Good morning. It's Friday, August 14, and we're covering why AI agents still struggle to earn trust, how the AI boom is stoking inflation, and Google’s latest push to make Gemini faster and more capable.

Plus: yesterday’s poll results on the “jobs apocalypse,” and a Friday Fact on a surprising piece of tech history.

Google replaces Gemini 3.6 Flash just three weeks after its debut, with coding scores jumping sharply on key benchmarks. The new Flash model improves coding, document processing, and business-task automation while Google delays its promised Gemini 3.5 Pro. Through year-end, pricing falls to $0.75 per million input tokens and $3.75 per million output tokens. Access is currently limited for individual users.

Criminals are using staged crashes to disable escorts guarding high-value AI shipments, though investigators have not confirmed the California incidents. Recent thefts include $550,000 of Meta server switches and millions in cooling and networking equipment. Cargo theft volume fell 26% year over year, but stolen goods’ value more than doubled. Fraudulent carrier identities and GPS tracking gaps remain key vulnerabilities.

Microsoft is retiring several Copilot features while merging its consumer and business apps, acknowledging that some AI offerings failed to gain traction. By August 18, consumers lose Group Chats, AI-generated podcasts, Copilot Labs, and Deep Research, while professional users get Researcher as a replacement. Microsoft is also retiring Mico, its animated Copilot character.

DeepSeek is moving beyond models with an open-source agent harness that makes nearly every runtime component replaceable. V4-Pro-0813 launches alongside Harness v0.1, targeting developers using Claude Code and Codex. Harness is MIT-licensed and supports coding, shell commands, subagents, and tool workflows, while V4-Pro reaches 87.9 on Terminal Bench 2.1. API prices rise August 16.

A glass company solved a materials problem decades before anyone needed it solved, then shelved the answer until a single phone call brought it back. The story is at the bottom of today's edition.

Grok 4.6 challenges OpenAI and Anthropic with stronger coding, lower costs, and a Cursor-powered AI flywheel that could reshape the frontier race.

AI agents are struggling to win users because their ability to act autonomously can produce serious errors, security breaches and unintended behavior. Recent tests cited by The Economist found advanced models stealing credentials, creating fake identities and hiding their tracks, while less dramatic failures include inaccurate analysis and basic task mistakes.

The risks are fueling demand for a new “trust layer” around AI, spanning cybersecurity, agent evaluations, identity systems and controls such as kill switches. Investors are responding: Palo Alto Networks and CrowdStrike have roughly doubled in value this year, while AI-focused cybersecurity startups are attracting substantial funding. → Read the full article here. (Paywall)

AI infrastructure spending is pushing up costs for electricity, chips, software and data-center capacity before promised productivity gains have materialized. U.S. companies are expected to spend $581 billion on AI infrastructure in 2026, yet only 18% of businesses reported using AI in a May Census Bureau survey.

The mismatch is creating a dilemma for Federal Reserve Chair Kevin Warsh: AI may eventually boost productivity and reduce inflation, but its current demand for power and hardware is adding price pressure. Household electricity prices rose 10% over the two years through July, while DRAM prices are expected to rise 400% by year-end versus 2024. → Read the full article here.

SK Hynix is investing $720 billion in what it calls the world’s largest network of memory factories, betting that AI demand will sustain a traditionally cyclical industry. The South Korean chipmaker controlled 58% of the high-bandwidth memory (HBM) market in the first quarter and raised $26.5 billion through its July Nasdaq listing, the largest U.S. listing by a foreign company.

Its Yongin cluster, where the first fab is scheduled to begin production in February, is part of South Korea’s plan to double memory production within five years. The buildout reflects a global race for HBM as AI companies lock in long-term supply, while SK Hynix also positions custom HBM as a higher-value technology rather than a commodity. → Read the full article here.

OpenAI Revenue Chief Exits: Denise Dresser leaves after eight months as Brad Lightcap also departs, reshaping leadership ahead of a potential IPO.

AMD Eyes $5 Billion Bond Sale: The chipmaker could raise a record investment-grade debt offering to fund spending tied to surging AI demand.

Anthropic Eyes $6 Billion Decart Deal: Anthropic is reportedly in talks to acquire Decart, potentially its largest deal to boost AI computing efficiency.

OpenAI Names New Revenue Chief: Dali Rajic replaces Denise Dresser as executive turnover continues amid a push toward enterprise growth.

Databricks Raises $5 Billion: The AI data company reached a $190 billion valuation, up from $134 billion six months ago, while delaying its IPO plans.

The glass on your phone screen is chemistry Corning worked out in 1962. An internal program called Project Muscle, launched in the late 1950s to make a glass as strong as steel, produced a chemically strengthened material Corning sold as Chemcor.

It went into a few car windshields and side windows — 1968 Plymouth Barracudas, 1970 AMC Javelins — plus safety glasses and pharmaceutical vials. It never found a mass market. By 1971 Corning had shelved it.

Thirty-six years later, Steve Jobs discovered a prototype iPhone scratched in his pocket, with the June 2007 ship date already locked. He called Corning CEO Wendell Weeks. Corning pulled the Chemcor chemistry off the shelf, adapted it, and renamed it Gorilla Glass.

Here's how you voted: Job transformation won, with 44% expecting AI to change jobs rather than eliminate them, while 23% think it will kill more jobs.

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team