Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Data Labs are projected to surpass $100B/year in data spend by 2030 as demand shifts from compute-limited to data-limited regimes. The bottleneck now lies in collecting private, high-quality datasets, as public internet data becomes insufficient for AI training needs. This scarcity positions data as a strategic asset pivotal to economic and scientific progress, potentially spurring efforts to cultivate new data sources akin to major compute investments.
Decagon runs roughly 90% of its workloads on open-source models because small, heavily fine-tuned models deliver the latency and task-specific performance required for customer service agents. Most enterprise AI use cases remain early-stage, so companies favor frontier models that maximize flexibility and intelligence. As deployments mature and workflows stabilize, many production workloads will likely migrate from closed models to specialized open-source alternatives.
Sponsored
AI fluency is quickly becoming a must-have skill in today's workplace. See how CoderPad's AI Interview Designer turns any job description into a realistic, role-specific interview that helps evaluate how candidates actually use AI. Register here [12].
Hy3 is a 295-billion-parameter Mixture-of-Experts model with 21 billion active parameters and 3.8 billion MTP layer parameters. Developed by Tencent, Hy3 outperforms similar-sized models and rivals flagship open-source models with two to five times more parameters. The full-sized model is available on Hugging Face. It is available for free on OpenRouter until July 21.
Loops are agents that repeat cycles of work until a stop condition is met. They can be categorized based on how they are triggered, how they are stopped, what Claude Code primitive is used, and what type of task is appropriate for each. This post covers the main loop types, when to use each, and how to maintain code quality while managing token usage.
Training state-of-the-art large language models with billions of parameters requires distributed training across hundreds or thousands of GPUs. At this scale, hardware failures are expected. PyTorch Monarch enables elastic, fault-tolerant distributed training on AMD GPUs. This post shows how Monarch dynamically recovers from node failures without halting the entire job and why this represents a significant step toward stable large-scale AI infrastructure.
PACE framework predicts costly agentic LLM benchmark performance using a small subset of atomic evaluation instances, achieving high accuracy at a fraction of the cost. It uses a regression model with selected instances from non-agentic benchmarks to predict scores on agentic benchmarks. PACE reduces evaluation costs by over 99% while maintaining a mean absolute error under 4%, aiding model development and selection.
xAI has rebranded to SpaceXAI. The rebranding brings new clarity to the business. The AI division is critical to Elon Musk's narrative about SpaceX's future. In this story, space infrastructure and exploration are inextricably linked to AI.
Claude Code, Codex CLI, and Omp are close enough in result quality that it wouldn't make sense to rank them. They can all read a serious repository, form a plan, edit across files, run checks, recover from failures, and land production-shaped patches. However, they differ in task clarity, repository hygiene, permissions, and whether the harness exposes the right tool at the right moment. OpenCode produces lower-quality results, but it tries the hardest to be good with every model.
Sponsored
Zoned out for a moment on the Zoom? Need extra context? Ask an LLM in real time - Granola already has the meeting notes. 1 month free with code [20]: TLDR1MO
Even Realities Technology, founded by ex-Apple exec Will Wang, has reached unicorn status thanks to a $150 million funding round led by Tencent and Meituan.
Initial capacity is expected to come online in the second half of next year, with a full buildout targeted for 2028.
There are levels to the art of agentic engineering that result in outlier agentic engineers with orders of magnitude more productivity than the median.
Sponsored
IBM Bob works directly in the codebase to generate and refine code in context. Describe what you need in natural language and get implementation where work is already happening. No switching between tabs, tools or half-finished ideas.
ByteDance is rumored to be launching its Dreamina Seedance 2.5 model on July 9. Availability is expected across Dreamina, CapCut, and other ByteDance partner platforms. The new model can output 180-second-long videos, but it is unclear whether it can keep character identity, motion, camera logic, and prompt intent stable.
Alibaba reportedly planned to prohibit employee use of Claude Code from July 10 after classifying it as high-risk software. Employees were instead directed toward Alibaba's Qoder tool amid Anthropic's efforts to prevent unauthorized access and model distillation.
