Quantum Computing. The Banque de France and the Monetary Authority of Singapore have announced the successful completion of a groundbreaking joint experiment in post-quantum cryptography (#PQC) conducted across continents over conventional Internet technologies.
The PQC experiment aims to strengthen communication & data security in the face of quantum computing advancements, and the successful experimentation marks a crucial milestone in the evolution of the protection of international electronic communications against the cybersecurity threats posed by quantum computing.
Using Microsoft Outlook as the email client coupled with a PQC email plugin, BdF and MAS successfully exchanged digitally-signed and encrypted emails using PQC algorithms, namely CRYSTALS-Dilithium and CRYSTALS-Kyber.
The PQC experiment aims to strengthen communication & data security in the face of quantum computing advancements, and the successful experimentation marks a crucial milestone in the evolution of the protection of international electronic communications against the cybersecurity threats posed by quantum computing.
Using Microsoft Outlook as the email client coupled with a PQC email plugin, BdF and MAS successfully exchanged digitally-signed and encrypted emails using PQC algorithms, namely CRYSTALS-Dilithium and CRYSTALS-Kyber.
Salesforce introduced CRMArena - a work-oriented benchmark for LLM agents to prove their mettle in real-world business scenarios
CRMArena features nine distinct tasks within a complex business environment filled with rich and realistic data, all validated by domain experts.
Code.
Leaderboard.
CRMArena features nine distinct tasks within a complex business environment filled with rich and realistic data, all validated by domain experts.
Code.
Leaderboard.
arXiv.org
CRMArena: Understanding the Capacity of LLM Agents to Perform...
Customer Relationship Management (CRM) systems are vital for modern enterprises, providing a foundation for managing customer interactions and data. Integrating AI agents into CRM systems can...
MIT introduced DART: breaking the barriers for robotic data collection by enabling anyone, anywhere in the world to control robots without even having a robot.
No need for resets or setting environments
Support for multiple robots
Robot bootstraps and autonomously collects data in simulation while you sleep!
This is the first step towards a fully crowdsourced and open-source foundation model for robotics.
No need for resets or setting environments
Support for multiple robots
Robot bootstraps and autonomously collects data in simulation while you sleep!
This is the first step towards a fully crowdsourced and open-source foundation model for robotics.
🆒2
A Comprehensive Survey of Small Language Models
Nice survey on small language models (SLMs) and discussion on issues related to definitions, applications, enhancements, reliability, and more.
What are SLMs?
Think of them as compact versions of large language models (like GPT-4), typically with fewer than 7 billion parameters. They're designed to be efficient while maintaining impressive capabilities.
Key Advantages:
• Run directly on mobile devices
• Better privacy (no cloud required)
• Lower computational costs
• Faster response times
• More energy-efficient
Use Cases:
• Question answering
• Code generation
• Recommendation systems
• Web search
• Mobile applications
• Domain-specific tasks
Why They're Revolutionary:
1. Privacy First: Process data locally without sending it to the cloud
2. Accessibility: Work on standard hardware without expensive GPUs
3. Cost-Effective: Lower operational costs for businesses
4. Eco-Friendly: Reduced energy consumption
Future Potential:
• Enhanced efficiency through specialized architectures
• Broader adoption in mobile apps
• Improved performance in specific domains
• Better integration with larger models
Nice survey on small language models (SLMs) and discussion on issues related to definitions, applications, enhancements, reliability, and more.
What are SLMs?
Think of them as compact versions of large language models (like GPT-4), typically with fewer than 7 billion parameters. They're designed to be efficient while maintaining impressive capabilities.
Key Advantages:
• Run directly on mobile devices
• Better privacy (no cloud required)
• Lower computational costs
• Faster response times
• More energy-efficient
Use Cases:
• Question answering
• Code generation
• Recommendation systems
• Web search
• Mobile applications
• Domain-specific tasks
Why They're Revolutionary:
1. Privacy First: Process data locally without sending it to the cloud
2. Accessibility: Work on standard hardware without expensive GPUs
3. Cost-Effective: Lower operational costs for businesses
4. Eco-Friendly: Reduced energy consumption
Future Potential:
• Enhanced efficiency through specialized architectures
• Broader adoption in mobile apps
• Improved performance in specific domains
• Better integration with larger models
arXiv.org
A Comprehensive Survey of Small Language Models in the Era of...
