Researchers evaluated 21 large language models across 16 benchmarks and found that combining their strengths can reduce the average error rate by 54% compared to the best individual model, and up to 82% with optimal model selection and a perfect judge. They also created an interactive dashboard called AI Frontier to explore the results.
GPT-6 Astra achieves 97.2% and 90.6% accuracy on short and medium-sized document extraction tasks, but is pricey and struggles with long documents and certain OCR tasks. Full results are available on ExtractBench.
Grok BotSpaceXAI introduces a new level of abstraction in agent configuration, making it easier to set up and use agents without requiring extensive coding knowledge. The platform provides a user-friendly interface, allowing users to create and manage agents with ease, and integrates with various tools and services. OpenClaw 2.0, on the other hand, offers more flexibility and customization options, but with a steeper learning curve.
The 2024 Nobel Prize in Physics has been awarded to Hopfield and Hinton, but the award is controversial due to allegations of plagiarism and incorrect attribution in computer science.
Perplexity Computer has added GPT-6 Astra for Pro and Max subscribers, expanding their AI capabilities.
Perplexity has exposed its inference capabilities through standardized APIs, utilizing a combination of Ivy, Tulip, and ROSE components to handle tasks such as parsing, scheduling, and forward passes. The system is implemented across multiple languages, including Rust and Python.
Deep Agents' virtual file system now supports MongoDB, allowing teams to choose their preferred storage layer, and its agent code remains compatible through the BackendProtocol.
Nathan updates his AI forecast, suggesting any verifiable domain can achieve weak superintelligence quickly, and notes Astra has saturated FrontierMath tier 4 with a third of the year left.
Perplexity has published research on building state-of-the-art serving infrastructure for embedding and ranking models, detailing their approach and findings.
Tulip uses CUDA graphs and LazyTensors to reduce latency and increase throughput by minimizing CPU overhead and enabling asynchronous result tracking. This approach allows the CPU to prepare the next batch while the GPU finishes the current one.
Emmimalpa created a prompt-dependency graph to help practitioners evaluate changes to composable prompts, making it easier to identify what needs to be re-evaluated when a shared component is changed.
Mostafa Ibrahim argues against the emerging consensus on splitting prefill and decode onto separate GPU pools for improved throughput, citing its inapplicability to most teams.
Physical Intelligence has made progress in solving tasks that are trivial for humans but hard for robots, such as making a peanut butter sandwich, with a 53% success rate, although still 10x slower than a human.
The past year has seen significant advancements in AI models, with notable improvements from last year's Opus 4.1, GPT-5, and Gemini 2.5 to the current Fable and Astra models.
Grok bot's architecture uses a real computer for its agent and server, allowing for simpler real-time updates and message syncing across devices.
OpenAI's 2016 project Universe has been revisited with the reincarnation of Astra, achieving reliable booking of a United flight, and the author reflects on lessons learned from the original project.
Shuai Guo explores building multimodal applications using a local LLM with Gemma 4 and Ollama, enabling private AI workflows beyond text.
Tesla's robotaxi development is progressing similarly to Waymo's, but about 3-4 years behind, with each scale increase introducing new edge cases to address. The company's vision for widespread robotaxi deployment is more complex than a simple push-button solution.
Vitest 5 has been released, featuring significant performance improvements, including up to 53% faster vm pools and an 18% boost across the board.
A serving infrastructure combining Ivy, Tulip, and ROSE improves latency and throughput for online and batch embedding workloads, and offers faster search at a lower cost than off-the-shelf solutions.
The Rust programming language has achieved a significant improvement in compilation speed, now compiling approximately 3.1 times faster than before.
EivindKjos discusses integrating Claude Code into efficient deployment workflows for faster code generation.
Gary Marcus and ilex_ulmus discuss the need for OpenAI to pause, with Marcus stating his reasons for this stance.
Ivy is an HTTP gateway that handles CPU-side request preparation, including parsing, tokenization, and templating, before sending requests to Tulip over gRPC, allowing for tuning of request formatting without affecting inference servers. This enables more efficient and flexible request processing.
The use of MCP connectors is anticipated to increase rapidly as agents begin to evaluate and select products, potentially replacing human decision-making.
Nathan discusses the potential of blobby, sensing appendages in robotics with Jean Nehme, inspired by a Google DeepMind roboticist's desire for better robotic hands. The conversation is part of the AI:AM series.
