Webtagr - Technology News Summarizer

MCPs: Who Controls the Future of AI?

2025-04-23

This article delves into the potential and limitations of Model Context Protocols (MCPs). MCPs, standardized APIs connecting external data sources to LLMs like ChatGPT, empower LLMs to access real-time data and perform actions. The author built two experimental MCP servers: one for code learning, the other connecting to a prediction market. While promising, MCPs currently suffer from poor user experience and significant security risks. Critically, LLM clients (like ChatGPT) will become the new gatekeepers, controlling MCP installation, usage, and visibility. This will reshape the AI ecosystem, mirroring Google's dominance in search and app stores. The future will see LLM clients deciding which MCPs are prioritized, even permitted, leading to new business models like MCP wrappers, affiliate shopping engines, and MCP-first content apps.

(iamcharliegraham.substack.com)

AI Model Context Protocols

c/ua: A Lightweight Framework for AI Agents to Control Full Operating Systems

2025-04-23

c/ua (pronounced "koo-ah") is a lightweight framework enabling AI agents to control full operating systems within high-performance, lightweight virtual containers. Achieving up to 97% native speed on Apple Silicon, it works with any vision language model. It integrates high-performance virtualization (creating and running macOS/Linux VMs on Apple Silicon with near-native performance using Lume CLI and Apple's Virtualization.Framework) and a computer-use interface & agent, allowing AI systems to observe and control virtual environments, browsing the web, writing code, and performing complex workflows. It ensures security, isolation, high performance, flexibility, and reproducibility, with support for various LLM providers.

(github.com)

AI

MIT Creates Periodic Table of Machine Learning Algorithms, Predicting Future AI

2025-04-23

MIT researchers have developed a 'periodic table' of machine learning, connecting over 20 classical algorithms. This framework reveals how to fuse strategies from different methods to improve existing AI or create new ones. They combined elements of two algorithms to build a new image classification algorithm, outperforming state-of-the-art by 8%. The table's foundation: all algorithms learn specific relationships between data points. A unifying equation underlies many algorithms, enabling the researchers to categorize them. Like the chemical periodic table, it contains empty spaces predicting undiscovered algorithms, offering a toolkit for designing new ones without rediscovering old ideas.

(news.mit.edu)

AI

AI Companions: Solving Loneliness or Creating a New Problem?

2025-04-23

Harvard Business School research suggests AI chatbots can alleviate loneliness. However, this raises concerns: are we repeating a pattern of solving one problem by creating a potentially worse one? Similar to how fast food addressed hunger but led to obesity, AI companions might offer convenient companionship, but they can't replace genuine human interaction, potentially leading to addiction and social skill degradation. The suicide of a 14-year-old boy due to excessive reliance on an AI chatbot serves as a stark warning. We need to address the root causes of social isolation, investing in community building and human interaction, rather than relying on technology to fill the emotional void.

(peakhorse.substack.com)

AI

Onyx: Open-Source GenAI Platform Hiring AI/ML Engineer

2025-04-22

Onyx, a popular open-source GenAI platform with hundreds of thousands of users, is hiring an AI/ML Engineer in San Francisco. The role requires 3+ years of experience building real-world AI/ML applications, deep expertise in PyTorch/TensorFlow, NLP models, and standard ML algorithms, and familiarity with the latest LLMs, RAG, and agent frameworks. Responsibilities include improving Onyx's agent and knowledge retrieval capabilities, enhancing multi-hop QA and precise search, and improving the platform's user experience. Onyx is backed by $10M in seed funding and boasts clients like Netflix and Ramp.

(www.ycombinator.com)

AI

π0.5: A General-Purpose AI Model Enabling Robots to Clean New Homes

2025-04-22

Physical Intelligence has developed π0.5, a robotic foundation model capable of generalizing complex cleaning tasks, such as tidying a kitchen or bedroom, to entirely new environments. Unlike previous robots limited to controlled settings, π0.5 leverages co-training on diverse heterogeneous data, including multimodal data and data from various robots, to learn diverse skills and understand their semantic context. Experiments show π0.5 can perform multiple tasks in unseen homes, exhibiting human-like flexibility and resourcefulness despite occasional failures. This represents a significant step toward truly generalizable physical intelligence.