Large language models (LLMs) have demonstrated emergent abilities in text generation, question answering, and reasoning, facilitating various tasks and domains. Despite their proficiency in...
eBook-How-to-Build-a-Career-in-AI.pdf
3.5 MB
Key Insights from Andrew Ng's "How to Build Your Career in AI"
Andrew Ng, the founder of DeepLearning.AI, shares his comprehensive guide on building a successful career in AI.
Here are the essential takeaways:
🎯 Three Core Steps to Career Growth:
Learning foundational skills
Working on projects
Finding the right job
🧠 Must-Have Technical Skills:
Machine Learning fundamentals
Deep Learning
Software Development
Mathematics (Linear Algebra, Statistics, Probability)
Data structures and algorithms
📚 Project Development Strategy:
Start small with learning projects
Graduate to personal projects
Build value-creating solutions
Show progression in complexity
Create a strong portfolio
💼 Job Search Tips:
Use informational interviews
Build a supportive network
Focus on one transition at a time (either role or industry)
Choose great teammates over exciting projects
Pay attention to company culture
🌟 Keys to Success:
Embrace teamwork
Build genuine connections
Maintain personal discipline
Practice continuous learning
Help others grow
💭 Overcoming Imposter Syndrome:
Remember: 70% of people experience it
Focus on your strengths
Find supportive mentors
Celebrate small wins
Keep learning and growing
Andrew Ng, the founder of DeepLearning.AI, shares his comprehensive guide on building a successful career in AI.
Here are the essential takeaways:
🎯 Three Core Steps to Career Growth:
Learning foundational skills
Working on projects
Finding the right job
🧠 Must-Have Technical Skills:
Machine Learning fundamentals
Deep Learning
Software Development
Mathematics (Linear Algebra, Statistics, Probability)
Data structures and algorithms
📚 Project Development Strategy:
Start small with learning projects
Graduate to personal projects
Build value-creating solutions
Show progression in complexity
Create a strong portfolio
💼 Job Search Tips:
Use informational interviews
Build a supportive network
Focus on one transition at a time (either role or industry)
Choose great teammates over exciting projects
Pay attention to company culture
🌟 Keys to Success:
Embrace teamwork
Build genuine connections
Maintain personal discipline
Practice continuous learning
Help others grow
💭 Overcoming Imposter Syndrome:
Remember: 70% of people experience it
Focus on your strengths
Find supportive mentors
Celebrate small wins
Keep learning and growing
BlackRock’s Bitcoin ETF Achieves Record $1.1 Billion in Single-Day Inflows
BlackRock’s iShares Bitcoin Trust (IBIT) has set a new benchmark with over $1.1 billion in net inflows recorded on Thursday, marking the highest single-day inflows for the fund.
This surge comes on the heels of a broader trend, as the 12 U.S. spot Bitcoin ETFs collectively reported total daily net inflows of $1.38 billion—also a record since their inception in January.
BlackRock’s iShares Bitcoin Trust (IBIT) has set a new benchmark with over $1.1 billion in net inflows recorded on Thursday, marking the highest single-day inflows for the fund.
This surge comes on the heels of a broader trend, as the 12 U.S. spot Bitcoin ETFs collectively reported total daily net inflows of $1.38 billion—also a record since their inception in January.
CoinMarketCap Academy
BlackRock’s Bitcoin ETF Achieves Record $1.1 Billion in Single-Day Inflows
BlackRock’s iShares Bitcoin Trust (IBIT) has set a new benchmark with over $1.1 billion in net inflows recorded on Thursday, marking the highest single-day inflows for the fund.
Nvidia unveiled new updates to Project GR00T, it's comprehensive humanoid robot development suite
It includes environment generation with 2,500+ 3D assets, motion learning, and advanced dexterity training
All trained in sim, deployable to real robots.
It includes environment generation with 2,500+ 3D assets, motion learning, and advanced dexterity training
All trained in sim, deployable to real robots.
