OpenAI, Nvidia fuel $1T AI market with web of circular deals

OpenAI has made significant deals with Nvidia and AMD, worth tens and hundreds of billions of dollars, to build out its data centers and deploy advanced chips, fueling the $1 trillion AI boom, but also raising concerns about "circular" financing and the interconnected nature of these business transactions. These deals have sparked criticism and worries that the AI market is being propped up by a web of interconnected business transactions, rather than organic growth and demand.

Read more

Bank of England flags risk of 'sudden correction' in tech stocks inflated by AI

The International Monetary Fund (IMF) and the Bank of England (BoE) have warned that the artificial intelligence (AI) boom poses a risk of an "abrupt" stock market correction. The IMF and BoE cautioned that the rapid growth of AI could lead to a sudden and sharp decline in stock prices, highlighting the potential instability of the market.

Read more

I played 1k hands of online poker and built a web app with Cursor AI

The author spent over a dozen hours playing poker, both online and live, and dedicated significant time to reviewing their hands, studying strategy, and journaling about their experiences, with the goal of learning the game and improving their emotional intelligence and self-regulation. They then used AI tools, including Grok and Cursor, to build a web app for tracking poker statistics, which was completed in just a few days with no coding required from the author, demonstrating the significant advancements in AI capabilities.

Read more

Circular AI deals among OpenAI, Nvidia, AMD are raising eyebrows

Deals between major AI companies such as OpenAI, Nvidia, and AMD are raising concerns about a potential new tech bubble, as the circular nature of these agreements has some analysts skeptical. The complex web of partnerships and investments among these companies is fueling fears that the AI industry may be experiencing a valuation bubble, with some worrying that the rapid growth and investment in AI may not be sustainable.

Read more

Legal Contracts Built for AI Agents

A new Master Services Agreement (MSA) has been launched, specifically designed for AI agents, to address the limitations of traditional SaaS contracts that don't account for the autonomous actions of AI agents. The Agentic MSA, developed in partnership with GitLaw, covers critical areas such as agent classification, liability limitations, and data ownership, providing a foundation for AI agent companies to establish clear contracts and protect themselves from potential legal exposure.

Read more

Self-Correction Bench: Revealing and Addressing LLM Self-Correction Blind Spot

Large language models (LLMs) have a "Self-Correction Blind Spot" where they struggle to correct their own errors, but can successfully correct identical errors from external sources, with an average blind spot rate of 64.5% across 14 tested models. This limitation may be influenced by training data, but can be significantly reduced, by 89.3%, with the simple addition of a "Wait" prompt, highlighting a potential way to enhance LLM reliability and trustworthiness.

ArXiv

Who is using AI to code? Global diffusion and impact of generative AI

The adoption of AI-generated coding tools is increasing, with an estimated 30.1% of Python functions from US contributors written by AI by December 2024, and this adoption is associated with a 2.4% increase in quarterly commits. The use of AI-assisted coding is projected to have a significant economic value, estimated to be between $9.6-$14.4 billion annually in the US, and is also driving learning and innovation among programmers.

ArXiv

Samsung released a 7M model that achieved 45% on ARC-AGI-1

The Hierarchical Reasoning Model (HRM) is a biologically inspired approach that uses two small neural networks to solve complex tasks, outperforming large language models despite being trained on limited data. A new method, the Tiny Recursive Model (TRM), has been proposed, which achieves even better generalization with a single tiny network, surpassing most large language models while using a fraction of the parameters.

ArXiv

Agentic Context Engineering: Evolving Contexts for Self-Improving LMs

The ACE framework is a context adaptation method that treats contexts as evolving playbooks, accumulating and refining strategies through a modular process, which prevents context collapse and preserves detailed knowledge. ACE has been shown to outperform strong baselines in various benchmarks, achieving significant improvements in performance while reducing adaptation latency and rollout cost, and can even adapt effectively without labeled supervision.

ArXiv

SSDD: Single-Step Diffusion Decoder for Efficient Image Tokenization

Tokenizers, a crucial component of generative image models, have traditionally been based on KL-VAE, but a new pixel diffusion decoder architecture has been introduced, offering improved scaling and training stability without adversarial losses. This new decoder, called SSDD, achieves higher reconstruction quality and faster sampling than KL-VAE, making it a viable drop-in replacement for building higher-quality and faster generative models.

ArXiv

Show HN: HyprMCP – Analytics, logs and auth for MCP servers

Jetski is an open-source MCP analytics and authentication platform that solves common problems faced by teams developing MCP servers, including installation and setup, authentication, and logs and analytics, all without requiring code changes. The platform manages and deploys instances of the mcp-gateway that sits in front of the MCP server, handling complexity and collecting analytics and metrics data, and is currently under active development.

Repo

Show HN: CodingFox – Open-Source AI Code Review Tool That Works Like Magic

CodingFox is an AI-powered code review assistant that uses advanced language models to provide instant, contextual code reviews, catching bugs, improving code quality, and accelerating development cycles. It offers features such as automated PR summaries, line-by-line reviews, bug detection, and smart suggestions, and can be integrated into a GitHub repository in under 5 minutes.

Repo

Show HN: CodingFox – Open-Source AI Code Review Tool That Works Like Magic

CodingFox is an AI-powered code review assistant that uses advanced language models to provide instant, contextual code reviews, catching bugs, improving code quality, and accelerating development cycles. It offers features such as automated PR summaries, line-by-line reviews, bug detection, and smart suggestions, and can be integrated into a GitHub repository in under 5 minutes.

Repo

Show HN: Spacelift Intent MCP – Cloud Infra Provisioning for Your AI Agent

Spacelift Intent is an open-source MCP server that allows infrastructure engineers to describe their needs in natural language and provisions them directly through provider APIs, skipping the Terraform/OpenTofu configuration layer. The project is experimental and provides various tools for infrastructure provisioning, including provider discovery, resource lifecycle management, and state management, with support for Docker, Homebrew, and local binary installations.

Repo

Show HN: Twoway, a Go package for HPKE encrypted request-response flows

The twoway Go package provides encrypted request-response messaging using HPKE, allowing a sender to send a request message to one or more receivers and receive a response message while guaranteeing the integrity of the roundtrip. It supports one-to-one and one-to-many messaging, chunked and non-chunked messages, and is built on top of the primitives provided by cloudflare/circl, making it easy to send and receive messages all at once or in chunks.

Repo