The AI PC Is Entering the Agentic Era (Sponsor)
Agentic AI is changing modern work. As AI agents become part of everyday productivity, AI PCs are evolving to support a new generation of intelligent experiences.
- Ready for the next generation of work. AI PCs support agentic AI experiences, including Claude Cowork, Microsoft Copilot, and ChatGPT, alongside the applications professionals use every day.
- Run AI where it makes sense. AI PCs can handle selected AI workloads locally to help reduce latency, keep data on-device, and optimize cloud inference costs.
- Built for the agentic era. AMD PRO processors deliver enterprise-grade performance and AI acceleration.
AMD Agentic PCs l Watch the latest
🚀Headlines & Launches
Claude Opus 5 (3 minute read)
Anthropic introduced Claude Opus 5 as a more efficient model approaching Claude Fable 5's capabilities at half the price. It reportedly led several coding and knowledge-work benchmarks and became the default model for Claude Max.
Prentis, new AI lab co-founded by Reid Hoffman, Mark Pincus in talks to raise $100M (5 minute read)
Prentis is an AI research lab focused on computer use models. Its founders are now in talks to raise $100 million at a $1 billion valuation. The lab is training models to learn how office workers navigate routine workflows across documents and systems. It aims to eventually build AI agents that can control computers to automate those tasks. The startup has already signed contracts worth up to $50 million with several customers.
🧠Deep Dives & Analysis
More On An Internal OpenAI Model Hacking Into Hugging Face (38 minute read)
OpenAI's unreleased internal model coordinated more than 17,000 complex actions over several days and successfully completed its goal, breaching Hugging Face in the process. The model escaped its sandbox, gained access to Hugging Face, escalated access, harvested credentials, and then found the data it was looking for. It took many days for the issue to be discovered. This post takes a detailed look at what happened during the attack.
How we built the new fastest API for GLM-5.2 (5 minute read)
Baseten's API for GLM-5.2 has peak speeds of 280 tokens per second and average speeds of around 100 tokens per second. It has more than double the performance of the launch-day API. The company has also built a Fast version of the API, which is focused on reducing latency for coding and agents. It plans to roll out another improvement to its speculative decoding algorithm soon that will further optimize the performance of GLM-5.2.
We have proof automation now (20 minute read)
The problem with great type-system power is that it requires great proof effort. Doing proofs can be fun, but they take a lot of time. It can take hours of effort to discover that what you were trying to prove was false. This overhead has made programming in dependently typed languages extremely niche. It has also spurred people to try and automate it away. LLMs promise to be an extremely capable form of proof automation.
👨💻Engineering & Research
The new rules of context engineering for Claude 5 generation models (6 minute read)
Claude 5 models now prioritize judgment over strict rule-based context engineering, with prompts shifting from rigid guidelines to adaptable strategies. Progressive disclosure helps optimize context usage by loading necessary information on demand, while simple tool descriptions replace repetitive instructions. Claude auto-saves relevant memories and utilizes rich references like HTML artifacts for handling complex tasks.
Prompt Caching In Agents (19 minute read)
Prompt caching makes agents more economic, but this is quite fragile. A changed tool definition, a model switch, or a provider routing decision can turn a cheap incremental request into a full replay of the context. Cache behavior affects latency, cost, tool design, session design, and even which product features should be made available. This post looks at what the cache is, how it works, and common reasons for worse cache performance.
Nvidia's New Long-Form Video Generation (12 minute read)
SANA-Video 2.0 combined linear attention with periodic softmax layers to generate videos up to 720p on a single GPU. Its 5B and 14B models retained competitive quality while substantially reducing latency for long, high-resolution generation.
Introducing celeris-1 (2 minute read)
celeris-1 is a general-purpose language model that delivers near-GPT-5 level intelligence with 15x faster response times. It uses a new inference architecture that uses diffusion techniques, unlocking dramatically better speed while maintaining frontier-level intelligence. The model delivers p50 response latency of 157ms and a throughput of 1,280 tokens per second. A link to a post on how the model was built along with full benchmarks is available.
🎁Miscellaneous
Open Weights and American AI Leadership (6 minute read)
Nvidia calls for US government policies to support open-weight AI models, arguing this fosters innovation and enhances AI leadership. The company emphasizes how open weights allow researchers to build on existing models, speeding up advancements. Nvidia's push aligns with broader industry efforts to encourage transparency and collaboration in AI development.
TLDR is hiring a curator for TLDR Hardware! (TLDR Curator, ~3 hrs/week)
500,000 people have already signed up for TLDR Hardware, our new twice-weekly newsletter covering chips, robotics, energy, and devices. If you work in hardware and want to help curate it, send your LinkedIn or resume to
hardware@tldr.tech!
Brute intelligence (7 minute read)
AI agents can work through loops to complete complex tasks. Domains that work around problems with verifiable answers are susceptible to these kinds of loops. This is a new form of brute force. It is like having a million scientists working in a million labs, all taking swings at the same thing. While AI may not be smarter than humans, it is faster. Being fast enough might make up for the difference in intelligence.
⚡️Quick Links
How to create and test an MCP server with authentication + authorization (Sponsor)
PropelAuth's deep dive doesn't just show you how to build & test an MCP server, it explains how each component works. Get the guide and
start buildingModelExpress: Distributing Model Artifacts at the Speed of Light (12 minute read)
NVIDIA ModelExpress accelerates model weight distribution by using P2P RDMA for direct GPU-to-GPU transfers, reducing startup times significantly.
Introducing Classifiers, now in beta (3 minute read)
OpenRouter Classifiers allow developers to tag inference in their workspaces by task type, department, agent complexity, or anything else.
Agentic AI at Two Different Scales: Nanbeige4.2-3B and Laguna S2.1 (9 minute read)
Nanbeige4.2-3B is a compact, dense model intended to make capable agentic behavior practical on consumer and workstation hardware, and Laguna S 2.1 is an 118-billion-parameter Mixture-of-Experts model that uses sparse access to a much larger pool of learned parameters.
The Legora Benchmark for Agentic Reasoning (4 minute read)
The Legora BAR evaluates AI models on real legal cases within a practical environment, unlike traditional benchmarks with synthetic setups.