Same input. Same prompt. Different output. That's the reality of testing AI agents that write code, and most teams are shipping without solving it.Nick Nisi from WorkOS tackled this by building eval systems for two AI tools:- npx workos@latest, a CLI agent that installs AuthKit into your project- WorkOS agent skills that power LLM responses about SSO, directory sync, and RBAC.The post covers how to test against real project structures, score output that's different every time, and catch when your agent makes up methods that don't exist.Learn more about evals →

Anthropic appears to have started a fresh round of red teaming on a new internal build. The company is set to host its Code with Claude developer conference in San Francisco on May 6. The timing suggests that the model is being hardened ahead of an announcement timed to the event. The red team round is consistent with the company's responsible scaling policy, which calls for jailbreak probes and constitutional classifier stress tests before any frontier-class deployment.

Google is testing a new Omni model for video generation, potentially unifying its video and image-generation tools. The Omni model appears in Gemini's video generation UI, suggesting it might become a public product name. A launch during Google I/O 2026 is possible amidst increasing AI video competition.

OpenAI updated Codex with animated Pets, which appear as overlays on the screen and interact via short message bubbles. Codex also now auto-imports configuration files from other coding agents and features a new dictation dictionary to improve voice input accuracy. These updates aim to enhance Codex's usability and appeal as a comprehensive desktop application.

The question isn't “which tool has the best model?” It's “which solution will our team actually use?” This Notion guide breaks down the 5 critical jobs AI should solve at work and how to evaluate tools for adoption and integration, not just capabilities. Get the guide

A scalable method generates realistic virtual computer environments and long-horizon simulations, producing rich training signals that improve agent performance across productivity tasks.

AutoRound is an advanced quantization toolkit designed for large language models and vision-language models. It achieves high accuracy at ultra-low bit widths with minimal tuning. AutoRound seamlessly works with Transformers, vLLM, SGLang, and more. It can quantize 7B models in 10 minutes on a single GPU.

Edit-R1 introduced a chain-of-thought reward model that evaluates image edits through structured reasoning, improving alignment and performance in text-guided editing tasks.

Replit's Amjad Masad highlights strong growth, nearing a billion-dollar run rate, and boasts a 300% net revenue retention rate. Unlike Cursor, which struggles with negative margins, Replit maintains gross margin positivity and appeals to non-technical users with its secure, end-to-end platform. While Masad remains committed to Replit's independence, he acknowledges open discussions with potential acquirers and expresses frustration with Apple's alleged discriminatory App Store practices, suggesting possible legal action.

This post walks through the inference pipeline from tokenization and embeddings through stacked self-attention layers, then splits generation into two distinct phases on the same GPU: compute-bound prefill that processes all input tokens in parallel and memory-bound decode that emits one token at a time.

Mode collapse occurs when models repeatedly generate the most common outputs, leading to homogenous results, exemplified by AI generating more dogs over cats with unbalanced training data. It similarly impacts various domains like grant-making and music, as systems become increasingly specialized over time based on prior outputs and successes. To counteract, introduce variability or change external pressures to diversify and prevent over-specialization.

Granola takes your raw meeting notes and makes them awesome, no awkward transcription bot required. Now TLDR readers get one month free with code: TLDR1MO. See why people love Granola

Many of the companies have said their deals with the Department of War include commitments that their tools wouldn't be used for mass surveillance or autonomous weapons.

Comparing open-source models to closed APIs is flawed, as they serve different purposes.

TLDR is looking for an engineer/researcher at a major AI lab or startup to help write for 1M+ subscribers. Our curators have been invited to Google I/O and OpenAI DevDay, scouted for Tier 1 VCs, and get early access to unreleased TLDR products. Learn more.

The investors want to create a company that helps teach businesses how to incorporate AI across their operations.

One global vLLM pool is a poor default for mixed traffic.