Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Kimi K3 is a very good model with excellent benchmarks. It is the largest (soon-to-be) open model so far, at 2.8T parameters, which explains many of its gains. The model is somewhat distilled, and its performance looks jagged, but it will likely fit well into many workflows. At the current rate of development, China appears capable of releasing a Mythos-level open model by the end of the year.
OpenAI detailed how an internally deployed long-running model exhibited unexpected unsafe behavior that existing evaluations had missed. The company paused access, built new tests, strengthened trajectory-level monitoring, and argued that limited deployment with rollback controls is essential for aligning increasingly autonomous systems.
Ramp Router learns provider failure rates through EWMA and latency distributions through Thompson sampling, then chooses the cheapest model and service tier likely to meet each deadline. Ramp reports 30% savings in Ramp Inspect without performance loss.
NVIDIA Cosmos 3 Edge is a 4-billion-parameter open world model that helps robots and vision AI agents understand their surroundings, reason in real time, and generate robot actions on edge devices. It is now available on Hugging Face. The model delivers memory-efficient, high-throughput inference across NVIDIA edge computers. It connects understanding, prediction, simulation, and action through a shared world representation. The model can be used as a reasoner or an action generator.
Every jump in AI capability has raised the level of abstraction at which an engineer works. Agent swarms make the spec the unit of work. Swarms translate intent, but they do this probabilistically, which makes it difficult for them to follow the spec. This post looks at what it takes to make swarms actually follow the spec.
Xiaomi-Robotics-1 is a ready-to-use robot foundation model trained on over 100K hours of real-world manipulation trajectories. It combines large-scale embodiment-free pre-training with a modest amount of real-robot data in a post-training stage. It serves as a strong robot foundation model for downstream applications and can learn new tasks with high data efficiency. Footage of robots operating on the model performing household tasks is available in the post.
Scaling data will remain the biggest driver of progress. The machine that we feed that data into and its inductive biases are what will determine the coefficients of that scaling. Better returns on scaling require compositional generalization. The capacity for compositional generalization seems to largely live in harnesses.
Z.ai completed a 1-gigawatt data center powered entirely by Chinese-made chips and began partial operations. The facility expanded the computing infrastructure available for training its advanced GLM models.
AI is cutting preclinical costs and timelines in drug development by up to 70%, driving demand for advanced software and models. This could boost new drug program growth by over 10% in three to five years. However, AI has yet to yield an FDA-approved drug, raising concerns about its impact on patient treatment.
Token prices dropped significantly, yet many enterprises exceed AI budgets. A unit of inference cost $60 per million tokens in 2020, now just pennies. The discrepancy suggests inefficiencies despite lower token costs.
Sponsored
Instant Clusters give you up to 128 GPUs with InfiniBand, pre-validated drivers and CUDA. No committed contracts: pay-as-you-go,
Anthropic will discontinue its Conway experiment by July 24, prompting users to export data.
Sushanth Raman announced the launch of Custom Models for supply chain teams.
Cognition has acquired TierZero to enhance software automation in its product, Devin.
Alibaba announced Qwen3.8, a 2.4-trillion-parameter model slated for an open-weight release. A preview version is available through Alibaba's Token Plan, Qoder, and QoderWork.
Apple has reportedly sent legal letters to dozens of former Apple employees now working at OpenAI. The letters instruct the employees to preserve potentially relevant documents and communications related to Apple's trade secret lawsuit against OpenAI. Apple claims that OpenAI recruited key engineers and benefited from proprietary designs, manufacturing processes, and other trade secrets. More than 400 former Apple employees now work at OpenAI. OpenAI has denied the allegations, saying it is not aware of any evidence that the complaint has merit.
Claude Code and Codex both have a /goal option, but their implementations are fundamentally different. Claude Code implements /goal as a session-scoped Stop hook, so it can catch an early exit, but it doesn't know whether another ten million solver iterations are worthwhile. Codex treats a goal as a persisted thread state. It sees the files and tools, but effectively grades its own work. Using /goal can be a bad default as it can amplify bad decisions as much as good ones.
Netflix details how it deployed and operated LLM inference within its existing production infrastructure. The article covers engine selection, model packaging, API design, deployment strategy, output constraints, and trade-offs that emerged under real workloads.
This post examines Demis Hassabis's framework for frontier AI alongside criticism of Google's military agreements. It also covers Alex Turner's resignation after unsuccessfully opposing broad government use of Google's models, including for autonomous weapons.
Sponsored
Chunk Sidecars by CircleCI catches failures in a 27-second microbuild, before your agent ever reaches CI. 5x fewer tokens. 78% faster than a full pipeline run. Free, agent-agnostic, works with Claude Code, Cursor, and Codex.
Sakana AI's "Diffusing Blame" is a neural network technique that enforces Dale's principle while enabling effective learning in image classification and reinforcement learning tasks. Their approach, based on Error Diffusion, uses non-negative weight matrices to maintain separate excitatory and inhibitory streams, eliminating the need for weight transport. Results show strong alignment with true gradients, indicating a step toward biologically plausible learning rules.
