AI News: July 4, 2026

1. Anthropic Reportedly in Talks With Samsung on a Custom AI Chip

Anthropic is discussing a custom AI chip with Samsung, according to reports, as frontier labs look to trim the cost of training and serving models by owning more of their silicon. The company reportedly framed the effort as complementary rather than a break from Nvidia, which it says still matters for its stack. The move would follow rivals that have paired frontier ambitions with in-house or partner-built accelerators. Source

2. OpenAI Reportedly Proposed Donating 5% of Its Equity to a US Sovereign Wealth Fund

OpenAI proposed giving 5% of its equity to a US sovereign wealth fund, per a report on Sam Altman’s suggestion to distribute ownership stakes to a public investment vehicle. The idea would tie a slice of the company’s upside to broad public benefit, and it lands amid ongoing debate over how much AI wealth should be socialized. No fund or terms have been finalized. Source

3. Microsoft Plans an AI Super App With an Overhauled Copilot and Paid AutoPilot Agents

Microsoft is moving to consolidate its consumer and enterprise Copilot into a single application and add paid AutoPilot agents that handle background tasks, following Anthropic and OpenAI into the AI super app race. The plan reflects a broader industry bet that a unified assistant plus delegated agents is the interface most users will pay for. Source

4. Chinese AI Video Maker Kling Raises $2 Billion Ahead of a Hong Kong IPO

Kling, Kuaishou’s AI video division, raised $2 billion as it prepares for a Hong Kong public listing, one of the largest single rounds for a generative video business. The funding signals continued investor appetite for text-to-video despite crowded competition and rising compute costs. Source

5. Bridgewater and Thinking Machines Show a Fine-Tuned Open Model Beating Frontier Labs on Finance

Bridgewater and Thinking Machines Lab tested AI models on six finance tasks drawn from investor workflows and found frontier models such as Gemini, Claude, and GPT scored around 50% with basic prompts, rising to the mid-70s with expert instructions. A fine-tuned Qwen3-235B reached 84.7% accuracy versus 78.2% for the best frontier model, at roughly one-fourteenth the cost. The result argues that fine-tuning open weights on proprietary data and expert judgment can outperform general models on specialized work. Source

6. UK AI Security Institute Finds Benchmarks Underestimate What Agents Can Do

The UK AI Security Institute reported that standard evaluations systematically understate agent capabilities, because scores climb substantially once agents are given larger compute budgets at test time. The finding complicates safety and capability assessments that rely on fixed-budget benchmarks, suggesting headline numbers can hide how much more an agent can accomplish with more attempts. Source

7. Security Vulnerability Reports Surge as AI Models Hunt for Bugs

Reports of software vulnerabilities have climbed sharply as AI-powered discovery tools proliferate, with high-severity CVE reports reaching around 1,500 in June 2026. The spike coincides with the launch of several AI bug-hunting tools and raises questions about triage capacity as automated discovery outpaces the humans who validate and patch issues. Source

8. Zuckerberg Tells Staff Meta’s AI Agents Are Progressing Slower Than Hoped

Meta CEO Mark Zuckerberg told employees that the company’s AI agent development is moving slower than he expected, despite a reorganization around the technology. The candid admission tempers expectations for a company that has staked much of its roadmap on autonomous agents, and it echoes a wider pattern of agent timelines slipping across the industry. Source

9. Meta Quietly Launches Pocket, a Vibe-Coded Mini-Game App

Meta released Pocket, an experimental app that lets users generate small games from text prompts, without a formal announcement. The launch is a low-key test of consumer generative gaming and of how far vibe-coded software can go when the end user is doing the prompting. Source

10. Tesla Caps Employee AI Spending at $200 Per Week

Tesla set a weekly limit of $200 on employee spending for AI tools, according to internal communications. The cap is a small but telling sign that even well-resourced companies are starting to meter the runaway cost of agentic and coding tools that bill per token. Source