, CrewAI, and others. New solutions and approaches appear weekly. But what's next? Where is this technology heading?
Why Agents Alone Are Not Enough
The current approach of trying to solve everything with "smart agents" is like trying to build a house using only a hammer. Yes, a hammer is a great tool, but for quality results, you need a whole set of tools.
The Future Lies in Compound Systems
The next stage of development is compound systems, where AI agents become just one component of a larger system.
Think of it as a construction set where each part plays its role:
- Traditional databases for information storage
- Classical algorithms where they're more efficient
- AI agents for tasks requiring "intelligent" analysis
- Standard interfaces for component interaction
Core Benefits
1. Flexibility. Mix and match components based on specific needs
2. Cost-effectiveness. Use expensive AI components only where necessary
3. Risk management. Not putting all eggs in one basket
4. Future-proofing. Easier to adopt new technologies as they emerge
The era of standalone AI agents is coming to an end. The future belongs to integrated solutions where artificial intelligence is an important, but not the only component of the system. This isn't a step backward, but rather a natural evolution of technology, making it truly applicable in the real world.
By Dr. Egor Kraev, Head of AI at Wise
Over the past year, we've witnessed a boom in AI agents. Many frameworks have emerged - Langchain, Autogen, CrewAI, and others. New solutions and approaches appear weekly. But what's next? Where is this technology heading?
Why Agents Alone Are Not Enough
The current approach of trying to solve everything with "smart agents" is like trying to build a house using only a hammer. Yes, a hammer is a great tool, but for quality results, you need a whole set of tools.
The Future Lies in Compound Systems
The next stage of development is compound systems, where AI agents become just one component of a larger system.
Think of it as a construction set where each part plays its role:
- Traditional databases for information storage
- Classical algorithms where they're more efficient
- AI agents for tasks requiring "intelligent" analysis
- Standard interfaces for component interaction
Core Benefits
1. Flexibility. Mix and match components based on specific needs
2. Cost-effectiveness. Use expensive AI components only where necessary
3. Risk management. Not putting all eggs in one basket
4. Future-proofing. Easier to adopt new technologies as they emerge
The era of standalone AI agents is coming to an end. The future belongs to integrated solutions where artificial intelligence is an important, but not the only component of the system. This isn't a step backward, but rather a natural evolution of technology, making it truly applicable in the real world.
👍5
Streamline AI Agent Evaluation with Synthetic Data
Databricks has introduced a new synthetic data generation API to streamline the testing and improvement of AI agents. This feature significantly reduces the time and cost associated with evaluating agent performance.
The synthetic data generation API automates the evaluation process by:
- Automatically generating test questions and answers from company documents
- Creating comprehensive evaluation datasets in minutes
- Enabling rapid quality assessment without SME bottlenecks
- Providing immediate feedback on agent performance
Databricks has introduced a new synthetic data generation API to streamline the testing and improvement of AI agents. This feature significantly reduces the time and cost associated with evaluating agent performance.
The synthetic data generation API automates the evaluation process by:
- Automatically generating test questions and answers from company documents
- Creating comprehensive evaluation datasets in minutes
- Enabling rapid quality assessment without SME bottlenecks
- Providing immediate feedback on agent performance
Databricks
Streamline AI Agent Evaluation with New Synthetic Data Capabilities | Databricks Blog
2025 Crypto Trends: a16z's Vision for Building a Resilient Future
Key takeaway: While these trends showcase technological advancement, they also represent a strategic shift toward building financial infrastructure that's more resilient to centralized control.
Leading VC firm a16z crypto shared their vision for 2025, and it's fascinating how these predictions align with addressing current challenges in the financial system:
1. AI Agents with Wallets
Imagine AI chatbots that can independently own and transact in crypto. This isn't just about technology - it's about creating autonomous financial actors that operate outside traditional banking constraints.
2. Proof of Personhood
In a world of deepfakes and AI-generated content, verifying real human identity becomes crucial. The industry is working on new ways to prove humanity while preserving privacy.
