Search the archive

ANTHROPIC ACQUIRES SDK STARTUP STAINLESS

Anthropic acquired developer tools startup Stainless, whose SDK automation platform was widely used by AI companies, including OpenAI, Google, and Cloudflare.

CURSOR RELEASED COMPOSER 2.5

Cursor introduced Composer 2.5, an updated coding agent trained with targeted reinforcement learning, synthetic data, and new distributed training techniques.

WHAT POLITICAL CENSORSHIP LOOKS LIKE INSIDE AN LLM'S WEIGHTS

Qwen3.5-9B's political censorship is a small circuit that can be read and turned off. The factual knowledge is already in pretraining. The censorship behavior is layered on top of the facts. The model never loses the knowledge, it just learns to route around it.

AGENT EVALUATION: A DETAILED GUIDE

LLM evaluation has shifted from static benchmarks to more dynamic, real-world agent systems. Effective evaluation now requires realistic harnesses to test agents over long time horizons in complex environments. This is crucial as agents increasingly adopt high-stakes roles, such as coding and medicine, necessitating rigorous performance measurement and outcome-oriented evaluation.

GENERALIZATION DYNAMICS OF LM PRE-TRAINING

Language models (LMs) undergo unpredictable switches between parroting patterns and exhibiting adaptive intelligence during pre-training, a phenomenon termed "mode-hopping." This behavior cannot be corrected by standard optimization techniques and presents as a competition for model capacity, influenced by data from each training window. Researchers propose leveraging these dynamics to better select pre-training checkpoints, curate data for stable generalization, and evaluate metrics predicting LM behavior.

FINE-TUNING NVIDIA COSMOS PREDICT 2.5 WITH LORA/DORA FOR ROBOT VIDEO GENERATION

NVIDIA Cosmos Predict 2.5 generates videos from text, adapting for specific tasks like robot manipulation using LoRA/DoRA to inject trainable adapters, minimizing memory use. These methods offer efficient fine-tuning on a single GPU, preventing catastrophic forgetting while generating synthetic trajectories quickly. Fine-tuning with LoRA and DoRA significantly improves video quality, with LoRA more suited for tight memory conditions and DoRA preferred for addressing training instability.

HRM-TEXT

HRM-Text is a 1B text generation model based on the HRM architecture. It can be trained with 130-600x less compute and 150-900x less data than foundation models, making foundation model pretraining accessible. The 0.6B parameter version of the model can be trained on 8 H100s on a single node in about 50 hours for around $800. The 1B parameter model can be trained on 16 H100s on two nodes in about 46 hours for around $1,472.

VERA ARRIVES: NVIDIA'S FIRST CPU BUILT FOR AGENTS LANDS AT TOP AI LABS

The first Nvidia Vera CPUs recently arrived at Anthropic, OpenAI, SpaceXAI, and Oracle. They were hand-delivered by Nvidia Vice President of Hyperscale and High-Performance Computing, Ian Buck. Vera features 88 custom Nvidia-designed Olympus cores, 1.2 TB/s of memory bandwidth, and 50% faster per-core performance. It is the host processor for Vera Rubin NVL72, which pairs via second-generation Nvidia NVLink-C2C to a pair of Rubin GPUs.

JURY DISMISSES ALL CLAIMS IN ELON MUSK'S LAWSUIT AGAINST OPENAI CEO SAM ALTMAN

Elon Musk's lawsuit against Sam Altman and OpenAI has been dismissed. A jury has decided that Musk waited too long to file his lawsuit. Musk says he plans to appeal.

to make it work in real deployments. SAVE YOUR SEAT

SKILLS IN WEB, IOS, AND ANDROID

xAI launched "Skills" for Grok, allowing users to teach it functions once, which it remembers across interactions.

LLM WIKI V2

This post contains a pattern for building personal knowledge bases using LLMs.

FULLY REMOTE)

TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [21].

TURN REPEATED INSTRUCTIONS INTO REUSABLE SKILLS IN LOVABLE

Skills in Lovable allow users to create reusable, markdown-based instructions to eliminate repetitive explanations.

INTRODUCING SCHEDULED TASKS 2.0

Scheduled Tasks 2.0 enhances automation by allowing tasks to run with context, ensuring continuity in workflows across different projects and apps.