NVIDIA Technical Blog
Advancing Humanoid Robot Sight and Skill Development with NVIDIA Project GR00T
Humanoid robots present a multifaceted challenge at the intersection of mechatronics, control theory, and AI. The dynamics and control of humanoid robots are complex, requiring advanced tools…
Coinbase and Langchain AI introduced Agentkit: a production-ready, model-agnostic framework for building AI agents with infinite onchain and web2 functionality
Based Agents on base were just the beginning. it’s time to change the way we interact onchain.
Langchain is a powerful framework with pre-existing integrations for all sorts of web2 APIs, including Gmail, browsing the internet, X, and more.
Imagine the potential of combining these APIs with onchain, autonomous actions, like this demo integrating Aave labs using AI tools.
What’s next?
Over the following weeks, team’ll be releasing Agentkit templates to motivate and inspire some of the use cases.
GitHub
Replit.
Based Agents on base were just the beginning. it’s time to change the way we interact onchain.
Langchain is a powerful framework with pre-existing integrations for all sorts of web2 APIs, including Gmail, browsing the internet, X, and more.
Imagine the potential of combining these APIs with onchain, autonomous actions, like this demo integrating Aave labs using AI tools.
What’s next?
Over the following weeks, team’ll be releasing Agentkit templates to motivate and inspire some of the use cases.
GitHub
Replit.
Coinbase Developer Documentation
Welcome to AgentKit - Coinbase Developer Documentation
Epoch AI launched FrontierMath, a benchmark for evaluating advanced mathematical reasoning in AI
Existing math benchmarks like GSM8K and MATH are approaching saturation, with AI models scoring over 90%—partly due to data contamination.
FrontierMath significantly raises the bar.
Existing math benchmarks like GSM8K and MATH are approaching saturation, with AI models scoring over 90%—partly due to data contamination.
FrontierMath significantly raises the bar.
Epoch AI
FrontierMath
FrontierMath is a benchmark of hundreds of unpublished and extremely challenging math problems to help us to understand the limits of artificial intelligence.
DeFi. Governor Christopher J. Waller of the Federal Reserve Board recently gave a speech on ‘Centralized and Decentralized Finance: Substitutes or Complements?’ at the Vienna Macroeconomics Workshop, Institute of Advanced Studies, Vienna, Austria.
Key Takeaways of the speech:
1. DeFi allows asset trading without intermediaries, distinguishing it from centralized finance, yet it also has applications that complement traditional finance.
2. DLT offers faster and more efficient recordkeeping, useful for 24/7 markets, and is being explored by traditional financial institutions.
3. Tokenizing assets and using DLT can speed up transactions and enable automated, secure trading through #smartcontracts, reducing #settlement and counterparty risks.
4. Smart contracts streamline transactions by automating multiple steps, enhancing #security and #efficiency in both #DeFi and #centralizedfinance.
5. Stablecoins, typically pegged to $, facilitate decentralized trading and have potential in reducing global #payment costs, though they require regulatory safeguards to address #risks.
6. DeFi poses unique risks, including the potential for funds to reach bad actors, raising questions about the need for #regulations similar to those in traditional finance.
7. DeFi technologies can enhance centralized finance by improving efficiency, benefiting households and businesses through a more effective financial system.
Key Takeaways of the speech:
1. DeFi allows asset trading without intermediaries, distinguishing it from centralized finance, yet it also has applications that complement traditional finance.
2. DLT offers faster and more efficient recordkeeping, useful for 24/7 markets, and is being explored by traditional financial institutions.
3. Tokenizing assets and using DLT can speed up transactions and enable automated, secure trading through #smartcontracts, reducing #settlement and counterparty risks.
4. Smart contracts streamline transactions by automating multiple steps, enhancing #security and #efficiency in both #DeFi and #centralizedfinance.
5. Stablecoins, typically pegged to $, facilitate decentralized trading and have potential in reducing global #payment costs, though they require regulatory safeguards to address #risks.
6. DeFi poses unique risks, including the potential for funds to reach bad actors, raising questions about the need for #regulations similar to those in traditional finance.
7. DeFi technologies can enhance centralized finance by improving efficiency, benefiting households and businesses through a more effective financial system.
Board of Governors of the Federal Reserve System
Centralized and Decentralized Finance: Substitutes or Complements?
