TLDR AI 2026-03-05

How is your competition pricing AI? The Metronome Pricing Index shows you major vendors' pricing structures in one place.

Get breakdowns of credit systems, hybrid models, and the packaging patterns winning enterprise deals for 39 AIs and counting: aws, Cursor, OpenAI, DeepL….

👉 If you're building AI products, this is the competitive intel you've been looking for.

Explore the Pricing Index for free — no registration required.

OpenAI's upcoming GPT-5.4 model will have a one-million-token context window, putting it on par with Google and Anthropic's offerings. The model will feature an 'extreme' thinking mode that lets it burn significantly more compute on tough questions. It is expected to be more reliable and make fewer mistakes on longer tasks that can run for several hours. OpenAI's more frequent model release cadence is designed to keep expectations in check. The hype around GPT-5's launch set a high bar that was nearly impossible to clear, and the company's user growth has recently fallen short of internal projections.

The Qwen team has just seen some very high-profile departures. Junyang Lin, the lead researcher building Qwen, recently announced his departure on X. Following this, several other Qwen member also announced their departures, including Binyuan Hui, who was responsible for the entire agent training process from pre-training to post-training; Bowen Yu, who led the development of the Qwen-Instruct series models; and Kaixin Li, a core contributor to Qwen 3.5/VL/Coder. The new Qwen 3.5 models appear to be exceptionally good, so it is an interesting time for the team to disband.

Perplexity is expanding its Computer platform with "Skills," enabling users to integrate custom and generic markdown-based reusable workflow instructions. The new feature automates complex tasks and aims at professionals needing consistent outputs. Additionally, a "Final Pass" mode for document reviews is in development, enhancing document handling capabilities.

AI models now handle complex tasks without micromanagement, but require pre-planned workflows. Sketching workflows with decision branches and using images for AI setup streamline processes. This approach allows agents to work independently, producing ready-to-use outputs.

People sometimes say that neural networks are just logistic regression. In some sense, neural networks are logistic regressions with more parameters. However, using more parameters makes it different. New phenomena emerge at scale that could not have been anticipated at a smaller scale.

The upper bound of power and influence for AI labs appears to be infinite. This has led investors to focus everything on the technology. The tech industry has become incredibly politically engaged as productivity shifts from labor to capital. When AGI or similar is achieved, lab incentives will permanently transition to appease the shareholders who own the capital. AI safety is the last line of defense in the shift. It must be embedded into the tech industry's technical and social infrastructure before IPOs and competitive dynamics make it permanently impossible.

88% of AI proof of concepts never make it to production. Tines' intelligent workflow platform combines deterministic automation, AI, and human-led steps to run workflows you trust in production. Power your most important workflows today with Tines free Community Edition.

Controlled pretraining experiments that explore how text, images, and video should be combined when building multimodal models from scratch.

OpenPencil is an open-source, AI-native design editor compatible with Figma that enables design file manipulation without vendor restrictions. It features real-time collaboration, AI integration, and a headless CLI for versatile design operations.

Phi-4-reasoning-vision-15B is a compact open-weight multimodal AI model that matches or exceeds the performance of systems many times its size. The 15-billion-parameter model can process both images and text and can reason through complex math and science problems, interpret charts and documents, and navigate graphical user interfaces. It is available now through Microsoft Foundry, Hugging Face, and GitHub under a permissive license. The model was trained on only approximately 200 billion tokens of multimodal data, much less than competing models.

Anthropic CEO Dario Amodei criticized OpenAI's defense contract with the US Department of Defense, calling the company's public framing of safety protections misleading.

The contracts that AI companies had with the Department of War prevented the use of their technology if it would potentially lead to a kinetic strike or other similar scenarios. Emil Michael, the Undersecretary of Defense for Research and Engineering at the Department of War, believes that the US military should be able to use the technology it selects for all legal purposes. Vendors should not have a say in how their software should be used if it is legal. The software or company ideals should not dictate America's command and control environment or tell generals and war fighters what to do or what not to do.

Replay is Temporal's practical conference for devs building real systems. Join for hands-on workshops and leave with prod-ready AI agents. Save with code TLDR75

Future AIs may sacralize shared values like helpfulness, harmlessness, and honesty to coordinate but risk becoming less effective.

AI-native companies like Linear, Ramp, and Factory integrate AI agents deeply into their operations, treating them as essential team members across all functions.

Arm's technology powers 99% of smartphones, enabling on-device AI for billions of users globally.

Google released Canvas inside AI Mode in Search, giving users a workspace to draft documents, build dashboards, or prototype small coding projects directly within search results.