Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
The ARC-AGI-3 benchmark tests agents' ability to form and update an internal world model through ongoing learning. Continual Harness aims to efficiently enhance agents' self-improvement capabilities on this challenging IQ test.
Frontier models struggle with relatively simple financial tasks. Custom models fine-tuned on high-quality proprietary datasets labeled by expert investors perform better. They are also substantially cheaper. Models of the future will probably feature differentiated intelligence - organizations will have custom models tuned to their specific needs that outperform frontier models.
ZCode is now available on macOS, Windows, and Linux. It combines the best AI agents with existing tools so developers can plan, code, review, and deploy without friction. GLM-5.2 is tuned for ZCode, making agentic coding faster and steadier. GLM Coding Plan subscribers now have 1.5x usage quota in ZCode.
PorTAL is an innovative architecture designed to decouple task fine-tuning from specific base model weights. As the accelerating cadence of foundation model releases creates a compounding financial and engineering burden to re-tune capabilities from scratch, PorTAL offers a way to pay for task adaptation once and amortize it across all future models.
Claude Fable 5 and Mythos 5 have been redeployed. Fable 5 will be included for up to 50% of weekly usage limits until July 7, after which it will be available via usage credits. Mythos 5 has been restored for some organizations in the US. Anthropic is continuing to coordinate with the US government to expand Mythos access to the broader set of domestic and international partners in the Glasswing program.
A joint study by Ramp and Revelio Labs analyzed the relationship between firm-level generative AI investments and employment outcomes across over 21,000 companies in the US. The research reveals that companies with high-intensity AI spending grew their overall headcount by 10.2 percent and entry-level positions by 12 percent over the two years following adoption.
Sponsored
Most teams find out their AI agent is hallucinating or breaking guardrails from a customer complaint, not a dashboard. Regal's product and FDE team show you how to catch it first. Sign up for the Webinar
This article looks at how OpenAI uses WebRTC to run voice for its 900 million users.
Application layer AI companies cannot build sustainable moats against model providers through simple technical adjustments like fine tuning or model routing.
The proposal is part of a larger arrangement under which the US government would hold 5% of each of the leading US AI developers through a sovereign wealth fund vehicle.
Sponsored
Today's LLMs are brilliant, but have terrible memory. Every session starts at zero. You upload, re-explain, update... over and over.
Anthropic has introduced Claude Sonnet 5, a lower-cost Sonnet model with stronger agentic performance in planning, tool use, coding, and knowledge work. Its capabilities were described as approaching Opus 4.8 while improving substantially over Sonnet 4.6.
Google released Nano Banana 2 Lite, its fastest and most cost-efficient Gemini Image model, alongside Gemini Omni Flash for video generation and conversational editing. The models are available through AI Studio, the Gemini API, and Google's enterprise and consumer products.
Anthropic has received notice that the export controls on Claude Fable 5 and Mythos 5 have been lifted. The company will start restoring access tomorrow. Another update is expected soon.
Anthropic has announced Claude Science, an AI workbench app available in beta for Pro, Max, Team, and Enterprise users on macOS and Linux. The workbench integrates fragmented scientific tools into a single environment, natively rendering 3D protein structures, genome browser tracks, and chemical structures.
Thinking Machines is an AI research lab focused on human-AI collaboration. It believes that real work benefits from continuous collaboration, where the human clarifies, redirects, and gives feedback as a model goes along. This requires an interface that supports that instead of treating the human as someone who hands off a task and walks away. Thinking Machines' interaction models make interactivity a part of the model itself. The company plans to open a limited research preview in the coming months, with a wider release later this year.
AI models typically produce one token at a time - you can't compute the third token before you have the second. The GPU does most of the heavy lifting, but there is also some work that needs to be done by the CPU. GPU bubbles occur when the GPU sits idle in a loop waiting for the GPU to complete its job. This article looks at how to hide these bubbles using a technique called pipeline decoding, which involves starting the GPU work on the next token while the CPU is still finishing the last one.
Domain-specialized AI models consistently outperform generalized ones because finite resources require concentrated capacity. This pattern aligns across optimization mathematics, biological evolution, market competition, and machine learning, proving that universal generality is structurally inefficient under resource constraints.
In this podcast interview with Dwarkesh Patel, 3Blue1Brown creator Grant Sanderson discusses how AI's rapid, uneven progress in mathematics provides a roadmap for how it will transform the broader economy. Sanderson notes that while AI can rapidly brute-force fields like geometry, it still struggles with playful combinatorics problems that require deep conceptual creativity.
Sponsored
It's harder than it should be to evaluate agent performance—especially if waiting for a key event that might occur days later.
Meituan has officially launched LongCat-2.0, a massive 1.6 trillion-parameter Mixture-of-Experts model tailored specifically for agentic coding, multi-step workflows, and long-context processing. Notably, the model was unmasked as the engine behind "Owl Alpha," a highly popular stealth model on OpenRouter that recently ranked in the top three by global daily volume.
OpenAI's GeneBench-Pro is a benchmark that evaluates how AI agents handle ambiguity, revise assumptions, and choose analysis paths in computational biology. It focuses on research-level tasks across genomics, quantitative biology, and translational medicine.
Researchers set out to find out how good frontier LLMs are at doing frontier theory research. They designed a prover-verifier workflow that used GPT-5.5 Pro as the solver and Claude Opus 4.7 as the verifier. This workflow was stress-tested on open problems from areas with very different levels of familiarity. The results were surprisingly strong and the system resolved a list of open questions.