GEMINI APP ROLLING OUT ‘EXTENDED' THINKING LEVEL, NEW 3RD-PARTY APP INTEGRATIONS

Google is rolling out a new 'Thinking level' option for Gemini. The option has appeared for some users when they select Fast or Gemini 3.1 Pro. Google is also preparing to add more integrations with third-party apps in Gemini. Support for Canva, Instacart, and OpenTable appears to be coming.

CODEX WILL SOON BE ABLE TO CONTROL OTHER DESKTOP DEVICES VIA COMPUTER USE

OpenAI is working on a capability that lets its coding agent operate macOS applications through Computer Use even when a laptop is locked or asleep. Computer Use currently requires an unlocked, awake session to see the screen, move the cursor, and type. Lifting the restriction will allow users to direct their agents without having to walk back to their machines to log in first. It is unknown when the feature will be released.

CHATGPT PERSONAL FINANCE

OpenAI released a preview of a new personal finance experience in ChatGPT for Pro users in the US. The feature lets users securely connect financial accounts, view spending dashboards, and ask questions grounded in their financial context and goals.

TOKENOMICS: THE 62.5-MINUTE RULE FOR CLAUDE'S CACHE

If you expect to need a cache before 62.5 minutes, refresh it. Otherwise, let it expire. This number stays the same between models, and it doesn't change, no matter the size of the cache. The amount of dollars may change, but the decision point is still the same.

AI ECONOMICS PART 2

AI labs are in an ongoing war over GPU resources. That article looks into demand and supply and how the infrastructure powering AI today may not be sufficient. Scaling GPUs doesn't scale compute linearly. Efficiency matters more at raw scale given finite supply.

PORTABILITY IS A MYTH: WHY THE BEST AI STACKS WILL NEVER BE HARDWARE-AGNOSTIC

AI kernel portability is structurally impossible because TPU's Pallas, NVIDIA's CuTile and CUTLASS, AWS's NKI, AMD's FlyDSL, and Tenstorrent's tt-Metalium each expose hardware-specific concepts that no universal DSL can unify. The evidence: MaxText's MoE grouped matmul ships as 282 lines of Pallas on TPU while flashinfer's equivalent for Blackwell SM100 takes 4 million lines of generated CUDA, with zero shared code because the algorithms themselves diverge across hardware.

graceful failure handling in the May 26 workshop.

HOW CLAUDE CODE WORKS IN LARGE CODEBASES: BEST PRACTICES AND WHERE TO START

Claude Code is now being used in production across multiple large codebases in organizations with thousands of developers. These environments bring challenges that smaller codebases don't. This article covers patterns that Anthropic has seen that have led to the successful adoption of Claude Code at scale. It looks at how Claude Code has been used in monorepos with millions of lines, legacy systems built over decades, and microservices across separate repositories.

NOTES ON PRETRAINING PARALLELISMS AND FAILED TRAINING RUNS

Pretraining runs often fail. This article looks at all the ways that things can go wrong and why training is such a precarious operation. The key culprits seem to be breaking causality and adding bias.

RECENT DEVELOPMENTS IN LLM ARCHITECTURES: KV SHARING, MHC, AND COMPRESSED ATTENTION

KV-cache size, memory traffic, and attention cost quickly become the main constraints as reasoning models and agent workflows keep more tokens around for longer. LLM developers are adding a growing number of architecture tricks to reduce costs. Most of the changes look like small tweaks, but some are quite intricate design changes. This article looks at these architecture changes with a focus on what changes inside the transformer block, residual stream, KV cache, and attention computation.

LIGHTHOUSE ATTENTION

Lighthouse Attention, a selection-based hierarchical attention, offers up to 17x faster forward and backward passes than standard attention models at large contexts. It utilizes FlashAttention on a dense sub-sequence, maintaining efficiency and compatibility with upstream improvements. By enabling efficient long-context training and retaining dense model competence, Lighthouse Attention achieves 1.4x to 1.7x speedup in pretraining while reducing computational costs.

RUNWAY STARTED BY HELPING FILMMAKERS — NOW IT WANTS TO BEAT GOOGLE AT AI

Runway's founders believe that the next form of AI will be built from video and world models that learn how the world works. The company is training models directly on observational data to reach the next frontier of AI. Runway was one of the first to develop AI video generation, but world models are a different race with deep-pocketed competitors. The company has raised $860 million to date, but it is going against incumbents like OpenAI and Google.

