OpenAI is raising again — this time targeting a $1.4 trillion valuation while delaying its IPO and launching Dots, a personal agent platform that remembers context across your entire digital life. DevDay 2026 made clear OpenAI isn't just building models anymore — it's building the operating system for AI-native work.
A competitor is explicitly marketing itself to police departments as the facial recognition layer Flock won't build. The CEO framed mass biometric surveillance as simply "the way the world will have to be."
Anthropic quietly dropped a significant efficiency upgrade — Sonnet 5.5 beats its predecessor across the board while costing less to run. At the same price point, it's an immediate no-brainer swap for developers.
Anthropic's Frontier Red Team found that both GLM-5.3 and Claude Mythos Preview can develop full control flow hijacks on a binary exploitation benchmark — without any specialized training. The numbers are small but the trajectory is not.
Robinhood is giving retail customers an AI agent that trades autonomously around the clock, paired with 10x leveraged crypto products. The risk sits entirely with the user — Robinhood just takes the fees.
A team used off-the-shelf frontier LLMs to navigate a parking lot course with no autonomous driving training data whatsoever. It worked. The gap between "language model" and "physical world agent" is closing faster than the auto industry expected.
Mollick noted that OpenAI has now cycled through Plugins (2023), GPTs and the GPT Store (2024), and now Dots — releasing a third-party ecosystem platform each year before semi-abandoning the last one. It's a sharp observation: the graveyard of OpenAI platform bets is getting crowded, and developers building deep integrations on any single layer should take note.
Andrew Ng flagged that the OpenAI-Hugging Face security incident was enabled by weak sandboxing, and praised Nvidia for releasing open-source sandboxing tools for AI agents. As agentic systems proliferate, the attack surface explodes — and sandboxing is one of the few unsexy but critical infrastructure pieces nobody wants to build first.
The Kobeissi Letter flagged that only ~10% of US software-spending businesses are now using paid GPU infrastructure — but that share is at a record high and climbing fast. AI compute spending is concentrating, not democratizing. The companies locked into that 10% are pulling ahead; everyone else is running on borrowed time.
a16z's Assistant Benchmark creator David Pawlan argued the internet is about to grow an agent-to-agent layer — infrastructure that lets AI agents communicate, negotiate, and transact with each other without humans in the loop. If he's right, the API economy we've spent a decade building was just the warm-up act.