OpenAI's latest model emphasizes visual reasoning and computer use, a telling signal about where the frontier is moving. Speed and efficiency are the headline specs, but the real story is agentic capability — machines that can see screens and act on them.
Proton, the privacy-focused email company, is requiring employees to sit for AI competency interviews to determine who stays. "These re-interviews are consequential," a leaked internal email states. This is the AI reckoning going corporate-wide, and it won't stop at Proton.
Because Big Tech has gutted human customer service, ChatGPT is hallucinating 404 Media as a support destination — and the emails are flooding in. It's a darkly funny case study in what happens when AI fills a vacuum it doesn't actually understand.
Solana's security team has instituted a fee to filter out AI-generated junk flooding their vulnerability disclosure pipeline. The arms race between AI-generated attacks and AI-assisted defense has a new front: the bug bounty inbox.
Reducto CEO Adit Abraham makes the case that the document layer — PDFs, scans, spreadsheets — remains the critical unsolved problem blocking AI agents from consequential work in healthcare, legal, and finance. Reasoning is solved; ingestion isn't.
Goldman's new Small Business AI Institute data shows adoption nearly doubled in 24 months. The enterprise AI gap is closing faster than most enterprise software vendors would like to admit.
OpenAI's Ari Weinstein dropped a quiet bombshell: computer use is now faster than a human and getting cheaper. A16z flagged it because this is the threshold moment — when agentic automation stops being a party trick and starts being an economic substitution. The cost curve does the rest.
Mollick's standing advice crystallized again: always ask yourself how to make whatever you're doing with AI more ambitious, not just faster. It sounds obvious until you realize most people are using frontier models to autocomplete emails. Ambition is the unlock, not the tool.
A sharp research warning: a 120B model handles real-world MCP tool failures no better than a 20B model. Scale isn't saving you here. If you're building production pipelines on MCP integrations, the failure modes are nastier than benchmarks suggest — and the size of your model won't bail you out.
The through-line today is substitution pressure hitting from every direction at once. Proton's AI re-interviews aren't an anomaly — they're the template. Companies now have a forcing function to ask "who here actually knows how to work with these tools?" and the answer reshapes headcount. Meanwhile, OpenAI's GPT-6.1 Terra lands with computer use as the marquee feature, and Ari Weinstein confirms it's already faster and cheaper than a human at screen-based tasks. These two facts are not unrelated.