OpenAI has moved GPT-5.6 into a narrow preview, splitting it into three tiers: Sol, Terra, and Luna. A notable change includes a new reasoning-effort control slider and an "ultra" mode for handling complex tasks. The release has implications for Codex users, amid Anthropic's Fable 5 changes, while broad access depends on US government review approvals.
Agents have made building cheap, but the cost just moved to verifying whether things actually work. The verification loop is the distance between a claim and its proof. Most teams close this loop with humans, but agents are starting to outpace humans. Compound Engineering is a plugin with skills for coding agents. This post discusses how /ce-dogfood, a skill within Compound Engineering, closes the loop and the persona strategy that gives it eyes.
Meta and SpaceX selling compute capacity could mean that there's excess compute capacity that two of the largest buyers no longer need. This may result in downward capex revisions for the hyperscalers. It's possible demand simply isn't there for what's being built with AI. However, this is unlikely, as anyone willing to sell capacity finds buyers immediately.
Pace Layers is a framework for organizing fields and categories by how fast they change. Each layer is a differently paced component that is functionally different from the others and operates somewhat independently, but each layer influences and responds to the layers closest to it in a way that makes the whole system resilient. This post applies the Pace Layers framework to the AI ecosystem to help readers process its dynamics.
Sponsored
Enterprise AI stalls when governance breaks down: sensitive data crosses boundaries, agents mask human credentials, and no one can answer basic compliance questions. CData Connect AI, [12] now HIPAA compliant, connects your AI tools to execute prompts across all your source systems. Join July 8th for a live walkthrough. Register here ->
This guide discusses the hardware used to run SOTA AI models locally and how to run speech-to-text (STT) locally. It also provides a ready-to-run configuration for running models within Docker containers. A budget of $2,000 can get you a setup that runs Qwen and good STT. $40,000 can get you a machine that can run almost-Opus-level models.
The difference between the map and the territory is the unknown. When Claude runs into an unknown, it needs to make a decision based on its best guess. The more work that is done, the more unknowns Claude might run into. This post discusses how to use Claude to help discover unknowns. Discovering your unknowns before starting a project is a cheap way to find out what you didn't know before it gets expensive to fix.
The Gap Map maps out the open-source AI stack so people can collectively build what's missing. The current open source AI stack is seriously robust, but fragmented, duplicative, and hard to see as a coherent whole. The Gap Map helps people identify where to build new, where to invest in capability, and where to open up the tools. It is an invitation for greater collaboration and coordination across the space.
As coding models commoditize, the real differentiator is the harness: the control loop that manages tools, workflows, orchestration, and model routing. Vendor-native harnesses offer the best performance today, but owning a portable, model-agnostic loop may become the more durable advantage as models evolve and costs fall.
Distillation started as a way to compress large models into smaller, cheaper ones, but has evolved into a core post-training technique for transferring instruction-following and reasoning abilities from frontier models. Methods used by DeepSeek, Qwen, and GLM now sit at the center of both open model development and ongoing disputes over whether training on proprietary model outputs constitutes unauthorized copying.
It's better to tell Fable to use its own judgment when deciding to write tests and when choosing other models for smaller tasks.
Leanstral is an open-source 119B-parameter theorem-proving and code-verification agent built on Mistral's general-purpose coding framework.
Google is testing a new dedicated inbox section within the Gemini app for Business and Workspace users.
Meta's superintelligence chief, Alexandr Wang, says that the company's upcoming model has caught up with OpenAI's GPT-5.5 on closely followed AI benchmarks. Codenamed Watermelon, the model is still in training. The model reportedly uses an order of magnitude more compute than Muse Spark. Meta has not given a timeline for Watermelon.
Anthropic reportedly discussed a custom AI chip collaboration with Samsung as it looked to diversify its compute stack. The company said chips from Google, Amazon, and Nvidia remained central to its hardware strategy.
Many folks claim to be able to use AI to do the work of dozens of people. This researcher used autoresearch to work on a problem where the path from unknown to success is a clear, gradient optimization. The researcher found the auto-research/loop style of work makes sense for problems with a robust, measurable, and well-constrained metric to optimize. However, finding a problem with these factors is often tricky.
