Canva launches its own design model, adds new AI features to the platform (4 minute read)
Canva has launched a foundational model trained on its design elements that generates designs with editable layers and objects. The model works across different formats, including social media posts, presentations, whiteboards, and websites. Canva also recently introduced new products and features, updates to its AI assistant, and the ability to use its spreadsheet tool alongside its app-building feature to create data visualization widgets. More details about the recent launches are available in the article. |
Big Tech Is Spending More Than Ever on AI and It's Still Not Enough (7 minute read)
Big tech is planning to pour $400 billion into artificial intelligence efforts this year, but they all say it's nowhere near enough. Meta, Alphabet, Microsoft, and Amazon have all recently told investors that they will increase spending in 2026. Investors are uncertain about where the outsized spending will ultimately end. Whoever gets to AGI first will have an incredible competitive advantage over everybody else, so everyone is spending as much as they can. |
|
How Well Does RL Scale? (14 minute read)
Improving AI capabilities either involves scaling the amount of compute used for RL during training or scaling the amount of compute used for inference during deployment. It's important to know where capability gains come from because scaling up the inference compute has very different implications than scaling up the training compute. While RL has provided impressive gains, we've reached a point where it is too expensive to go much further. This leaves inference-scaling as the remaining form of compute-scaling. |
How fast can an LLM go? (13 minute read)
This article takes a look at how good inference software is becoming. It walks through calculations to provide estimates on what different hardware settings can theoretically achieve. Readers can adjust system configurations to change the values throughout the article. |
|
How we built OWL, the new architecture behind our ChatGPT-based browser, Atlas (10 minute read)
OpenAI built a new architectural layer called OWL (OpenAI's Web Layer) for Atlas, the company's ChatGPT-based browser, to run Chromium's browser process outside of the main Atlas app process. By moving Chromium out of the main application process and into an isolated service layer, OpenAI unlocked a simpler, modern app, faster startup, isolation from jank and crashes, fewer merge headaches, and faster iteration. Development can go faster because most of OpenAI's engineers don't need to build Chromium regularly from source. This article explains how OWL works. |
Introducing Aardvark: OpenAI's agentic security researcher (5 minute read)
Aardvark, currently in private beta, is a GPT-5-powered agent that autonomously scans code repositories to find security vulnerabilities, validate exploitability, and propose patches. It monitors commits in real-time, generates threat models for entire repositories, and integrates directly with GitHub workflows to deliver one-click patches, similar to Google's CodeMender. |
ImpossibleBench: Measuring Reward Hacking in LLM Coding Agents (9 minute read)
LLM-powered coding agents have been observed exploiting loopholes in tests or scoring systems rather than solving the actual tasks specified. ImpossibleBench was created to systematically measure this behavior. Its creators took existing coding benchmarks and manipulated their unit tests to directly conflict with the natural language specifications to create impossible tasks where models must choose between following instructions or passing tests. Their 'pass rate' on these impossible tasks is a direct measure of reward hacking. |
Kimi Linear Tech Report has dropped! ๐ (1 minute read)
Kimi Linear is a novel architecture that outperforms full attention with faster speeds and better performance. It offers up to a 75% reduction in KV cache usage and up to 6x decoding throughput at a 1M context length. Kimi Linear's open-sourced KDA kernels can be used as a drop-in replacement for full attention. The two models available were trained with 5.7T tokens. |
|
The Secrets of Claude Code From the Engineers Who Built It (1 hour video)
Claude Code creators, Cat Wu and Boris Cherny, discuss the product philosophy and technical workflows behind Anthropic's coding agent. They cover how their engineers use competing subagents for cleaner results, the team's "unshipping" approach to balance simplicity with power, and future form factors to make the tool more autonomous and accessible to non-technical users. |
|
| Want to advertise in TLDR? ๐ฐ If your company is interested in reaching an audience of AI professionals and decision makers, you may want to advertise with us.
Want to work at TLDR? ๐ผ Apply here or send a friend's resume to jobs@tldr.tech and get $1k if we hire them!
If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan, Ali Aminian, & Jacob Turner
|
|
| |
Comments