3. Enterprise Stablecoins
Major companies are expected to adopt stablecoins for payments - not just because they're the cheapest way to send dollars globally, but as a hedge against potential banking restrictions.
4. User Experience Revolution
Crypto is finally prioritizing accessibility over infrastructure. New crypto-native app stores will emerge, similar to the App Store but without centralized control.
5. Asset Tokenization
From government bonds to biometric data - more assets will move to blockchain. This creates a parallel financial system less dependent on traditional banking.
Interesting times ahead!
Key takeaway: While these trends showcase technological advancement, they also represent a strategic shift toward building financial infrastructure that's more resilient to centralized control.
Leading VC firm a16z crypto shared their vision for 2025, and it's fascinating how these predictions align with addressing current challenges in the financial system:
1. AI Agents with Wallets
Imagine AI chatbots that can independently own and transact in crypto. This isn't just about technology - it's about creating autonomous financial actors that operate outside traditional banking constraints.
2. Proof of Personhood
In a world of deepfakes and AI-generated content, verifying real human identity becomes crucial. The industry is working on new ways to prove humanity while preserving privacy.
3. Enterprise Stablecoins
Major companies are expected to adopt stablecoins for payments - not just because they're the cheapest way to send dollars globally, but as a hedge against potential banking restrictions.
4. User Experience Revolution
Crypto is finally prioritizing accessibility over infrastructure. New crypto-native app stores will emerge, similar to the App Store but without centralized control.
5. Asset Tokenization
From government bonds to biometric data - more assets will move to blockchain. This creates a parallel financial system less dependent on traditional banking.
Interesting times ahead!
a16z crypto
A few of the things we’re excited about in crypto (2025) - a16z crypto
A list of crypto, blockchains, web3 trends for 2025 -- including AI x crypto, prediction markets, stablecoins, voting, and much more
OpenAI just officially released its Sora AI video generation model
Sora has its own interface, where users can:
— Organize and view their generated videos
— See other users’ prompts and featured content
Much like Midjourney’s web UI, this feed style will lead to some awesome inspiration and discoverability of effective prompts.
Sora can generate up to 20-sec videos, in several different aspect ratios.
Generation time was a previous concern with early Sora versions, and it appears OpenAI has gotten it down significantly.
Sora will be available today to Plus subscribers, with Pro users getting 10x usage and higher resolution.
The rollout looks to exclude the EU, UK, China at launch.
Sora has its own interface, where users can:
— Organize and view their generated videos
— See other users’ prompts and featured content
Much like Midjourney’s web UI, this feed style will lead to some awesome inspiration and discoverability of effective prompts.
Sora can generate up to 20-sec videos, in several different aspect ratios.
Generation time was a previous concern with early Sora versions, and it appears OpenAI has gotten it down significantly.
Sora will be available today to Plus subscribers, with Pro users getting 10x usage and higher resolution.
The rollout looks to exclude the EU, UK, China at launch.
Sora
Transform text and images into immersive videos. Animate stories, visualize ideas, and bring your concepts to life.
👍4
The Willow chip breakthrough is cool!
Google announced that Willow has reached 105 qubits with improved error rates.
This Google’s roadmap from the blog is even more exciting.
A few more years of progress can lead to commercially relevant applications for a quantum computer and maybe even applications we haven't conceived yet.
Google announced that Willow has reached 105 qubits with improved error rates.
This Google’s roadmap from the blog is even more exciting.
A few more years of progress can lead to commercially relevant applications for a quantum computer and maybe even applications we haven't conceived yet.
The cryptocurrency market has entered a transformative phase marked by unprecedented growth and institutional adoption.
Here's a comprehensive analysis of the key developments shaping the industry in 2024.
1. Record-Breaking Bitcoin Performance
Bitcoin has achieved remarkable milestones, shattering previous records with two significant price breakthroughs:
- March 2024: Surpassed $73,000
- December 2024: Exceeded $100,000
2. DeFi Renaissance
Decentralized Finance (DeFi) has experienced a remarkable resurgence, demonstrating its resilience and potential. The sector's revival began in late 2023, reaching new heights of activity and showcasing its capacity to reshape traditional financial services. This growth indicates DeFi's transition from an experimental concept to a legitimate financial infrastructure.