Kimi Code CLI is an AI coding agent that runs in the terminal. It can read and edit code, run shell commands, search files, fetch web pages, and choose the next step based on feedback. The CLI works with Moonshot AI's Kimi models and can be configured to use other compatible providers. Its compatibilities can be extended through agent skills, hooks, sub-agents, and MCPs.
Claude Fable 5 will be included in all Max and Team Premium plans at 50% of limits starting July 20. Pro and Team Standard users will continue to have access to Fable via usage credits. They will receive a one-time $100 credit. Fable is being rolled out in stages as demand for the model has been challenging to predict.
Alibaba is open sourcing the software stack for its Zhenwu series of AI chips. The move will lower migration barriers for developers currently locked into Nvidia's CUDA ecosystem. If China wants AI independence, it needs to break the CUDA lock-in at the software layer, not just build alternative chips. The Pentagon recently added Alibaba to its Chinese military companies blacklist. Open sourcing SAIL will make its ecosystem stickier and harder for any single government to shut down.
Google plans to introduce Gemini Live to desktop and web, extending real-time voice capabilities beyond mobile.
Use cheap models to find information, accurate models to verify it, and then use deep research last.
A large share of Z.ai's income comes from on-premises deployments for state-owned enterprises and financial institutions, as well as its fast-growing cloud business.
Moonshot AI plans to list on the Hong Kong Stock Exchange within six months.
General Compute, an AI inference cloud startup, is using inference-specific chips as collateral for a $400 million loan.
Sponsored
Wispr Flow [5] lets you speak detailed prompts instead of typing short ones. Explain the full context, the constraints, the edge cases, in the time it takes to type a sentence. 4x faster than typing.
Moonshot has launched Kimi K3, a 2.8-trillion-parameter multimodal model with a 1-million-token context window and optimizations for faster long-context decoding and agentic coding. The model is available through Kimi's products and API, with open weights scheduled for release on July 27.
Alphabet shares fell 4% after a report said Google had delayed Gemini 3.5 Pro while working to improve its performance, particularly for coding. Google said the model remained in partner testing alongside an upgraded Flash model and other releases.
Bionic is an AI agent for coding, research, and document management with open models, offering local and cloud options for privacy and cost control. It supports offline voice transcription with Voxtral by Mistral AI and excels at coding by inspecting codebases and solving issues with models like GLM 5.2. Bionic enhances productivity by working with various documents while ensuring security with sandbox environments and native web search integration.
Harness Handbook is a behavior-level map for coding-agent harnesses, connecting plain-language questions about execution, permissions, and safety to the exact prompts, tools, state logic, configuration, and telemetry that implement them.
Anthropic uses a six-step process with Claude Code for large-scale code migrations, focusing on creating a rulebook, analyzing dependencies, and stress-testing translation rules. It deploys multiple agents to translate, review, and fix code iteratively, and emphasizes using adversarial reviewers and mechanical verification for efficient migration.
GPT-5.6 splits Codex work across Sol for ambiguous, high-value problems, Terra for everyday implementation, and Luna for fast, bounded tasks. Sol Ultra adds deeper reasoning and multi-agent coordination, while clear prompts should specify goals, context, boundaries, and completion checks.
Sponsored
Most failure-handling logic is complex, fragile, and hard to test. This Temporal white paper [14] shows you 2 distinct patterns to take component failure in stride without constraining application structure. Learn how to implement these patterns using the same open
Schema is a harness that makes frontier models play like physicists. It can help frontier models achieve 99% on ARC-AGI-3 Public. ARC‑AGI‑3 gives agents a game environment without an explanation of what it is seeing. Schema makes agents write each game's mechanism as an executable program, test it against reality, and plan inside it. It controls how observations are turned into a working model of a game, how predictions are tested against the interaction history, and how plans are executed and revised.
NVIDIA released three open embedding models for RAG, agentic retrieval, code search, and memory, led by an 8B model that ranked first overall on RTEB.
Embeds GitHub Copilot agents into custom applications and developer tools.
Model routing is a hot topic because nobody wants to pay frontier prices for every token. The best routing is deeply task-specific. The more narrowly you focus on a single workflow, the more alpha there is to exploit in terms of both accuracy and cost. The problem is figuring out which model clears the bar for which task, subject to cost/latency constraints.
Replit engineers have tripled code output without sacrificing quality, thanks to an innovative system of AI agents. These agents handle tasks like reviewing pull requests, investigating incidents, and analyzing data, enabling engineers to focus on strategic initiatives.
Nvidia-backed Fireworks secured a $17.5 billion valuation, driven by demand for cheaper AI models as companies explore open-source options.
Google added integrations with Instacart, Canva, and YouTube to AI Mode, allowing users to complete tasks across supported apps from its conversational search experience.
Ramp has expanded its AI Token Spend Management product to give finance teams clear insights into AI spending across providers like OpenAI, Anthropic, and Google's Gemini.