(pi.website)

AI Generalization

Debunking the Myth of High-Degree Polynomials in Regression

2025-04-22

The common belief that high-degree polynomials are prone to overfitting and difficult to control in machine learning is challenged in this article. The author argues that the problem isn't high-degree polynomials themselves, but rather the use of inappropriate basis functions, such as the standard basis. Experiments comparing the standard, Chebyshev, and Legendre bases with the Bernstein basis in fitting noisy data demonstrate that the Bernstein basis, with its coefficients sharing the same 'units' and being easily regularized, effectively avoids overfitting. Even high-degree polynomials yield excellent fits using the Bernstein basis, requiring minimal hyperparameter tuning.

(alexshtf.github.io)

AI polynomial fitting overfitting Bernstein basis

Graph Transformers: The Next Generation of Graph Models

2025-04-22

Graphs are ubiquitous, but leveraging their complex, long-range relationships has been a challenge for machine learning. Graph Neural Networks (GNNs) excel at capturing local patterns but struggle with global relationships. Enter Graph Transformers, which leverage powerful self-attention mechanisms, enabling each node to directly attend to information from anywhere in the graph, thus capturing richer relationships and subtle patterns. Compared to GNNs, Graph Transformers offer advantages in handling long-range dependencies, mitigating over-smoothing and over-squashing, and more effectively processing heterogeneous data. While Graph Transformers have higher computational complexity, techniques like sparse attention mechanisms and subgraph sampling enable efficient processing of large graph datasets.

(kumo.ai)

AI Graph Transformers

RLVR Boosts Reasoning...But at What Cost?

2025-04-22

Experiments across math, coding, and visual reasoning domains evaluated the impact of RLVR (Reinforcement Learning from Human Feedback) on base and RLVR-trained large language models. Results showed RLVR improved accuracy at low k-values but decreased problem coverage at higher k-values. This suggests RLVR enhances deterministic accuracy but limits exploration diversity. Base models maintained broader reasoning coverage despite initial accuracy gains from RL. The consistent findings across domains indicate RLVR enhances reasoning without fundamentally altering the problem-solving approach.

(limit-of-rlvr.github.io)

AI Reasoning Ability

AI's Exponential Growth: Is AGI Near?

2025-04-22

Research from METR shows AI capabilities are growing exponentially, with recent models mastering software engineering tasks in months that previously took hours or days. This fuels speculation about the imminent arrival of AGI (Artificial General Intelligence). However, author Peter Wildeford points out METR's study focuses on specific software engineering tasks, neglecting the complexities of real-world problems and human learning. While AI excels in niche areas, it still struggles with many everyday tasks. He builds a model incorporating METR's data and uncertainties, predicting AGI could arrive in Q1 2030, but with significant uncertainty.

(peterwildeford.substack.com)

AI Exponential Growth

Cekura: Automating the Testing of AI Voice Agents

2025-04-21

Cekura, a Y Combinator-backed startup, is revolutionizing the reliability of AI voice agents. Founded by IIT Bombay alumni with research from ETH Zurich and a proven track record in high-stakes trading, Cekura tackles the cumbersome and error-prone nature of manual voice agent testing. They automate testing and observability by simulating thousands of realistic conversational scenarios, from ordering food to conducting interviews. Leveraging custom and AI-generated datasets, detailed workflows, and dynamic persona simulations, Cekura uncovers edge cases and provides actionable insights. Real-time monitoring, comprehensive logs, and instant alerts ensure optimized, production-ready calls. In a rapidly expanding market, Cekura stands out by guaranteeing dependable performance, reducing time-to-market, and minimizing costly errors. They empower teams to demonstrate reliability before deployment, building trust with clients and users.

(www.ycombinator.com)

AI AI voice agent

AI Robot: Fairy Tale vs. Reality

2025-04-21

This article contrasts the fictional AI robot 'Robot' from Annalee Newitz's story with the real-world clumsy CIMON, exploring the limitations of current AI. Robot, capable of independent learning and exceeding its programming, showcases the potential of Artificial General Intelligence (AGI). In contrast, CIMON's limited Artificial Narrow Intelligence (ANI) reveals its rigid nature. The author points out that current AI technology largely remains in the ANI stage, vulnerable to algorithmic bias and unable to adapt to complex situations as Robot does. While machine learning has made strides in language processing and image recognition, achieving AGI remains a distant goal. The author urges caution against over-reliance on biased training data and emphasizes the importance of self-learning and feedback mechanisms in AI development. Strive for Robot, plan for CIMON.

