TLDR AI 2025-10-21

Did you know 40% of AI pilots never achieve lift off? Thoughtworks' latest whitepaper uncovers why — and reveals how the FOREST framework can turn promising pilots into production-scale success.

The framework spans six key dimensions of AI readiness: foundational architecture, operating models, data readiness, human-AI experiences, strategic alignment and trusted AI governance — enabling you to scale AI with confidence.

Discover how leading organizations are achieving measurable impact, including boosting sales, driving down costs and increasing efficiency.

Read the Thoughtworks whitepaper and start building for sustainable AI success.

It is estimated that Anthropic spent $2.66 billion on AWS through September against $2.55 billion in estimated revenue, before accounting for Google Cloud costs or operations. Cursor's AWS bills doubled from $6.2 million to $12.6 million in June after Anthropic introduced Priority Service Tiers that sharply increased caching costs. Anthropic's costs appear to scale linearly with revenue - the path to profitability might have to be dramatic price increases.

Anthropic has launched a browser-based version of Claude Code. The beta allows users to connect GitHub repositories and delegate multiple tasks simultaneously across different repositories, with each session running in isolated sandboxes.

Anthropic launched a web app for Claude Code yesterday. It is now rolling out to Pro and Max users. Anthropic is attempting to evolve Claude Code beyond a command-line interface tool. It hopes developers will spin up AI coding agents in more places now that Claude Code is on the web. Claude Code has grown 10x in users since its broader launch in May. It now accounts for more than $500 million of Anthropic's annual revenue.

Pro-AI-bubble arguments are remarkably weak, given the amount of bubble talk. The recent series of circular deals done in close succession has raised concerns. If there is a massive financial crisis or bubble pop that affects AI, it will be mostly for mundane economic reasons rather than the technology not making rapid progress. Such a crisis will have, at most, modest effects on timelines.

Testing ChatGPT, Claude, and DeepSeek across English, Chinese, Arabic, Hindi, French, and Spanish revealed striking consistency in liberal, secular values regardless of language. All models strongly disagreed that sons deserve education priority over daughters, despite many humans worldwide answering yes. AIs seem to share a mostly cohesive, single underlying worldview across languages rather than adapting to local cultural values.

BERT-style models are essentially just text diffusion models trained on one masking rate. Masked language models can be repurposed into full generative engines by interpreting variable-rate masking as a discrete diffusion process. The standard masked language modeling objective can be turned into a step-by-step generation procedure by gradually corrupting text with mask tokens and training the model to iteratively denoise at increasing mask intensities. A fine-tuned RoBERTa can generate coherent-looking text after slightly modifying the training objective, even without architectural changes.

Geospot Infinity is a photo-to-GPS model. It retrieves 10 candidate coordinates for each upload and ranks them for the user. Users then pick the closest. The model attempts to learn from interactions via online reinforcement learning. It turned out that over 65% of users would just click the first guess regardless of accuracy, effectively nullifying the model's ability to learn anything.

Accelerate your AI/ML initiatives with AWS Marketplace solutions. Explore data architecture tools for secure scaling, discover implementation guides, and start your free trial today to transform your AI development from proof-of-concept to production-ready applications.

DeepSeek-OCR was created to investigate the feasibility of compressing long contexts via optical 2D mapping. It consists of DeepEncoder and DeepSeek3B-Moe-A570M. DeepEncoder serves as the core engine, while DeepSeek3B-Moe-A570M acts as a decoder. DeepSeek-OCR can generate training data for LLMs and VLMs at a scale of over 200,000 pages per day. Experiments show that when the number of text tokens is within 10 times that of vision tokens, the model can achieve a decoding precision of 97%.

Researchers built malware that uses Windows' built-in AI model to autonomously find and exploit security vulnerabilities without connecting to external servers. The technique only works on systems with NPUs or GPUs since CPU inference would be prohibitively slow, limiting deployment to CoPilot+ PCs and high-end computers until AI hardware becomes standard.

OmniVinci is a multi-modal architecture with novel techniques for aligning and encoding image, video, audio, and text data. The 9B-parameter model outperforms Qwen2.5-Omni across perception benchmarks while requiring just a fraction of the training tokens.

Adobe has released AI Foundry, a new service that allows enterprises to train Firefly-based generative models on proprietary branding data. The offering supports output across text, images, video, and 3D, and shifts pricing to usage-based billing for greater flexibility.

Anthropic is pushing into life sciences following OpenAI and Google with connectors to Benchling, PubMed, BioRender, and genomics platforms, alongside dedicated Agent Skills for tasks like single-cell RNA sequencing quality control. The company aims to support the entire research workflow from early discovery through regulatory compliance and commercialization, rather than just individual tasks like code writing or paper summarization.

Pinterest rolled out new tools that let users adjust how much generative AI content appears in their feed.

LAVA is a scheduling algorithm that continuously re-predicts and adapts to the actual lifetimes of virtual machines to optimize resource efficiency in large cloud data centers.

OpenEvidence raised $200 million at a $6 billion valuation three months after raising $210 million at $3.5 billion.

AI agents embed untrusted actors within themselves in their training, so fixing hallucinations is insufficient because even if an AI accurately interprets its inputs and produces corresponding output, it can be fully corrupt.