3. Stablecoin Dominance
Stablecoins have emerged as the backbone of cryptocurrency transactions, accounting for:
- 50-75% of all on-chain transaction volume
- Crucial financial tools in regions with unstable currencies
- Essential infrastructure for global digital commerce
Historic ETF Breakthrough
The launch of spot Bitcoin ETFs in U.S. markets marks a watershed moment for cryptocurrency adoption:
- Nearly $10 billion in daily trading volume (March 2024)
- Fastest-growing ETF launch in history
- Unprecedented institutional accessibility to cryptocurrency exposure
Real-World Asset (RWA) Tokenization
The tokenization of real-world assets represents a quiet revolution in asset management:
- Market capitalization exceeding $100 billion
- Major financial institutions like Franklin Templeton and Goldman Sachs entering the space
- Assets ranging from treasury bills to real estate and art
- Enhanced liquidity and accessibility for traditionally illiquid assets
Ecosystem Maturity Indicators
The cryptocurrency ecosystem shows clear signs of sustained growth:
- Over 400 million active cryptocurrency wallets
- Consistent growth despite market volatility
- Deeper integration with traditional financial systems
- Growing institutional participation and infrastructure development
Here's a comprehensive analysis of the key developments shaping the industry in 2024.
1. Record-Breaking Bitcoin Performance
Bitcoin has achieved remarkable milestones, shattering previous records with two significant price breakthroughs:
- March 2024: Surpassed $73,000
- December 2024: Exceeded $100,000
2. DeFi Renaissance
Decentralized Finance (DeFi) has experienced a remarkable resurgence, demonstrating its resilience and potential. The sector's revival began in late 2023, reaching new heights of activity and showcasing its capacity to reshape traditional financial services. This growth indicates DeFi's transition from an experimental concept to a legitimate financial infrastructure.
3. Stablecoin Dominance
Stablecoins have emerged as the backbone of cryptocurrency transactions, accounting for:
- 50-75% of all on-chain transaction volume
- Crucial financial tools in regions with unstable currencies
- Essential infrastructure for global digital commerce
Historic ETF Breakthrough
The launch of spot Bitcoin ETFs in U.S. markets marks a watershed moment for cryptocurrency adoption:
- Nearly $10 billion in daily trading volume (March 2024)
- Fastest-growing ETF launch in history
- Unprecedented institutional accessibility to cryptocurrency exposure
Real-World Asset (RWA) Tokenization
The tokenization of real-world assets represents a quiet revolution in asset management:
- Market capitalization exceeding $100 billion
- Major financial institutions like Franklin Templeton and Goldman Sachs entering the space
- Assets ranging from treasury bills to real estate and art
- Enhanced liquidity and accessibility for traditionally illiquid assets
Ecosystem Maturity Indicators
The cryptocurrency ecosystem shows clear signs of sustained growth:
- Over 400 million active cryptocurrency wallets
- Consistent growth despite market volatility
- Deeper integration with traditional financial systems
- Growing institutional participation and infrastructure development
Chainalysis
Growth Opportunities: Stablecoins, TradFi Engagement, and Tokenization Lead Web3’s Next Wave
With bitcoin prices at all-time highs, we explore how TradFi engagement, stablecoins and tokenization are driving web3 growth opportunities.
OpenAI Canvas (now out of beta and broadly available) can provide line-by-line comments.
It almost looks like an AI-superpowered Google Docs.
Canvas to become a more personalized tutor that you can interact in much richer ways.
You can ask ChatGPT to explain some mathematical concept, write code to plot that will help you learn it in a more visual way.
Canvas is a new way to collaborate with ChatGPT for writing and coding. Today it'll launch to everyone, with built-in python execution and integration with Custom GPTs.
It almost looks like an AI-superpowered Google Docs.
Canvas to become a more personalized tutor that you can interact in much richer ways.
