TLDR AI 2025-08-26

Generative AI has turned financial fraud into an industrial operation. Voice cloning now takes just an hour of YouTube footage and an $11 subscription. Synthetic identity fraud already costs banks $6 billion annually. And fraudsters are scaling attacks using AI tools that traditional defenses can't handle.

Plaid and MIT Technology Review's new research reveals the true scope of AI-powered fraud—and why current prevention methods are like "bringing a stick to a gunfight."

The report looks at current fraud tactics—from credential-stuffing automation to AI-generated deepfakes targeting elderly victims—and how organizations are fighting back with their own AI systems.

Get your free copy from Plaid

Perplexity has launched a $5 standalone subscription that gives users access to premium publisher content by sharing revenue with participating publishers based on human visits and AI interactions.

Elon Musk's xAI filed a lawsuit accusing Apple and OpenAI of colluding to suppress competition in AI, citing Apple's partnership to embed ChatGPT in its ecosystem as anticompetitive behavior.

Modern language models use a building block called sliding window attention to handle long texts efficiently. The effective memory of these models has fundamental limitations. They struggle to use information from more than about 1,500 words ago, far less than the theoretical 100,000. This is because information gets diluted as it spreads through the network, and residual connections create an exponential barrier that blocks distant information.

In this interview, Jeff Huber from Chroma talks about what matters in vector databases, why modern search for AI is different, and how to ship systems that don't rot as context grows. Context engineering is the job of figuring out what should be in the context window at any given LLM generation step. It involves an inner loop, setting up the context window 'this time', and an outer loop that involves getting better at filling the context window with only relevant information.

Training frontier models? Turing remains the independent choice for the world's top AI teams — from multimodal training to RL gyms. 1000s of domain-trained experts across 60+ languages, 700+ benchmark tasks built from real-world business and STEM use cases, 50+ frontier projects supported successfully. Ask about free migration paths.

This framework enhances Claude Code with 21 new slash commands, 14 specialized AI agents, and 6 MCP server integrations to create structured workflows for development tasks. The project has gained 14k GitHub stars in a month and claims to reduce context usage by 30-50%.

ThinkMesh is a Python library for running diverse reasoning paths in parallel, scoring them with internal confidence signals, reallocating compute to promising branches, and fusing outcomes with verifiers and reducers. It works with offline Hugging Face Transformers, vLLM/TGI, and hosted APIs.ThinkMesh is confidence‑gated, strategy‑driven, and offline‑friendly. It is still in an early development stage.

Memento is a memory-based framework that enables LLM agents to learn continually without altering model weights. It uses a planner–executor loop with case-based reasoning and supports a wide range of tools.

The IVA framework trains Vision-Language-Action models to detect and respond to instructions based on false premises, using clarification and alternative grounding strategies.

Perplexity will start paying publishers for news articles from a $42.5 million revenue pool. The money will be distributed when Perplexity's AI assistant or search engine uses a news article to fulfill a task or answer a search request. The payments come out of the subscription revenue generated by Perplexity's new news service, Comet Plus, which will roll out widely in the fall. Publishers will get 80% of Comet Plus revenue.

Sebastien Bubeck, an employee at OpenAI, claims that gpt-5-pro can prove new interesting mathematics. Bubeck took the first version of a paper on convex optimization with a clean open problem in it and asked gpt-5-pro to work on it. The model proved a better bound than what was in the paper. The proof is correct - the second version of the paper (already published) had already reached this conclusion.

AI value is difficult to capture and quantify - but every AI startup has to get it right. Read the complete guide by Metronome: The Monetization Operating Model

Google's Gemini Live AI assistant will soon highlight items on your screen using your camera, initially available on the new Pixel 10 and other Android devices.

Andreessen Horowitz and OpenAI President Greg Brockman are among those launching the "Leading the Future" super-PAC network, which has already raised over $100 million to oppose strict AI regulations and support candidates favorable to the industry in the 2026 congressional races.

AI now leads unicorn creation, dominating over half of new $1B+ valuations in 2025.

MCPMark is a challenging benchmark built to stress-test models with comprehensive tasks that balance different CRUD (create, read, update, and delete) operations while enforcing strict environment management.