THE HAVES AND HAVE NOTS OF THE AI GOLD RUSH

The AI boom has created a wealth divide, with an estimated 10,000 individuals from companies like OpenAI and Nvidia achieving over $20M in wealth, while others face uncertain futures with stagnant job prospects and layoffs. Software engineers express concerns about their skills becoming obsolete, raising anxiety about career paths. This disparity fuels tension in San Francisco's tech scene as some criticize the dual role of AI as a wealth source and a career threat.

FULLY REMOTE)

TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [21].

APPLE SILICON COSTS MORE THAN OPENROUTER

Openrouter costs about 1/3 the price at around 2x the speed for comparable models.

HEADROOM

Headroom compresses everything an agent reads before it reaches the LLM to produce the same answers at a fraction of the tokens.

DEEPSEEK-V4-FLASH MEANS LLM STEERING IS INTERESTING AGAIN

Steering is the idea that LLM outputs can be guided by directly manipulating the activations of a model mid-flight.

OPENAI QUIETLY BOUGHT VOICE-CLONING STARTUP WEIGHTS.GG, THEN FOLDED THE TEAM

OpenAI acquired the six-person team and its intellectual properties, then shut down Weights.gg and dispersed its team across multiple OpenAI groups.

Curious? See how it works in the live webinar on May 19th

INTRODUCING GROK BUILD

Grok Build is a coding agent that runs from the terminal. It is now in early beta for SuperGrok Heavy subscribers. AGENTS.md, plugin, hooks, skills, and MCP servers all work out of the box. Grok Build supports subagents for larger tasks, and it also supports deep worktree integrations, so users can launch subagents in their own worktrees. There is a headless mode that allows the easy running of agents inside scripts and automations.

CLOUD AGENT DEVELOPMENT ENVIRONMENTS

Cursor detailed a new system for configuring cloud-based development environments tailored to autonomous coding agents. It supports multi-repo, environment configuration as code, automated setup workflows, and governance controls for managing fleets of parallel agents.

OPENAI EXPLORES LEGAL ACTION AGAINST APPLE

Bloomberg reported that OpenAI explored legal options against Apple over dissatisfaction with how deeply ChatGPT was integrated into Apple's ecosystem and the limited subscriber growth that followed.

2028: TWO SCENARIOS FOR GLOBAL AI LEADERSHIP

Anthropic outlines two possible 2028 global AI leadership scenarios: one where the US retains its compute advantage and shapes AI norms, and another where China competes closely due to policy inaction. The US currently leads due to strong export controls and advanced chip technology preventing China from keeping pace. Closing loopholes on compute access and restricting distillation attacks are crucial for maintaining the US lead and ensuring democracies shape AI governance.

HOW WE BUILT SECURE, SCALABLE AGENT SANDBOX INFRASTRUCTURE

There are two ways to sandbox an agent that can execute code: isolate the tool or isolate the agent. Agents should have nothing worth stealing and nothing worth reserving. Isolating the agent requires an extra network hop on every operation and more services to deploy, but there are no secrets to steal, no state to preserve, and agents can be killed, restarted, and scaled independently.

Learn more → [11] | Book a demo →

CODEX IS GETTING EASIER TO AUTOMATE AND CUSTOMIZE AROUND YOUR CODE

Codex has implemented hooks and programmatic tokens to make it easier to automate and customize code. Hooks can customize the Codex loop with scripts that run at key points in a task. Programmatic access provides scoped credentials for Business and Enterprise teams. A video showing how to create access tokens for Codex automations is available.

RAINDROP WORKSHOP

Raindrop Workshop gives Claude Code the ability to read traces, write evals against codebases, and fix what's broken. It provides livestreamed traces, coding-agent integration, a self-healing eval loop, and local replay. Raindrop Workshop is compatible with TypeScript, Python, Go, and Rust, and most popular SDKs, providers, and coding agents.

GENKIT MIDDLEWARE

Genkit is a framework for building full-stack, AI-powered and agentic applications for any platform. It supports TypeScript, Go, Dart, and Python. Genkit uses composable hooks that intercept generation calls to implement retries and fall-backs for maximum reliability, human approval before destructive tool calls, and observability across every layer. Its middleware system runs a tool loop that repeats until the model is done. The Genkit Developer can be used to inspect, test, and debug applications and middleware execution.

UNLOCKING ASYNCHRONICITY IN CONTINUOUS BATCHING