TLDR AI 2025-10-02

WorkOS presents Enterprise Ready Conf: the flagship full-day event for product and engineering leaders shaping the future of Enterprise AI.

Join us in San Francisco on October 22 to connect with founders, executives, and developers leading AI adoption in the enterprise. Learn hard-won lessons from startups achieving breakout growth, and take away best practices in security, automation, and infrastructure at scale.

If you're responsible for the enterprise roadmap, this is where you get the insights that matter.

Register for Enterprise Ready Conf →

Thinking Machines has launched Tinker, a flexible API for fine-tuning open-weight language models. It abstracts infrastructure complexity, supports large MoE models like Qwen-235B, and comes with a low-level API and open-source Tinker Cookbook for implementing advanced post-training techniques.

Anthropic's 'Skills' will let users share repeatable and customizable instructions. It appears the company is formalizing a file-based standard for user-defined capabilities. Anthropic has yet to clarify which types of tasks or automations will be possible. No official launch date has been mentioned.

The launch of Sora 2, a TikTok-style feed for AI-generated videos, prompted mixed reactions from OpenAI researchers, fearful of repeating social media's addictive patterns. Sam Altman defended the launch by claiming that consumer revenue will fund AGI research, but many internally and externally are skeptical that OpenAI can grow entertainment products without compromising its nonprofit charter.

Anthropic claims that Claude Sonnet 4.5 is the best coding model in the world. While there are still clear use cases where versions of GPT-5 seem likely to be better, Claude Sonnet does seem like the best coding model in the world right now. GPT-5 still seems to make better use of extended thinking time than Claude Sonnet 4.5 does. Unless you must have an open model or need to keep inference costs very low, there is no reason to consider anything other than Claude Sonnet 4.5 or GPT-5.

Claude Code is a terminal-based application that trades accessibility for native Unix command integration. The commands that power Unix happen to be perfectly suited for use by large language models. Having file system access allows Claude Code to write notes to itself, accumulate knowledge, keep running tallies, and think beyond a single conversation. Claude Code works as a blueprint for building reliable agentic systems because it captures model capabilities instead of limiting them through over-engineered interfaces.

Running AI on your PC is transformative for productivity - but like all emerging tech, it comes with some security risk. Learn about the tactics attackers use to gain entry to your endpoints, and how you can stay secure. Read the eBook by Dell and Intel

Google DeepMind has built a custom generative image model fine-tuned for prototyping capable of producing design outputs that reflect specific stylistic features while enabling exploration of novel concepts.

Tunix is a JAX-native library that makes the path from pre-trained model to a fully aligned, production-ready large language model simpler. It gives developers a comprehensive and developer-friendly toolkit for aligning models at scale. Tunix is built for performance on TPUs, especially when combined with MaxText. The initial release provides modular and easy-to-use APIs for the most common post-training workflows.

Slack is reshaping how AI agents access and use enterprise data with a new platform that allows developers to tap directly into the rich conversational data flowing through workplace channels. Its new real-time search API and Model Context Protocol server give third-party developers secure, permission-aware access to Slack conversations, messages, and files. Slack is in increasingly intense competition with Microsoft Teams. Both companies are embedding AI throughout their collaboration platforms while taking markedly different approaches. Slack is focused on becoming an integration hub for other software experiences, while Microsoft is building a comprehensive suite of productivity tools.

YouTube plans to leverage AI to transform video creation, enabling creators to use AI tools like DeepMind's Veo 3 for video enhancements. CEO Neal Mohan highlights the potential for AI-generated content to democratize creation, although it raises questions about authenticity. Despite concerns, YouTube remains committed to AI advancements, aiming to stay at the forefront of video innovation.

The deal gives Meta access to Nvidia's latest GB300 systems and offers much-needed diversification to CoreWeave, which relies heavily on Microsoft and OpenAI. CoreWeave shares jumped 15% on the news.

GenAI is now mainstream, and companies are placing an increased focus on security, data quality, and cost efficiency. Read the Google Cloud research report: 2025 State of AI infrastructure

After receiving a cease-and-desist from Disney, Character.AI has removed various Disney-owned characters from its chatbot platform.

The AI chip maker claims 20x faster inference speeds than Nvidia GPUs and serves trillions of tokens monthly to major companies and government agencies.

Samsung and SK Hynix say that OpenAI's anticipated demand could grow to 900,000 DRAM wafers monthly.