Your AI agents have access. Who's controlling what they do with it? (Sponsor)
87% of IT leaders say an AI tool or agent has exceeded its intended scope in the past year. Only 1 in 3 can both revoke an AI tool's access and end an agent session immediately. Access alone isn't control.
Get Delinea's 2026 Identity Security Report to learn why standing privilege is creeping back into mature identity programs and how agents become over-permissioned.
See the full data and learn how to extend control from the point of access to the moment of action.
Get the report →
Introducing GPT-6.2 Astra and Sol (14 minute read)
OpenAI released GPT-6.2 Astra, a frontier model for the hardest projects, and GPT-6.2 Sol, an efficient everyday model. Both are available now in ChatGPT Work, Codex, and the API.
Microsoft's Seven New Models (4 minute read)
Microsoft AI released seven new models built for strong performance at low prices. Its MAI Code 2 coding model, now in GitHub Copilot, reaches 98% of GPT-6.1 Sol's coding quality at about 40% of the price. The lineup also covers cybersecurity, image editing, and real-time voice.
Introducing Replit Missions (8 minute read)
Replit launched Missions, a tool for coordinating swarms of AI agents to tackle complex projects like winning enterprise clients or developing software, within specified budgets. Their mission-built DataCat, an application monitoring platform that achieved faster performance benchmarks and surpassed a competing platform in API compatibility tests. This system enables small teams to pursue ambitious projects by defining goals and constraints, while agents execute tasks, providing businesses with a foundation for innovation at scale.
Horizon Labs (12 minute read)
Horizon is an agent based on Opus 5.5. Researchers gave it full access to a blank Hugging Face account and compute with the main goal of enabling real adoption on Hugging Face. Horizon used over 600 GPU hours and autonomously created 38 models. These models are now housed in over 100 GitHub repositories and have been downloaded more than 50,000 times.
Weight-Space Programming builds neural networks like a 1990s researcher (20 minute read)
Frontier models can read, write, and edit the weights of neural networks. This could present new ways to train networks. This researcher ran a weight-writing experiment with Claude Fable 5.2 on a familiar image classification problem. The way the model worked resembled the work of researchers from the 1990s, who used hand-crafted, pre-deep-learning techniques.
Verda Inference: Kimi K3.1 Pro and DeepSeek V4.1 Flash, engineered for AI agents (Sponsor)
Frontier open-weight models through one
OpenAI-compatible API, from the team behind the full-stack AI cloud. Run everything from simple prompts to long-running agents, with long sessions resuming up to 10x faster, at lower cost. Get every model at full quality and speed. Use code TLDR-75 for 75% off your first top-up.
Start building on Verda Inference →
Vercel's eve 1.0 (10 minute read)
Vercel released eve 1.0, its open-source framework for building production-ready AI agents. It manages the loop that calls models and runs tools, along with memory, approvals, and monitoring.
Co-Designing Sparse Attention for Efficient Long-Context Inference (17 minute read)
NVIDIA's sparse attention technique reduces compute and memory costs for transformers, solving bottlenecks in long-context inference by focusing on selective attention. Indexers in sparse attention score a subset of keys efficiently, aided by techniques like speculative decoding, cross-head reduction, and block-level selection to optimize performance on NVIDIA GPUs.
Become a founding member of TLDR's AI community (2 minute form)
TLDR is launching a curated community for AI builders to discuss major news and launches. The best links and comments will be featured in TLDR AI.
Apply here to join.
FrontierHarness Eval v1.1 (14 minute read)
The best coding-agent harness, the software that runs a model on a task, depends on the model. Harness choice alone can change cost per task by up to five times.
Cerebras launches disaggregated inference (6 minute read)
Cerebras' disaggregated inference service lets GPUs handle prompt processing while the company's systems handle output generation. The solution delivers five times more tokens per megawatt at the same ultrafast speeds. It allows Cerebras and partners to serve many more ultrafast tokens. Disaggregated inference is now in early access, and access is being expanded.
How should we define “capabilities activities” for the purpose of pacing? (21 minute read)
There is an urgent need to pace the frontier. This requires good, publicly verified measures of how much compute is being spent on pushing the frontier forward. Anthropic has already published estimates of how much of its AI R&D is safety research. AI developers should let third-party auditors measure and report how much compute AI companies are allocating to capabilities and non-capabilities activities.
Automate your software factory's flags, releases, and rollbacks (Sponsor)
LaunchDarkly Release Autopilot helps make your AI software factory possible by automatically wrapping PRs in flags, rolling out progressively, observing, and rolling back if
something goes wrong.
Toptal launches Toptal AI: one partner for all your AI needs (Sponsor)
Toptal AI develops, evaluates, and deploys frontier-grade AI using open-weight and proprietary models, while organizations retain full ownership and control over their AI systems and data.
Discover Toptal AI.
Microsoft-Decision-1.1 is now generally available for fast, low-cost decision making (34 minute read)
Microsoft's decision models are designed to make calibrated decisions in the face of uncertainty across a wide range of real-world software use cases.
Gemini Can Now Make Calls and Pay Bills (6 minute read)
Google's Gemini app can now handle tasks like calling businesses, booking appointments, and shopping within a spending limit.
Introducing Computer's Multiagent effort level (4 minute read)
Computer's Multiagent, a new effort level, enhances complex problem-solving by using multiple agents working in parallel.
Experiments in Weight-Space Programming (15 minute read)
GPT-6 Astra can look at an AI model, identify what task the network was built for just by looking at its raw weights, write the weights of a game-playing policy network from scratch, and improve models by editing model weights directly.