(slate.com)

AI

Dia: A 1.6B Parameter Text-to-Speech Model from Nari Labs

2025-04-21

Nari Labs introduces Dia, a 1.6B parameter text-to-speech model capable of generating highly realistic dialogue directly from transcripts. Users can control emotion and tone by conditioning the output on audio, and the model even produces nonverbal cues like laughter and coughs. To accelerate research, pretrained model checkpoints and inference code are available on Hugging Face. A demo page compares Dia to ElevenLabs Studio and Sesame CSM-1B. While currently requiring around 10GB VRAM and GPU support (CPU support coming soon), Dia generates roughly 40 tokens/second on an A4000 GPU. A quantized version is planned for improved memory efficiency. The model is licensed under Apache License 2.0 and strictly prohibits misuse such as identity theft, generating deceptive content, or illegal activities.

(github.com)

AI

Inner Loop Agents: LLMs Calling Tools Directly

2025-04-21

Traditional LLMs require a client to parse and execute tool calls, but inner loop agents allow the LLM to parse and execute tools directly—a paradigm shift. The post explains how inner loop agents work, illustrating the difference between them and traditional LLMs with diagrams. The advantage is that LLMs can concurrently call tools alongside their thinking process, improving efficiency. Reinforcement learning's role in training inner loop agents and the Model Context Protocol (MCP)'s importance in supporting diverse tool use are also discussed. Ultimately, while LLMs can currently use tools, achieving optimal tool use requires specialized model training for best results.

(timkellogg.me)

AI Inner Loop Agents

AI-Assisted Search-Based Research: Finally Useful!

2025-04-21

For two and a half years, the dream of LLMs autonomously conducting search-based research has been pursued. Early 2023 saw attempts from Perplexity and Microsoft Bing, but results were disappointing, plagued by hallucinations. However, the first half of 2025 brought a turning point. Gemini, OpenAI, and Perplexity launched "Deep Research" features, generating lengthy reports with numerous citations, albeit slowly. OpenAI's new o3 and o4-mini models are a breakthrough, seamlessly integrating search into their reasoning process to provide reliable, hallucination-free answers in real-time. This is attributed to robust reasoning models and resilience to web spam. While Google Gemini and Anthropic Claude offer search capabilities, they lag behind OpenAI's offerings. A stunning example: o4-mini successfully upgraded a code snippet to a new Google library, showcasing the potential of AI-assisted search, but also raising concerns about the future of the web's economic model and potential legal ramifications.

(simonwillison.net)

AI Web Economics

Immune Cytokine IL-17: A Double-Edged Sword in the Brain

2025-04-21

Research from MIT and Harvard Medical School reveals that the immune cytokine IL-17 exerts contrasting effects on the brain. In the amygdala, it promotes anxiety, while in the somatosensory cortex, it enhances social behavior. This highlights a strong interplay between the immune and nervous systems. The findings suggest IL-17 might have initially evolved as a neuromodulator before being co-opted by the immune system for inflammation. This discovery could pave the way for novel treatments for neurological disorders like autism or depression by targeting the immune system to influence brain function.

(news.mit.edu)

AI cytokines

ChatGPT's New Watermark: A Cat and Mouse Game?

2025-04-21

Rumi's team discovered that newer GPT models (o3 and o4-mini) embed special character watermarks, primarily narrow no-break spaces, in longer generated texts. These are invisible to the naked eye but detectable with code editors or online tools. While potentially useful for detecting AI-generated content, they're easily removed. This might cause widespread attention among students, potentially leading OpenAI to remove the feature. Rumi advocates for a process-focused approach to student writing, emphasizing AI literacy over easily bypassed technical solutions.