The SGLang team outlined how agent workflows are being turned into reusable SKILL.md files, benchmark contracts, review loops, and production debugging playbooks. The post frames agent value around procedural engineering knowledge that can be executed, tested, and reviewed.
State-of-the-art block-wise Diffusion Large Language Models (dLLMs) rely on a remasking mechanism that decodes only the most confident tokens and discards the rest. Recycling computation from the discarded tokens is beneficial, as these tokens retain contextual information useful for subsequent decoding iterations. Residual Context Diffusion is a module that converts these discarded token representations into contextual residuals and injects them back for the next denoising step. It consistently improves frontier dLLMs in terms of accuracy with minimal extra computation overhead across a wide range of benchmarks.
Devin Security Swarm is a cost-effective and accurate way to find security vulnerabilities in complex codebases. It uses a new architecture for whole-codebase reasoning called Agentic MapReduce. Devin maps relevant signals across a repository, fans out focused agents over bounded shards, reduces their findings to one report, then verifies serious vulnerabilities in isolated sandboxes before marking them confirmed. Cognition has published extensive documentation and technical materials about Agentic MapReduce.
Laguna XS 2.1 is a 33B parameter Mixture-of-Experts model optimized for agentic coding and long-horizon tasks, showing a 5.4-point improvement on SWE-bench Multilingual to 63.1%. It supports various platforms and offers three quantized checkpoints for resource-efficient deployment. Licensed under OpenMDW-1.1, it enables open model distribution and is available for download on Hugging Face or via API.
Seed2.0 focused on long-tail knowledge, complex instruction following, reasoning, visual understanding, and search for real-world tasks. The model card described an evaluation-driven approach built around user needs and complex usage scenarios.
Woodside Energy leverages AI to enhance operational efficiency by integrating agentic AI systems in complex industrial workflows, particularly in LNG plant startups through their "Startup Advisor". These AI tools, built upon years of investment in predictive analytics and machine learning, augment human expertise rather than replace it.
In late June, AI hardware saw massive advancements with custom chips transitioning from concepts to actual products. OpenAI, Etched, Amazon, and SambaNova led these significant developments, marking a pivotal shift in the industry. These innovations promise to enhance AI processing capabilities and efficiency.
CursorBench evaluates agents on ambiguous, multi-file tasks from real Cursor sessions.
Claude Enterprise introduces enhanced admin analytics, model-level entitlements, and spend alerts to better manage usage and costs.
The Ramanujan Challenge for AI is designed to test whether AI systems can move from a concrete formula to a valid proof or symbolic derivation.
Sponsored
Coding agents only move as fast as the inference behind them. FriendliAI is the FRONTIER INFERENCE CLOUD FOR AGENTS — serving frontier open-weight models with the speed, reliability, and cost efficiency agents need in production.
Google is testing a new Gemini Flash upgrade on LM Arena, reportedly offering incremental improvements over the current Flash version. Despite no official confirmation, past Arena tests have often preceded public launches, drawing attention to potential new labels like "Gemini 3.6 Flash" or "Gemini 4 Flash." The Flash tier attracts interest due to its importance for everyday users and cost-conscious developers, handling most free and pay-as-you-go traffic faster than pricier Pro models.
Meta Platforms is developing an internal cloud infrastructure initiative to sell access to its surplus AI computing power and hosted models to external developers. This strategic pivot aims to generate a new revenue stream from the company's massive multi-billion-dollar investments in data centers and chips, directly challenging dominant hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud.
Dwarkesh Patel posted some big questions about AI a couple of months ago and received 600 essay submissions for the contest. The three winners of the competition were Jassi Pannu, an Assistant Professor at Johns Hopkins University; Ege Erdil, the co-founder of Mechanize; and Michael Li, a Master of Public Policy candidate at Harvard Kennedy School. The full essays are available in the post.
Autoresearch involves building an 'outer loop' where agents help maintain and improve the primary system using feedback signals, evals, and human input to make progress over time. One of the startups building infrastructure for these self-improving systems is Introspection. This post features an interview with Introspection's co-founder and CEO, Roland Gavrilescu. The discussion covers the shift from agent harnesses to feedback loops, the role of the open-source Pi framework, and why autonomous software factories must first learn from humans.