Thank you for inviting me to speak today.1 I have participated in this conference for nearly 20 years and have often presented my research on monetary theory, banking, and payments. So, I believe this is the right audience to speak to regarding the role of…
Alibaba introduced Qwen2.5-Coder-32B-Instruct: A New Era in AI Coding
Meet the groundbreaking family of coding models that's revolutionizing AI-assisted programming!
The results are nothing short of incredible.
The flagship Qwen2.5-Coder-32B-Instruct achieves remarkable benchmark scores:
✨ HumanEval: 92.7
✨ MBPP: 86.8
✨ CodeArena: 68.9
✨ LiveCodeBench: 31.4
Key highlights that make it special:
- Outperforms GPT-4 in several benchmarks!
- Available in multiple sizes: 0.5B, 1.5B, 3B, 7B, 14B, and 32B
- Supports popular quantization formats: GPTQ, AWQ, GGUF
- Seamless integration with Ollama for local deployment
- Fully open source
Get your hands on it now:
📍 Hugging Face
📍 ModelScope
📍 Kaggle
📍 GitHub
Meet the groundbreaking family of coding models that's revolutionizing AI-assisted programming!
The results are nothing short of incredible.
The flagship Qwen2.5-Coder-32B-Instruct achieves remarkable benchmark scores:
✨ HumanEval: 92.7
✨ MBPP: 86.8
✨ CodeArena: 68.9
✨ LiveCodeBench: 31.4
Key highlights that make it special:
- Outperforms GPT-4 in several benchmarks!
- Available in multiple sizes: 0.5B, 1.5B, 3B, 7B, 14B, and 32B
- Supports popular quantization formats: GPTQ, AWQ, GGUF
- Seamless integration with Ollama for local deployment
- Fully open source
Get your hands on it now:
📍 Hugging Face
📍 ModelScope
📍 Kaggle
📍 GitHub
Qwen
Qwen2.5-Coder Series: Powerful, Diverse, Practical.
GITHUB HUGGING FACE MODELSCOPE KAGGLE DEMO DISCORD
Introduction Today, we are excited to open source the “Powerful”, “Diverse”, and “Practical” Qwen2.5-Coder series, dedicated to continuously promoting the development of Open CodeLLMs.
Powerful: Qwen2.5-Coder…
Introduction Today, we are excited to open source the “Powerful”, “Diverse”, and “Practical” Qwen2.5-Coder series, dedicated to continuously promoting the development of Open CodeLLMs.
Powerful: Qwen2.5-Coder…
AlphaFold 3 is now open source!
AlphaFold 3 is a revolutionary AI model developed by Google DeepMind and Isomorphic Labs that can predict the 3D structures and interactions of virtually all biological molecules (including proteins, DNA, RNA, and small molecules) with crazy accuracy, achieving at least 50% improvement over previous existing methods.
AlphaFold 3 is a revolutionary AI model developed by Google DeepMind and Isomorphic Labs that can predict the 3D structures and interactions of virtually all biological molecules (including proteins, DNA, RNA, and small molecules) with crazy accuracy, achieving at least 50% improvement over previous existing methods.
GitHub
GitHub - google-deepmind/alphafold3: AlphaFold 3 inference pipeline.
AlphaFold 3 inference pipeline. Contribute to google-deepmind/alphafold3 development by creating an account on GitHub.
Justin Drake proposed a new consensus layer upgrade proposal "Beam Chain" at the Devcon conference, which is called "Ethereum 3.0" by the community.
The proposal aims to achieve faster block times, lower validator staking requirements, "chain snarkifaction" and quantum security improvements.
It is expected to formulate specifications in 2025 and enter the full testing phase in 2027.
The proposal aims to achieve faster block times, lower validator staking requirements, "chain snarkifaction" and quantum security improvements.
It is expected to formulate specifications in 2025 and enter the full testing phase in 2027.
The Block
Justin Drake proposes 'Beam Chain,' an Ethereum consensus layer redesign
Ethereum Foundation researcher Justin Drake announced Beam Chain, a new consensus layer upgrade proposal for Ethereum at Devcon on Tuesday.
Sanofi, OpenAI, Formation debut patient recruiting tool, will use in Phase 3 multiple sclerosis studies
Muse is AI tool for patient recruitment strategy & content creation.
AI systems like Muse will enable to drastically reduce cost + time of bringing new medicines to patients.