You can ask ChatGPT to explain some mathematical concept, write code to plot that will help you learn it in a more visual way.
Canvas is a new way to collaborate with ChatGPT for writing and coding. Today it'll launch to everyone, with built-in python execution and integration with Custom GPTs.
👍7
Meta researchers introduced COCONUT
It's an interesting new AI reasoning approach that allows AI models to think more naturally rather than through rigid language steps, leading to better performance on complex problem-solving tasks.
It's an interesting new AI reasoning approach that allows AI models to think more naturally rather than through rigid language steps, leading to better performance on complex problem-solving tasks.
arXiv.org
Training Large Language Models to Reason in a Continuous Latent Space
Large language models (LLMs) are typically constrained to reason in the language space, where they express the reasoning process through a chain-of-thought (CoT) to solve complex problems....
👍2
CBInsights just dropped the 2024 Digital Health 50, spotlighting AI and advanced care.
Here are 4 interesting themes:
1. AI is becoming foundational infrastructure: 36 of 50 companies are building AI, from operational AI (Laguna) to LLMs (Hippocratic AI).
2. Workflow efficiency is a key priority
19 of the companies are streamlining admin/clinical tasks, from document tools like Tennr to ambient AI like Abridge.
The rise in automation shows the focus on efficiency, shifting providers from paperwork to patient care.
3. Diagnostic innovations dominate
11 companies comprised this year’s largest category, focusing on imaging (Airs Medical), pathology (Proscia), and non-invasive tests (Alimetry).
They aim to improve access, use non-invasive methods, and enable early detection.
4. More specialized platforms
Virtual and hybrid care representation doubled in this year’s cohort, reflecting the shift from general telemedicine toward condition-specific virtual models in areas like mental health with Talkiatry and cancer care startups such as Resilience.
Here are 4 interesting themes:
1. AI is becoming foundational infrastructure: 36 of 50 companies are building AI, from operational AI (Laguna) to LLMs (Hippocratic AI).
2. Workflow efficiency is a key priority
19 of the companies are streamlining admin/clinical tasks, from document tools like Tennr to ambient AI like Abridge.
The rise in automation shows the focus on efficiency, shifting providers from paperwork to patient care.
3. Diagnostic innovations dominate
11 companies comprised this year’s largest category, focusing on imaging (Airs Medical), pathology (Proscia), and non-invasive tests (Alimetry).
They aim to improve access, use non-invasive methods, and enable early detection.
4. More specialized platforms
Virtual and hybrid care representation doubled in this year’s cohort, reflecting the shift from general telemedicine toward condition-specific virtual models in areas like mental health with Talkiatry and cancer care startups such as Resilience.
Maya is a New Multimodal Multilingual Vision-Language Model.
Maya is completely open source, open weight and open dataset, designed to handle 8 languages, cultural diversity, and nuanced real-world contexts in vision-language models.
Hf.
Maya is completely open source, open weight and open dataset, designed to handle 8 languages, cultural diversity, and nuanced real-world contexts in vision-language models.
Hf.
huggingface.co
Paper page - Maya: An Instruction Finetuned Multilingual Multimodal Model
Join the discussion on this paper page
❗️Google Announces Gemini 2.0: A New Era of AI Agents
What's New in Gemini 2.0?
The new model brings groundbreaking capabilities to the table. Unlike its predecessor, Gemini 2.0 can not only understand multiple types of input but also generate various forms of output. It can create images, produce audio, and seamlessly integrate with tools like Google Search. Perhaps most impressively, it operates at twice the speed of the previous version while delivering superior performance.
3 Groundbreaking Projects
Google is showcasing Gemini 2.0's capabilities through three innovative projects:
1. Project Astra: An enhanced universal AI assistant that can understand multiple languages, remember conversations for up to 10 minutes, and naturally integrate with Google tools like Search, Lens, and Maps.
2. Project Mariner. A browser-based AI agent that can navigate web interfaces and complete complex tasks. Early tests show an impressive 83.5% success rate on real-world web tasks.
