Codex now works directly in Chrome on macOS and Windows (2 minute read)
OpenAI Codex now works directly in Chrome on macOS and Windows. It works in parallel across tabs in the background without taking over the browser. The implementation can quickly move through repetitive browser work like navigating structured pages and complex data flows. It writes code to navigate and complete tasks under the hood. Video demos showing what users can do with the feature are available in the article.
|
OpenAI Released Realtime Audio Models (8 minute read)
OpenAI released a new set of real-time audio models, including GPT‑Realtime‑2 for conversational reasoning, GPT‑Realtime‑Translate for live multilingual translation, and GPT‑Realtime‑Whisper for streaming transcription.
|
Meta prepares Hatch AI Agent with waitlist and social skills (2 minute read)
Meta is developing Hatch, an AI agent positioned as a consumer-grade competitor to OpenAI's OpenClaw, with features for image and video generation, shopping, and learning integrated deeply into social platforms like Instagram and Facebook. Internal tests are expected by June, using simulated environments, with a potential wide release gated by a waitlist. Meta also plans a shopping tool for Instagram by Q4, aiming to keep users engaged within their existing social spaces.
|
|
Improving token efficiency in GitHub Agentic Workflows (12 minute read)
GitHub Agent Workflows significantly improve repository hygiene and quality, but costs are becoming a growing concern for developers. AI jobs like agentic workflows are automatically scheduled and triggered, so costs can accumulate out of view. GitHub started systematically optimizing the token usage of many workflows last month. This post describes what the team instrumented, the optimizations it applied, and its preliminary results.
|
The Six-Hour Codex Run That Survived a Five-Hour Pause (10 minute read)
/goal is a headline feature that shipped in Codex on April 30 that introduced persisted goals, goal states that survive terminal restarts, laptop sleeps, and multi-hour passes without re-prompting. The feature injects a developer message on resume rather than waiting for the user to time anything. It allows users to turn off their devices and pick up exactly where they left off without re-prompting anything.
|
Good QC for RL Data (18 minute read)
There needs to be a higher bar for what good quality looks like. Any vendor selling into a frontier lab is being judged implicitly during the purchase decision, and most are failing multiple quality control gates at once. Standardizing QC for evaluating how well data tests something on the performance-cost-latency Pareto curve is of the utmost importance. The vendors who don't internalize a higher QC bar will start running into problems this year.
|
|
AlphaEvolve: How our Gemini-powered coding agent is scaling impact across fields (9 minute read)
AlphaEvolve is a Gemini-powered coding agent that can design advanced algorithms. It can help make new discoveries on open problems across mathematics and computer science. The agent has been upgraded with the ability to help explain the physics of the natural world to help accelerate progress for scientists and businesses across a variety of fields. This post shares the areas AlphaEvolve has had the most significant impacts on to date.
|
Building Fast & Accurate Agents with Prime-RL Post Training (22 minute read)
Ramp Sheets built Fast Ask to handle its spreadsheet agent's information retrieval loop. It can navigate a workbook, read the relevant ranges, and return a compact answer for the main agent to use. This post presents Fast Ask as a case study for reinforcement learning. It covers where it is worth training a specialized agent, how to design the environment, and how to evaluate whether post-training worked.
|
ds4.c (GitHub Repo)
ds4.c is an intentionally narrow, small native inference engine for DeepSeek V4 Flash. The project aimed to create one local model that felt finished from end to end. It is Metal-only, but the development team may implement CUDA support in the future. The project is still in alpha, so there are still bugs in the implementation.
|
Natural Language Autoencoders (9 minute read)
Anthropic introduces Natural Language Autoencoders (NLAs) to translate AI model activations into human-readable text, aiding in understanding model thoughts. NLAs have been used to detect safety concerns and hidden motivations in AI behavior, improving model alignment auditing. Despite limitations like hallucinations and high costs, NLAs advance AI auditing techniques. Anthropic has released training resources for further development.
|
|
Notes from inside China's AI labs (18 minute read)
Chinese and American labs look similar when you look at their outputs and ingredients, but there are major differences in how they are organized and conditioned. Chinese scientists are much more willing to do non-flashy work to improve a model rather than push their own ideas. This results in less gamification of the system and more flexibility in adapting modern techniques. The Chinese AI community feels more like an ecosystem than battling tribes, with all the labs having nothing but respect for their peers. They are also much less focused on the business side of the technology.
|
Long AI Short AGI (3 minute read)
Silicon Valley's narrative emphasizes AGI as the ultimate scarce resource, but the rapid commoditization of AI models challenges this. Intelligence now follows the same path as compute, bandwidth, and storage, where market forces drive competition and reduce costs. The real winners in AI won't necessarily have superior models but will own customer relationships and proprietary data, much like past tech giants.
|
|
|
Love TLDR? Tell your friends and get rewards!
|
|
Share your referral link below with friends to get free TLDR swag!
|
|
|
|
|
Track your referrals here.
|
|
|
|