Muse is AI tool for patient recruitment strategy & content creation.
AI systems like Muse will enable to drastically reduce cost + time of bringing new medicines to patients.
PR Newswire
Formation Bio collaborates with Sanofi and OpenAI to Introduce Muse, a first of its kind AI tool to accelerate patient recruitment…
/PRNewswire/ -- Today, Formation Bio, together with OpenAI and Sanofi, introduced Muse, an advanced AI-powered tool developed to accelerate and improve drug...
👍3
New paper on scaling laws in primate vision modeling
Researchers trained and analyzed 600+ neural networks to understand how bigger models & more data affect brain predictivity.
Researchers trained and analyzed 600+ neural networks to understand how bigger models & more data affect brain predictivity.
2411.04330v1.pdf
1.4 MB
Precision-Aware Scaling Laws: A New Perspective on Language Model Training and Inference
A groundbreaking paper from researchers at Harvard, Stanford, MIT, and CMU reveals crucial insights into the relationship between model precision, training data, and performance in language models.
Key Findings:
1. Post-Training Quantization Challenge
The researchers discovered a counterintuitive phenomenon: models trained on more data become increasingly sensitive to post-training quantization. This means that after a certain point, additional training data can actually harm performance if the model will be quantized for inference.
2. Optimal Training Precision
The study suggests that the current standard practice of training in 16-bit precision may be suboptimal. Their analysis indicates that 7-8 bits might be the sweet spot for training, challenging both current high-precision (16-bit) and ultra-low precision (4-bit) approaches.
3. Unified Scaling Law
The team developed a comprehensive scaling law that accounts for:
- Training precision effects
- Post-training quantization impacts
- Interactions between model size, data, and precision
4. Practical Implications
- Larger models can be trained effectively in lower precision
- The race to extremely low-precision training (sub-4-bit) may face fundamental limitations
- There's an optimal precision point that balances performance and computational efficiency
5. Methodology
The research is backed by extensive experimentation, including:
- 465+ pretraining runs
- Models up to 1.7B parameters
- Training datasets up to 26B tokens
This work provides valuable insights for ML engineers and researchers working on large language models, suggesting that precision choices should be carefully considered based on model size and training data volume rather than following a one-size-fits-all approach.
The findings have significant implications for future hardware design and training strategies, potentially influencing how we approach model scaling and efficiency optimization in the AI field.
A groundbreaking paper from researchers at Harvard, Stanford, MIT, and CMU reveals crucial insights into the relationship between model precision, training data, and performance in language models.
Key Findings:
1. Post-Training Quantization Challenge
The researchers discovered a counterintuitive phenomenon: models trained on more data become increasingly sensitive to post-training quantization. This means that after a certain point, additional training data can actually harm performance if the model will be quantized for inference.
2. Optimal Training Precision
The study suggests that the current standard practice of training in 16-bit precision may be suboptimal. Their analysis indicates that 7-8 bits might be the sweet spot for training, challenging both current high-precision (16-bit) and ultra-low precision (4-bit) approaches.
3. Unified Scaling Law
The team developed a comprehensive scaling law that accounts for:
- Training precision effects
- Post-training quantization impacts
- Interactions between model size, data, and precision
4. Practical Implications
- Larger models can be trained effectively in lower precision
- The race to extremely low-precision training (sub-4-bit) may face fundamental limitations
- There's an optimal precision point that balances performance and computational efficiency
5. Methodology
The research is backed by extensive experimentation, including:
- 465+ pretraining runs
- Models up to 1.7B parameters
- Training datasets up to 26B tokens
This work provides valuable insights for ML engineers and researchers working on large language models, suggesting that precision choices should be carefully considered based on model size and training data volume rather than following a one-size-fits-all approach.
The findings have significant implications for future hardware design and training strategies, potentially influencing how we approach model scaling and efficiency optimization in the AI field.
❤5👍2
❗️Endocisternal minimally invasive neural interfaces
In a first-of-its-kind demonstration, researchers from The University of Texas Medical Branch and Rice University a wireless neural interface through a cistern, a space filled with CSF that provides an alternative to endovascular delivery.
They also showed neuromodulation, recording, and explantation!
In a first-of-its-kind demonstration, researchers from The University of Texas Medical Branch and Rice University a wireless neural interface through a cistern, a space filled with CSF that provides an alternative to endovascular delivery.