3. Jules: An AI-powered coding assistant that integrates with GitHub, helping developers plan and execute coding tasks under their supervision.
Google's ambitions extend beyond traditional computing interfaces. They're exploring applications in gaming, where AI agents can provide real-time guidance by analyzing gameplay, and even in robotics, where Gemini 2.0's spatial reasoning capabilities could revolutionize physical automation.
Understanding the significant implications of these advancements, Google emphasizes their commitment to responsible AI development. They've implemented extensive safety measures, including:
- Comprehensive risk assessments
- AI-assisted red teaming
- Privacy controls
- Protection against potential misuse
What's New in Gemini 2.0?
The new model brings groundbreaking capabilities to the table. Unlike its predecessor, Gemini 2.0 can not only understand multiple types of input but also generate various forms of output. It can create images, produce audio, and seamlessly integrate with tools like Google Search. Perhaps most impressively, it operates at twice the speed of the previous version while delivering superior performance.
3 Groundbreaking Projects
Google is showcasing Gemini 2.0's capabilities through three innovative projects:
1. Project Astra: An enhanced universal AI assistant that can understand multiple languages, remember conversations for up to 10 minutes, and naturally integrate with Google tools like Search, Lens, and Maps.
2. Project Mariner. A browser-based AI agent that can navigate web interfaces and complete complex tasks. Early tests show an impressive 83.5% success rate on real-world web tasks.
3. Jules: An AI-powered coding assistant that integrates with GitHub, helping developers plan and execute coding tasks under their supervision.
Google's ambitions extend beyond traditional computing interfaces. They're exploring applications in gaming, where AI agents can provide real-time guidance by analyzing gameplay, and even in robotics, where Gemini 2.0's spatial reasoning capabilities could revolutionize physical automation.
Understanding the significant implications of these advancements, Google emphasizes their commitment to responsible AI development. They've implemented extensive safety measures, including:
- Comprehensive risk assessments
- AI-assisted red teaming
- Privacy controls
- Protection against potential misuse
Google
Introducing Gemini 2.0: our new AI model for the agentic era
Today, we’re announcing Gemini 2.0, our most capable AI model yet.
🆒4
Google introduced Deep Research, your personal agentic AI research assistant. Rolling out starting today in Gemini Advanced.
With Deep Research, you can create in-depth research reports on complex topics, complete with source links, giving you hours of research at your fingertips in just minutes.
Visit gemini.google.com and select Deep Research from the Gemini Advanced model drop-down to get started.
With Deep Research, you can create in-depth research reports on complex topics, complete with source links, giving you hours of research at your fingertips in just minutes.
Visit gemini.google.com and select Deep Research from the Gemini Advanced model drop-down to get started.
Agents are going mainstream. Google launched an AI that takes over Chrome and does things for you.
It scores an insane 90.5% w/ tree search on the WebVoyager benchmark which has tasks like:
"Book a flight from Paris to Berlin, departing on March 5 and returning on the 12"
It scores an insane 90.5% w/ tree search on the WebVoyager benchmark which has tasks like:
"Book a flight from Paris to Berlin, departing on March 5 and returning on the 12"
Google DeepMind
Project Mariner
Project Mariner is a research prototype exploring the future of human-agent interaction, starting with browsers. It automates tasks to help boost productivity.
SubCell a suite of vision transformer models for microscopy to capture single-cell biology
SubCell, a suite of self-supervised deep learning models for fluorescence microscopy that are designed to accurately capture cellular morphology, protein localization, cellular organization, and biological function beyond what humans can readily perceive.
SubCell provides deep, image-driven representations of cellular architecture across biological contexts & datasets.
Researchers showcase the power of SubCell by constructing the first proteome-wide hierarchical map of proteome organization that is directly learned from image data.
This vision-based multiscale cell map defines cellular subsystems with large protein-complex resolution, reveals proteins with similar functions, and distinguishes dynamic and stable behaviors within cellular compartments.
Code, data and tutorial.