ROSE is a model engine that reuses kernels for large language models (LLMs) and embeddings, using ragged attention for embeddings and supporting multiple attention backends. It optimizes performance based on model shape and sequence length.
Shuai Guo discusses how structured outputs can make local large language models easier to integrate into real-world workflows, addressing the integration problem of local models.
The U.S. government considered banning TikTok and there are concerns that OpenAI may cause harm.
Anthropic is soliciting input on a potential overhaul of Claude Code to make it more extensible, and has shared early details.
Spotify's Portal by Spotify introduces AiKA Modes, which can reduce token usage for AI coding agents like Claude Code by up to 90% by delegating grunt work to cheaper models, and two modes, bulk-reader and code-writer, are available for public use, allowing for reusable and shareable model routing. The modes are part of the Portal plugin, which can be installed and set up in a new Claude Code session to start using the modes.
OpenAI's GPT-6 Astra shows promising results in code review, catching 4% more labeled bugs than GPT-5.6 Sol and 22% more than Opus 5, with larger gains in harder cross-file reviews. Astra's pricing is $10 per million input tokens and $50 per million output tokens, with potential applications in research synthesis, operational investigation, and document analysis. The model was also used to build a game, NIGHTSHIFT, demonstrating its ability to reason about complex systems and balance gameplay mechanics.
Artificial Analysis has released Intelligence Index v4.2, an update to their benchmarking tool, featuring new evaluations such as AA-Briefcase and GDP.pdf, and improved grading infrastructure. The update brings the Index closer to real-world use cases with more challenging and realistic tasks. Anthropic's Claude Fable 5.1 and OpenAI's GPT-6 Astra lead the Index, with Meta and Z.AI also performing well. The update includes a private held-out test set to prevent gaming and improved robustness of grading sandboxes.
GPT-6 Astra outperformed Claude Fable 5 and Fable 5.1 in a robotic manipulation task, successfully placing a block in a bowl 19 out of 20 times, and demonstrated comparable performance to Fable 5.1 in a puzzle task. The results show Astra's improved performance and efficiency, with lower estimated costs per run.
OKF Agent Memory is a domain-neutral, Git-native persistent project memory for AI agents based on the Open Knowledge Format (OKF) v0.2, providing a standardized, vendor-neutral memory layer that lives directly in the repository as plain Markdown files with YAML frontmatter. The project includes a zero-dependency Go toolchain, a native Model Context Protocol (MCP) server, and a comprehensive onboarding guide.
The increasing use of AI-assisted incident response tools may lead to a loss of practice for human responders, making them less prepared to handle complex incidents. The aviation industry's approach to training pilots for rare failures can serve as an inspiration for the software industry, which needs incident simulators to help engineers develop and maintain their skills. Rootly has partnered with Uptime Labs to create realistic incident simulations, allowing engineers to practice their skills in a safe environment.
Anthropic has published updated system prompts for their Claude consumer applications, including new rules for reproducing song lyrics and copyrighted characters, as well as changes to Claude's answering style and handling of abusive conversations. The updates also include recommended substance support sites and a reliable knowledge cutoff date of June 2026.
The author criticizes the use of large language models (LLMs) to generate content on LinkedIn, arguing that it leads to low-quality writing and undermines the authenticity of the poster. The author suggests that LLMs are useful for other tasks, but not for writing, and encourages users to write their own content instead. The author also notes that LinkedIn's own AI-powered rewriting tool can contribute to the problem.
The adoption of AI in companies is hindered by the complexity of existing workflows and the need for institutionalized solutions, despite the potential for AI to automate tasks and increase productivity. Companies are experimenting with AI pilots, but scaling AI adoption is a challenge. AI is expected to expand existing apps and create new vertical apps, but its impact will depend on how companies adapt to the new technology. The article discusses the limitations of AI in automating tasks and the need for a more nuanced approach to AI adoption, including the importance of understanding the company's existing workflows and the need for change management and training. The article also highlights the role of professional services firms in helping companies deploy AI solutions.
Cloud in a Bottle is an open-source personal cloud platform that allows users to self-host containerized apps with unified auth and good UX, aiming to make self-hosting accessible to everyone. The platform is available in managed and self-hosted versions, with a curated app catalog and a goal to catalyze the creation of more open-source web software.