Surfstudio podcast

Tracking new developments in markets, science, technology, politics and more, the annual 'The World Ahead' issue from The Economist gives you an unparalleled understanding of the major forces at play across the globe. Open up a mind map to quickly visualise the breadth of commentary and reporting in this notebook, or listen to an Audio Overview guiding you through the most interesting predictions for the year ahead from expert forecasters. Or, dive into whatever trends you need to understand by asking questions in the chat – you'll have some of the world's most respected analysis on tap to help you see the big picture.

Subscribe to get full access to The Economist's weekly edition, app, podcasts and more.

Get health and longevity advice by drawing on the expertise of bestselling author Eric Topol, professor at the Scripps Research Institute and one of the most cited researchers in all of medicine. Ask lifestyle questions about diet, sleep or exercise, and get detailed, science-backed recommendations, based on Topol's latest book, Super Agers and his popular Substack, Ground Truths. Listen to Audio Overviews explaining the importance of healthspan over lifespan and the new possibilities of an AI diet, or explore a mind map of all the longevity-related topics covered in the notebook.

For more essential health advice from Eric Topol, pick up a copy of his book Super Agers or subscribe to his Substack, Ground Truths.

Draw on the latest scientific research and the work of classic philosophers to help you lead a happier and more fulfilling life, with bestselling author and Atlantic columnist Arthur Brooks as your guide. Navigating the fine line of work-life balance or contemplating a change of careers or struggling to find meaning in your daily routine? You'll find thoughtful, science-backed wisdom in this notebook, based on Brooks' acclaimed 'How To Build A Life' column, paired with ancient wisdom from the Greeks. Listen to an Audio Overview on the latest science on human happiness or just use the chat to ask for advice on whatever life challenges you're facing.

Subscribe to The Atlantic to experience all of its award-winning reporting and writing on politics, society, culture, technology and more.

While both Mass and MaAS significantly advance the automation of multi-agent system (MAS) design, how do their core methodologies diverge in addressing the challenge of efficiently adapting MAS to queries of varying complexity and resource constraints, and what are the distinct benefits and trade-offs of each approach in achieving adaptive and cost-effective performance?

This study investigates Large Reasoning Models (LRMs), which generate "thinking processes" like Chain-of-Thought, using **controllable puzzle environments** to avoid data contamination and analyze "thinking traces".

Key findings reveal **three performance regimes**:
*   **Low complexity**: Standard LLMs often surprisingly outperform LRMs with greater token efficiency.
*   **Medium complexity**: LRMs show an advantage due to their "thinking" mechanisms.
*   **High complexity**: **Both LRMs and standard LLMs experience complete accuracy collapse**.

Counter-intuitively, LRMs **reduce their reasoning effort** (thinking tokens) as problems approach accuracy collapse, despite having ample budget. Furthermore, LRMs exhibit **limitations in exact computation and consistently following explicit algorithms**, as providing the algorithm did not improve performance. These findings suggest current LRMs face **fundamental barriers to generalizable and robust reasoning**.

Welcome to an episode that delves into the revolutionary concept of the Darwin Gödel Machine (DGM). We'll explore how today's AI systems typically operate with human-designed, fixed architectures, which fundamentally limits their capacity for autonomous and continuous self-improvement1.

Our discussion will focus on the DGM as a groundbreaking alternative: a self-improving system that iteratively modifies its own code1. Crucially, unlike the theoretical Gödel machine that aimed for provably beneficial changes, the DGM empirically validates each modification using coding benchmarks1.

Inspired by both Darwinian evolution and open-endedness research, the DGM operates by maintaining an archive of generated coding agents1. It continually expands this archive by sampling an existing agent and then utilizing a foundation model to create a new, interesting version of that agent1. This dynamic process allows for open-ended exploration, fostering a growing tree of diverse, high-quality agents and enabling the parallel exploration of many different improvement paths1.

We'll highlight the DGM's empirical successes, such as its ability to automatically improve its coding capabilities, including better code editing tools and long-context window management1. The DGM has demonstrated significant performance increases on benchmarks like SWE-bench (from 20.0% to 50.0%) and Polyglot (from 14.2% to 30.7%), markedly outperforming baselines that lack self-improvement or open-ended exploration1.

Finally, we'll touch upon the safety precautions taken during DGM experiments, such as sandboxing and human oversight, emphasizing that these are crucial considerations for developing such advanced systems1. Join us to understand why the DGM is considered a significant step toward truly self-improving AI, capable of paving its own way to endless innovation1.

Subject Area: It falls under Computer Science, specifically Artificial Intelligence (cs.AI) and Computation and Language (cs.CL)1.

•

Core Concept: The paper proposes a "Self-Challenging framework" for training intelligent agents1. This framework allows an agent to generate its own high-quality tasks1.

•

Task Generation: The agent acts as a "challenger" to generate tasks after interacting with given tools. These tasks are defined as a novel class of problems called "Code-as-Task," which include an instruction, a verification function, and solution/failure cases that act as tests to filter for high-quality tasks1.

•

Training: After generating tasks, the agent takes on an "executor role" and trains on these self-generated tasks using reinforcement learning, with evaluation feedback serving as a reward1.

•

Performance: The Self-Challenging framework showed significant improvement (over two-fold) in Llama-3.1-8B-Instruct on two existing multi-turn tool-use agent benchmarks, M3ToolEval and TauBench, despite using only self-generated training data1.

•

Availability: The paper, submitted on June 2, 2025, is available in PDF and experimental HTML formats12.