SubCell, a suite of self-supervised deep learning models for fluorescence microscopy that are designed to accurately capture cellular morphology, protein localization, cellular organization, and biological function beyond what humans can readily perceive.
SubCell provides deep, image-driven representations of cellular architecture across biological contexts & datasets.
Researchers showcase the power of SubCell by constructing the first proteome-wide hierarchical map of proteome organization that is directly learned from image data.
This vision-based multiscale cell map defines cellular subsystems with large protein-complex resolution, reveals proteins with similar functions, and distinguishes dynamic and stable behaviors within cellular compartments.
Code, data and tutorial.
bioRxiv
SubCell: Vision foundation models for microscopy capture single-cell biology
Cells are the functional units of life, and the wide range of biological functions they perform are orchestrated by myriad molecular interactions within an intricate subcellular architecture. This cellular organization and functionality can be studied with…
Google introduced Trillium: sixth-generation and most performant TPU to date
Google used Trillium TPUs to train the new Gemini 2.0, and now everyone can take advantage of the same powerful, efficient, and sustainable infrastructure.
Google used Trillium TPUs to train the new Gemini 2.0, and now everyone can take advantage of the same powerful, efficient, and sustainable infrastructure.
Google Cloud Blog
Trillium TPU is GA | Google Cloud Blog
Trillium, Google’s sixth-generation Tensor Processing Unit (TPU) is now GA, delivering enhanced performance and cost-effectiveness for AI workloads.
Microsoft Phi-4 is announced
Phi-4 is in Llama 3.3-70B category (win some lose some) with 5x fewer parameters, and notably outperforms on pure reasoning like GPQA (56%) and MATH (80%).
It's a 14B parameter LM trained heavily on synthetic data, with very strong performance, even exceeding GPT-4o on GPQA and MATH benchmarks!
Currently available on Azure AI Foundry, will be on HuggingFace next week.
Model.
Paper.
Phi-4 is in Llama 3.3-70B category (win some lose some) with 5x fewer parameters, and notably outperforms on pure reasoning like GPQA (56%) and MATH (80%).
It's a 14B parameter LM trained heavily on synthetic data, with very strong performance, even exceeding GPT-4o on GPQA and MATH benchmarks!
Currently available on Azure AI Foundry, will be on HuggingFace next week.
Model.
Paper.
TECHCOMMUNITY.MICROSOFT.COM
Introducing Phi-4: Microsoft’s Newest Small Language Model Specializing in Complex Reasoning | Microsoft Community Hub
Today we are introducing Phi-4, our 14B parameter state-of-the-art small language model (SLM) that excels at complex reasoning in areas such as math, in...
❤3🆒3
Stablecoins will eat payments
Today the payment landscape is dominated by gatekeepers & extractive networks - stifling competition, limiting the creativity of builders & taxing the profitability of every business
All businesses want to add 2% to their bottom line.
Stablecoin adoption likely won't start with the big business.
Intl payments are painful for small businesses. Intermediaries take up to $150 in fees to send $1000 from Mexico to Vietnam. A 1.5% improvement to transaction costs is substantial (and a ways off), but it’s possible because stablecoins enable more competitive, efficient payments
There are no gatekeepers on the blockchain + stablecoins are composable, programmable & cheap. It's much easier to build a payment platform with stablecoins.
Today the payment landscape is dominated by gatekeepers & extractive networks - stifling competition, limiting the creativity of builders & taxing the profitability of every business
All businesses want to add 2% to their bottom line.
Stablecoin adoption likely won't start with the big business.
Intl payments are painful for small businesses. Intermediaries take up to $150 in fees to send $1000 from Mexico to Vietnam. A 1.5% improvement to transaction costs is substantial (and a ways off), but it’s possible because stablecoins enable more competitive, efficient payments
There are no gatekeepers on the blockchain + stablecoins are composable, programmable & cheap. It's much easier to build a payment platform with stablecoins.
a16z crypto
How stablecoins will eat payments, and what happens next
Businesses that are most harmed by the current payment options will adopt stablecoins, disrupting the payments industry
❤7