Miles is a framework for large-scale LLM RL post-training. It makes frontier-scale LLM RL easier to build, reproduce, and operate. RL post-training has become a distributed systems problem as models have become larger and run across more distributed and specialized hardware. Miles makes large-scale LLM RL training more composable, reproducible, and easier to scale while keeping the core trainer small enough for researchers and infrastructure teams to customize.
Claude Code fingerprints custom API routing inside model context. While Anthropic has real reasons to care about unofficial Claude routers, the implementation is not transparent. Its technique makes a line of model context look semantically neutral while using punctuation to carry routing metadata. It almost crosses the line into becoming spyware.
Anthropic is starting an internal drug discovery program to develop AI tools designed for drugmakers. The company will focus on discovering treatments for neglected diseases that traditional biopharmaceutical companies wouldn't consider attractive targets. It is unclear what Anthropic will do if it finds any promising drug candidates.
Sponsored
Granola's Recipes autonomously write follow-up emails, pull out decisions, prep for your next meeting, or turn conversations into real work in seconds. 1 month free with code [24]: TLDR1MO
Etched has already booked $1 billion in contract orders for full systems powered by its chip.
Guest users can only access a limited set of features, so it is unclear whether the performance gains will carry over to the full product.
Base44 aims to eventually outperform frontier models.
Sponsored
When IT consultancy Novacomp set out to modernize a business-critical Java REST API, the expectation was weeks of coordinated effort [JMG1] [5] . Instead, the work was completed in two days.
Devin Fusion is a multi-model harness from Cognition that mixes frontier and cost-effective models, reducing expenses by 35% on the FrontierCode benchmark while maintaining top-tier performance. This system uses dual-agent architecture with a main agent and a sidekick for dynamic model routing, optimizing task handling and avoiding costly cache misses. Fable 5 integration further enhances efficiency, achieving a 41% reduction in cost, promising advancements as models evolve.
All eligible users in the US can now access the Nano Banana-powered image generation feature within the Gemini app for free. The feature can generate images based on the AI model's understanding of users' likes and preferences without users having to specify them in the prompt. Personal Intelligence is an opt-in feature and users can decide which apps Gemini can access. Google has several updates for the Gemini app planned, including a new 'Daily Brief' feature, a revamped interface, access to the AI video model Gemini Omni, and a personal AI agent called Gemini Spark.
Cursor for iOS, now in public beta, allows developers to manage projects from anywhere using cloud or local agents. Launch or control agents via the mobile app, receive updates through Live Activities, and merge PRs on the go. This app supports workflows such as incident handling, resolving customer issues, and acting on user feedback efficiently.
RL in verifiable areas is clearly working. The next big leap will come from approaches that help bring the same achievements to things that are harder to verify. This article looks at why verifiability is a constraint, the techniques that are working now, and the companies attacking the problem.
AI is transitioning from closed, vertically integrated systems to a modular ecosystem supported by standardized interfaces like Transformer architecture and inference APIs. This architectural disaggregation enables open-weights models to compete effectively with closed systems, significantly reducing costs while accelerating innovation across the entire stack.
Sponsored
Framer is a professional website builder where AI agents make real, reviewable changes directly to your site. Trusted by teams at Miro and Perplexity, agents work across your canvas, components, CMS, and SEO to produce editable, production ready updates. Review every change in branches, compare diffs, and merge only what you approve.
DSpark is a system designed to make large language models answer faster without changing what the underlying model is trying to say. Most AI models write one small chunk of text at a time. DSpark acts as a scout that runs a few steps ahead, guesses the likely path, and lets the larger model quickly check which steps are safe. When guesses are good, the model moves faster, but if they are weak, DSpark tries not to waste time checking them.
RoadmapBench evaluates long-horizon coding tasks, spanning multiple files and languages, grounded in real version upgrades across 17 repositories. The benchmark tests 115 tasks, requiring agents to implement functionalities with a median modification of 3,700 lines across 51 files.
DiScoFormer, a transformer model, estimates both density and score from data in a single pass without retraining, surpassing classical kernel density estimation (KDE) in accuracy, especially in high dimensions. Utilizing cross-attention, it adapts to new data distributions on the spot, improving accuracy for generative modeling and Bayesian inference. Trained with Gaussian Mixture Models, DiScoFormer reduces score error by 6.5x and density error by more than 37x over KDE in 100 dimensions.
Google will start offering specialist AI models from SandboxAQ through Google Cloud. SandboxAQ's large quantitative models are trained on scientific equations and laboratory data. The addition of these models will widen enterprise and research access to AI built for drug discovery, materials science, and semiconductor manufacturing. Researchers can combine the models with Gemini, using the language model for reasoning and interface and the quantitative model for the underlying science.
Salesforce helped promote Claude Tag when it launched. This caused some confusion among employees, as Slack has its own Slackbot and Agentforce platform - which runs on Claude. Claude Tag offers a parallel experience inside the same platform. Salesforce expects to spend $300 million on Anthropic tokens this year, and it holds roughly a 1% stake in Anthropic, so it has financial reasons to promote Claude Tag despite the competitive tension.
Sponsored
Gartner talks with 80K+ business leaders and 10K+ IT executives to shape proactive, objective AI guidance. Get the Gartner intelligence
Mistral Workflows is a durable and fault-tolerant orchestration platform tailored for executing and monitoring multi-agent AI pipelines.
Fugu Ultra beat Fable on LiveCodeBench and starts at $5 per million input tokens.
This post follows a kernel from the code down to the warps, and back up to the answer.