They also showed neuromodulation, recording, and explantation!
Nature
Endocisternal interfaces for minimally invasive neural stimulation and recording of the brain and spinal cord
Nature Biomedical Engineering - Endocisternal neural interfaces with wireless and magnetoelectrically powered electronics that approach brain and spinal cord targets through spaces filled with...
⚡4
Can a tiny startup’s 70 billion parameter model beat OpenAI’s o1 model?
Nous Research just launched the Forge Reasoning Engine, and it even managed to beat o1 on the American Invitational Math Exam.
Forge uses a combination of:
A) Monte Carlo Tree Search
B) Chain of Code
C) Mixture of Agents
D) Code Intepreter use
to get Nous’ Hermes 70B model close to o1’s performance on several math and science benchmarks.
This is a significant development as it is one of the first inference time scaling releases post o1 release.
They also point out that Forge allows “advancement in inference time scaling that can be applied to any model or a set of models”.
This means that they can swap out and upgrade the LLM piece over time, while keeping the rest of the Engine constant.
Nous is famous in the open source community for having released some of the best early open source fine tunes in 2023 and 2024.
Forge though is not open sourced, and is currently available via API to a small group of beta testers.
It is interesting to note that fairly small models maybe able to scale to the intelligence of very large models just by taking the time to think more at inference.
Inference time compute may finally level the playing field between the GPU poor and GPU rich.
Try Nous Chat today here for free.
Nous Research just launched the Forge Reasoning Engine, and it even managed to beat o1 on the American Invitational Math Exam.
Forge uses a combination of:
A) Monte Carlo Tree Search
B) Chain of Code
C) Mixture of Agents
D) Code Intepreter use
to get Nous’ Hermes 70B model close to o1’s performance on several math and science benchmarks.
This is a significant development as it is one of the first inference time scaling releases post o1 release.
They also point out that Forge allows “advancement in inference time scaling that can be applied to any model or a set of models”.
This means that they can swap out and upgrade the LLM piece over time, while keeping the rest of the Engine constant.
Nous is famous in the open source community for having released some of the best early open source fine tunes in 2023 and 2024.
Forge though is not open sourced, and is currently available via API to a small group of beta testers.
It is interesting to note that fairly small models maybe able to scale to the intelligence of very large models just by taking the time to think more at inference.
Inference time compute may finally level the playing field between the GPU poor and GPU rich.
Try Nous Chat today here for free.
NOUS RESEARCH
Introducing the Forge Reasoning API Beta and Nous Chat: An Evolution in LLM Inference - NOUS RESEARCH
The Forge Reasoning API contains some of our latest advancements in inference-time AI research, building on our journey from the original Hermes model.
🔥7❤1
Supermaven is merging with Cursor
This union brings together two innovative forces in AI-powered development tools.
Who is Supermaven? Founded by Jacob Jackson, the pioneer behind Tabnine (2019) and former OpenAI innovator, Supermaven has developed a lightning-fast, context-aware AI coding assistant that's been pushing the boundaries of what's possible in development tools.
Why does this matter?
• Combined expertise in AI and development tools
• Faster delivery of innovative features
• Shared vision for revolutionizing software development
• Enhanced capabilities through unified technologies
What's next?
The teams are already working on exciting improvements, including a next-generation Tab model featuring:
• Enhanced speed and responsiveness
• Superior context awareness
• Advanced intelligence for handling complex code changes
This union brings together two innovative forces in AI-powered development tools.
Who is Supermaven? Founded by Jacob Jackson, the pioneer behind Tabnine (2019) and former OpenAI innovator, Supermaven has developed a lightning-fast, context-aware AI coding assistant that's been pushing the boundaries of what's possible in development tools.
Why does this matter?
• Combined expertise in AI and development tools
• Faster delivery of innovative features
• Shared vision for revolutionizing software development
• Enhanced capabilities through unified technologies
What's next?
The teams are already working on exciting improvements, including a next-generation Tab model featuring:
• Enhanced speed and responsiveness
• Superior context awareness
• Advanced intelligence for handling complex code changes
Cursor
Supermaven joins Cursor
We’re teaming up to build the next phase of AI coding.
❤4🤔1