TLDR AI 2025-09-12

It's unrealistic to completely avoid AI content in 2025. But without quality guardrails in place, it can end up surfacing irrelevant or poor quality results in site search — and badly frustrating users.

This Algolia white paper looks at practical ways to avoid this:

• Combining quantitative and qualitative methods to measure content clarity
• Systems and automations to evaluate readability, relevance, and accuracy
• Identifying and removing poor content that pollutes your information ecosystem

Read it here > Evaluating GenAI content for the optimum search experience

Microsoft and OpenAI will extend their partnership. The move could help OpenAI move towards becoming a for-profit corporation. OpenAI plans to create a new for-profit company in which Microsoft and OpenAI will each initially receive a roughly 30% stake. The rest will go to employees and investors. It plans to keep the nonprofit's control over the new for-profit and endow it with a stake valued at more than $100 billion.

Qwen3-Next is a new model architecture featuring a sparse Mixture-of-Experts design with hybrid attention and multi-token prediction. Its 80B-parameter base model activates only 3B parameters at inference, enabling 10x faster throughput for long-context tasks.

Claude starts conversations with a blank slate and only activates memory when users explicitly invoke it. It recalls by only referring to users' raw conversation history. ChatGPT's memory system exposes interaction metadata, recent conversation context, model set context, and user knowledge memories alongside the system prompt. That the two top assistants have built completely opposing memory systems shows that memory in AI has a massive design space with no right answer or one-size-fits-all technique.

Anthropic engineers found that agents perform better with fewer, more thoughtful tools rather than wrapping every API endpoint. Claude-optimized tools significantly outperformed human-written versions in internal tests, with agents able to automatically improve their own toolsets through evaluation loops.

The AI Security Summit will be held in person from October 22, 2025 - October 23, 2025! Join an exclusive gathering of AI innovators and execs confronting the AI Security Chasm—and building a foundation of trust for future-forward AI initiatives.

Register now

Infrastructure choices matter a lot for distributed training. Network and storage configurations can easily create 6-7x performance differences. This post looks at network storage and configurations that directly impact both training time and costs.

Cursor Tab is a system that predicts users' next actions across their codebase to make suggestions. The Tab model runs on every user action, handing over 400 million requests per day. This has given Cursor a lot of data about which suggestions users accept and reject. Cursor uses this data to improve Tab using online reinforcement learning. Its approach involves rolling out new models to users frequently throughout the day and using that data for training.

OpenAI's recently released GPT-OSS series of models features novel techniques like MXFP4 quantization, efficient kernels, a brand-new chat format, and more. The library has been upgraded considerably. This post looks at all of the upgrades in-depth. The updates make it very efficient to load, run, and fine-tune the models.

GPT-5's launch was so botched that Sam Altman apologized for it. However, it appears business adoption has continued to increase after a slowdown earlier in the summer. It looks like GPT-5's release drove a sharp rebound for OpenAI's tools. Adoption growth is outpacing Anthropic for the first time since May.

Meta's TBD Lab operates in a special badge-access area near Zuckerberg's desk, hidden from internal org charts. New hires routinely leverage competing offers from other frontier labs to secure massive compensation packages. ChatGPT co-creator Shengjia Zhao quit within a week of joining but returned after Meta tripled his pay and made him chief scientist. Meanwhile, existing employees lobby for raises while being excluded from the elite unit's secretive operations.

Rovo is Atlassian's AI-powered core, helping every team work smarter and faster with AI-powered search, chat agents, and more. Unlike alternatives, it comes with the security, compliance, and guardrails that build a trusted foundation. Read the ebook

Perplexity has secured another $200 million in funding, pushing its valuation to $20 billion.

Switzerland's open-source AI model, Apertus, is an alternative to proprietary models like ChatGPT.

Deep Loop Shaping, a novel AI method, enhances the stability of LIGO's control systems, reducing noise by 30 to 100 times.

This article outlines three key perspectives on AI's future, focusing on its development trajectory and societal impact.