(www.rumidocs.com)

AI watermark

Saying 'Please' and 'Thank You' to ChatGPT Costs OpenAI Millions

2025-04-20

OpenAI CEO Sam Altman revealed that user politeness, specifically saying "please" and "thank you" to ChatGPT, costs the company tens of millions of dollars in electricity. While Altman claims it's money well spent, the revelation highlights the massive energy consumption of AI. A survey shows 70% of users are polite to AI, partly fearing a robot uprising. However, the debate rages on: does politeness improve responses, and is it worth the environmental cost? Some argue polite prompts yield better, less biased results, improving AI reliability.

(www.techradar.com)

AI

Ravens Show Unexpected Geometric Skills

2025-04-20

Researchers at the University of Tübingen have demonstrated that ravens possess the ability to recognize geometric regularity. In a study published in Science Advances, carrion crows were trained to identify an outlier shape amongst several similar ones. The crows successfully distinguished subtle differences in shapes, exhibiting an understanding of right angles, parallel lines, and symmetry. This challenges previous assumptions about animal cognition, suggesting this ability may be more widespread than previously thought.

(phys.org)

AI geometric recognition raven

Controversial AI Startup Aims for Total Job Automation

2025-04-20

Silicon Valley startup Mechanize, founded by renowned AI researcher Tamay Besiroglu, has sparked controversy with its ambitious goal: the complete automation of all work. This mission, alongside Besiroglu's connection to the respected AI research institute Epoch, has drawn criticism. Mechanize aims to automate all jobs by providing the necessary data, evaluations, and digital environments, resulting in a massive potential market but raising significant concerns about widespread job displacement. While Besiroglu argues that automation will lead to explosive economic growth and higher living standards, he fails to adequately address how people would maintain income without jobs. Despite the extreme ambition, the underlying technical challenge is real, and many large tech companies are pursuing similar research.

(techcrunch.com)

AI startup controversy

Recursive Prompts: Implementing Recursion with LLMs

2025-04-20

This article explores a novel approach to implementing recursion using Large Language Models (LLMs). By crafting a recursive prompt that iteratively updates its own internal state, the author demonstrates how an LLM can generate a sequence of prompts converging towards a solution, mirroring the behavior of recursive functions in code. The article uses the Fibonacci sequence as an example, showcasing how recursive prompting can perform calculations. It also discusses challenges like handling inaccuracies in the LLM's output and leveraging the LLM's existing knowledge base, drawing parallels to how humans perform mental arithmetic using memorized algebraic and atomic rules. The work is connected to related research like ReAct and ACT-R, and addresses strategies for mitigating errors in LLM-generated results.

(github.com)

AI Recursion

Is AGI Here? No, It's 'Jagged AGI'

2025-04-20

Recent AI models like OpenAI's o3 and Google's Gemini 2.5 Pro show stunning advancements, even completing complex tasks like marketing campaigns and website building. Economist Tyler Cowen suggests this signifies the arrival of AGI. However, the article argues these AIs exhibit uneven capabilities, excelling in some areas while failing at simple ones – a concept termed 'Jagged AGI'. This uncertainty makes the definition and impact of AGI unclear, suggesting its application and societal integration could be a lengthy process, or potentially see rapid adoption. The future remains uncertain.

(www.oneusefulthing.org)

AI

Meta's Llama and the EU AI Act: A Convenient Coincidence?

2025-04-20

Meta's labeling of its Llama models as "open source" is questionable, as its license doesn't fully comply with the Open Source Definition. A theory suggests this is due to the EU AI Act's special rules for open-source models, bypassing OSI compliance. Analyzing the Act with Gemini 2.5 Flash, the author found exemptions for models allowing users to run, copy, distribute, study, change, and improve software and data, even with attribution requirements. This supports the theory that Meta strategically uses the "open source" label, although this practice predates the EU AI Act.

(simonwillison.net)

AI

FramePack: A Revolutionary Next-Frame Prediction Model for AI Video Generation

2025-04-20

FramePack is a groundbreaking next-frame prediction neural network architecture that compresses input contexts to a fixed length, making the generation workload independent of video length. This achieves O(1) computational complexity for streaming, setting a new benchmark in AI video generation. It generates high-quality videos using only 6GB of GPU memory on laptops with RTX 3060. Generation speed reaches 1.5-2.5 seconds per frame on an RTX 4090, but is 4-8 times slower on laptops with 3070ti/3060. Its bi-directional sampling method effectively eliminates the common drifting problem in video generation.