•

Associated Tools/Platforms: The sources also mention various bibliographic tools, citation tools, and platforms associated with the paper, such as NASA ADS, Google Scholar, Semantic Scholar, alphaXiv, CatalyzeX Code Finder, DagsHub, Huggingface, Papers with Code, ScienceCast, Replicate, TXYZ.AI, Influence Flower, and CORE Recommender34. Some of these are part of arXivLabs, which supports experimental projects45.

 Welcome to an episode where we explore SmolVLA, a groundbreaking development in the field of robotics that's making advanced AI more accessible and affordable1.

Traditional Vision-Language-Action (VLA) models, which enable natural language-driven perception and control for robots, are often massive, typically boasting billions of parameters1. This leads to significant challenges like high training costs and limited practical deployment in real-world scenarios1. These larger models also tend to rely on academic and industrial datasets, overlooking the valuable, growing pool of community-collected data from more affordable robotic platforms1.

SmolVLA offers a powerful solution to these challenges1. It's designed as a small, efficient, and community-driven VLA model that drastically cuts down on both training and inference expenses1. A key innovation is its ability to be trained on a single GPU and deployed on widely available consumer-grade GPUs or even standard CPUs, making it incredibly accessible for a broader range of users and applications1.

Furthermore, SmolVLA introduces an asynchronous inference stack, which enhances responsiveness by separating the processes of perception and action prediction from the actual execution of actions1. This allows for higher control rates through the generation of action chunks1. Despite its compact size, SmolVLA demonstrates performance comparable to VLAs that are ten times larger, proving that efficiency doesn't have to compromise capability1. This model has been rigorously evaluated on both simulated and real-world robotic benchmarks, with its code, pretrained models, and training data openly released to foster community engagement1. Join us as we delve into how SmolVLA is paving the way for more affordable and efficient robotics, bringing advanced AI control to a wider audience.

Foundation-Sec-8B is a cybersecurity-specialized large language model (LLM) developed by Foundation AI – Cisco Systems Inc., built on the Llama 3.1 architecture.
Key aspects of Foundation-Sec-8B:
•
Purpose: It was created to address limitations in the adoption of LLMs within cybersecurity, such as the lack of specialized training data and the complexity of representing cybersecurity knowledge. The goal is to accelerate the progress and adoption of AI-driven tools in both public and private cybersecurity contexts.
•
Development Process: The model was enhanced through continued pretraining on a carefully curated cybersecurity corpus. This involved a multi-stage data collection and processing pipeline, starting with 4 TiB of raw web content, which was then filtered, cleaned, and tokenized to produce approximately 25 GiB (about 5 billion tokens) of high-quality training data.
•
Performance: Foundation-Sec-8B demonstrates significant performance gains over its base model, Llama 3.1-8B, across various cybersecurity benchmarks. It matches Llama 3.1-70B and GPT-4o-mini in certain cybersecurity-specific tasks, particularly in cyber threat intelligence (CTI) knowledge, and achieves state-of-the-art performance for its size class on selected CTI tasks. It also retains general knowledge without severe degradation, consistent with prior research on continued pretraining.
•
Use Cases: It is being piloted or deployed for real-world applications across the security lifecycle, including:
◦
SOC Acceleration: Automating tasks like summarizing alerts, generating incident timelines, and drafting analyst reports.
◦
Proactive Threat Defense: Extracting Tactics, Techniques, and Procedures (TTPs), prioritizing vulnerabilities, generating attack path hypotheses, and drafting penetration test reports.
◦
Engineering Enablement: Helping teams interpret security policies, validate configurations, assess compliance evidence, and analyze security policies.
•
Availability: Foundation-Sec-8B is being released publicly to foster broader experimentation, advancement, and practical deployment of AI-driven security tools within the cybersecurity and AI research communities.

Foundation-Sec-8B is a cybersecurity-specialized large language model (LLM) developed by Foundation AI – Cisco Systems Inc., built on the Llama 3.1 architecture12.

Key aspects of Foundation-Sec-8B:

•

Purpose: It was created to address limitations in the adoption of LLMs within cybersecurity, such as the lack of specialized training data and the complexity of representing cybersecurity knowledge13. The goal is to accelerate the progress and adoption of AI-driven tools in both public and private cybersecurity contexts14.

•

Development Process: The model was enhanced through continued pretraining on a carefully curated cybersecurity corpus1. This involved a multi-stage data collection and processing pipeline, starting with 4 TiB of raw web content, which was then filtered, cleaned, and tokenized to produce approximately 25 GiB (about 5 billion tokens) of high-quality training data5....

•

Performance: Foundation-Sec-8B demonstrates significant performance gains over its base model, Llama 3.1-8B, across various cybersecurity benchmarks2. It matches Llama 3.1-70B and GPT-4o-mini in certain cybersecurity-specific tasks, particularly in cyber threat intelligence (CTI) knowledge, and achieves state-of-the-art performance for its size class on selected CTI tasks110. It also retains general knowledge without severe degradation, consistent with prior research on continued pretraining11.

•

Use Cases: It is being piloted or deployed for real-world applications across the security lifecycle, including:

◦

SOC Acceleration: Automating tasks like summarizing alerts, generating incident timelines, and drafting analyst reports1213.

◦

Proactive Threat Defense: Extracting Tactics, Techniques, and Procedures (TTPs), prioritizing vulnerabilities, generating attack path hypotheses, and drafting penetration test reports1314.

◦

Engineering Enablement: Helping teams interpret security policies, validate configurations, assess compliance evidence, and analyze security policies1415.

•

Availability: Foundation-Sec-8B is being released publicly to foster broader experimentation, advancement, and practical deployment of AI-driven security tools within the cybersecurity and AI research communities14.