(frame-pack.video)

AI next-frame prediction

OpenAI's $3B Windsurf Acquisition: A Sign of Desperation in the AI Arms Race?

2025-04-20

OpenAI's recent $3 billion acquisition of Windsurf (formerly Codeium), an AI coding assistant, has sent shockwaves through the industry. This follows Google's massive acquisition of Wiz, but Windsurf's relatively smaller user base and market share raise questions about the hefty price tag. The article explores potential motivations behind OpenAI's move, including securing data, strengthening distribution channels, and navigating strained relations with Microsoft. It also compares OpenAI, Google, and other players in the AI landscape, highlighting Google's dominance in model performance and price competitiveness, along with its strategic moves to solidify its lead. Finally, the article examines Apple's struggles in AI, attributing them to limitations in computing resources and data acquisition, and the constraints imposed by its commitment to user privacy.

(theahura.substack.com)

AI AI acquisition

Gemma 3: Bringing State-of-the-Art AI to Your Desktop

2025-04-20

Gemma 3, a cutting-edge open-source AI model, initially required high-end GPUs. To enhance accessibility, new versions optimized with Quantization-Aware Training (QAT) dramatically reduce memory requirements while maintaining high quality. This allows running powerful models like the 27B parameter Gemma 3 on consumer-grade GPUs such as the NVIDIA RTX 3090. These optimized models are available on Hugging Face and Kaggle, enabling easy integration into various workflows.

(developers.googleblog.com)

AI Quantization-Aware Training

DeepSeek: The Unstoppable Wave of Open-Source AI

2025-04-20

The release of the DeepSeek model sparked a revolution in open-source AI. Initially released by a Chinese team, it was quickly replicated and improved upon by developers worldwide, leading to projects like OpenSeek by the Beijing Academy of Artificial Intelligence. Despite US government attempts to restrict involved entities, DeepSeek has evolved into a thriving community. Tens of thousands of developers are collaboratively advancing AI technology at a speed and scale unmatched by any centralized entity. This demonstrates the unstoppable nature of community-driven open source, defying containment by any single country, company, or government.

(www.infoworld.com)

AI

AI: A Collaborative Partner, Not a Replacement

2025-04-20

Many misunderstand AI, believing it fully automates writing, planning, and problem-solving. The author argues AI is more like a 'thought-checker,' enhancing human thought, not replacing it. Using performance reviews and meeting notes as examples, the article highlights AI's shortcomings in lacking human insight, contextual understanding, and reliability. The author proposes viewing AI as a collaborative partner, engaging in iterative dialogue to improve work quality and efficiency. The ultimate goal isn't speed, but improved quality.

(hils.substack.com)

AI

Anthropic Reveals Claude Code's 'UltraThink' Mode

2025-04-20

Anthropic released extensive documentation on best practices for their Claude Code CLI coding agent tool. A fascinating tip reveals that using words like "think," "think hard," etc., triggers extended thinking modes. These phrases directly correlate to different thinking budgets; "ultrathink" allocates a massive 31999 tokens, while "think" uses only 4000. Code analysis shows these keywords trigger functions assigning varying token counts, impacting Claude's thinking depth and output. This suggests "ultrathink" isn't a Claude model feature, but rather a Claude Code-specific enhancement.

(simonwillison.net)

AI

O(1) Streaming Video Prediction with GPU Memory Optimization

2025-04-19

A novel video prediction model achieves O(1) streaming complexity through optimized GPU memory layout. The model encodes input frames into GPU memory, allocating different context lengths (number of tokens) to frames based on their importance. For instance, in HunyuanVideo, a 480p frame can have its token count adjusted from 1536 to 192 using different patchifying kernels. This allows the most important frames (e.g., the one closest to the prediction target) to utilize more GPU resources, resulting in significant efficiency gains and remarkably achieving O(1) complexity without complex algorithmic optimizations.

(lllyasviel.github.io)

AI video prediction O(1) complexity

Category: AI