OpenAI rushes out GPT 5.2 while somehow convincing Disney to pay them $1B
🎄Happy holidays from Handy AI! Celebrate the season by grabbing a month-long trial of Handy AI Premium, which includes an extended weekly update and some other festive benefits.
last week’s top stories
🧠 OpenAI ships GPT-5.2. OpenAI released GPT-5.2 as its flagship “professional work” model family, tuned for long-running agent workflows and deeper tool use. The release framed GPT-5.2 as an upgrade in coding reliability, planning, and long-context execution, which matters more than benchmark vanity. Expect the competitive surface area to shift toward orchestration, evals, and cost curves, since model deltas keep shrinking. Read me
🎬 Disney and OpenAI signs a $1B pact. Disney committed $1B to OpenAI under a multi-year agreement that licenses a catalog of Disney, Marvel, Pixar, and Star Wars characters for Sora generation. The deal signals that top-tier IP holders now treat model platforms as distribution rails, with licensing as a core moat. The risk sits in governance around creator rights, brand safety, and downstream confusion between fan media and sanctioned media. Read me
🏛️ The White House orders a single AI rulebook push. A new executive order aimed to preempt state-level AI regulation in favor of a national framework, with federal agencies directed to challenge “obstructive” state laws. The intent is speed for builders and buyers, since compliance fragmentation taxes deployments. The fight moves to courts and Congress, where funding leverage and preemption theory will get stress-tested. Read me
👓 Google and Warby Parker set 2026 AI glasses timing. Google and Warby Parker disclosed a 2026 launch window for AI-powered smart glasses using Android XR and Gemini. Two paths: a screen-free assistant with sensors, and a version with an in-lens display for navigation and translation. Wearables really only matter once they stay on faces all day (so ergonomics and privacy indicators will decide the outcome). Read me
🔎 Google launches an embeddable deep research agent. Google released a developer path to embed its Deep Research capability based on Gemini 3 Pro into apps. Instead of a standalone “report generator,” the focus shifted to an API surface for agent interactions and controllable research workflows. This is a direct push toward agent platforms as infrastructure, with app builders owning the product loop. Read me
🗞️ White House signals Congress talks on AI framework. A White House adviser said the administration planned to work with Congress to craft unified AI legislation following the executive order. This matters because durable rules tend to come from statute, not memos, and companies plan capex around durable constraints. Expect heavy lobbying from both “ship fast” and “safety-first” coalitions. Read me
🪖 DoD launches GenAI.mil with Gemini for Government. The Department of Defense rolled out GenAI.mil as a controlled platform to deliver frontier models to military and civilian personnel, starting with Google’s Gemini for Government. The hard part will be prompt injection, data boundary control, and auditability under operational pressure. Read me
🧑🎓 Purdue requires AI competency for graduation. Purdue’s trustees approved an “AI working competency” requirement for all undergrads as part of a broader AI strategy. The signal: AI literacy shifts from elective to baseline, like spreadsheets or basic statistics. Universities that treat this as tooling plus critical thinking will produce graduates who can ship, audit, and govern systems. Read me
📱 Hinge’s leader left to build an AI dating app. Hinge CEO Justin McLeod stepped down to start Overtone, an AI-first dating product. The bet is that LLM-mediated coaching, matching, and conversation scaffolding can change retention and outcomes. Dating apps already run on ranking systems; this pushes the interface toward agent-like personalization, with safety and manipulation risk in the center. Read me
🧩 Salesforce expands Agentforce into a partner marketplace. Salesforce opened Agentforce 360 so partners can build, package, and sell agents, with deeper AppExchange provisioning and usage reporting. Partners ship agent “apps” while Salesforce supplies data pipes, identity, billing, and guardrails. Agent ecosystems will look like SaaS ecosystems (plus eval harnesses and tool-permission design). Read me
🧑💻 Mistral launches Devstral 2 for coding. Mistral released Devstral 2 in two sizes, targeting agentic software engineering and coding workflows. The licensing posture leaned permissive, aimed at adoption by enterprises and indie builders who want controllable stacks. This keeps pressure on closed labs, as coding assistants become commodity unless workflow integration stays ahead (though Mistral’s drop off in performance doesn’t help them). Read me
⚖️ OpenAI hit a trademark trap inside Sora. Reporting described a trademark dispute over OpenAI using “Cameo” as a feature name in Sora, followed by a forced rename. Product velocity collides with brand law once consumer apps ship at scale. Expect more of this as AI apps mint features faster than legal review cycles. Read me
🧪 AI Research of the Week
Multimodal AI generates virtual population for tumor microenvironment modeling
From Microsoft Research, Providence Cancer Institute, and University of Washington
Jake’s Take: This paper teaches an AI system (GigaTIME) to take the standard cancer microscope slide most hospitals already produce (H&E stain) and predict what a far pricier lab test would show: which proteins are active, and where, across the tissue.
The team trained the model on paired examples where they had both the cheap slide and the expensive “protein lighting” images, then scaled it across 14,256 Providence patients to generate about 300,000 predicted protein maps and hunt for patterns tied to biomarkers, disease stage, and survival.
The upside is scale. Biology studies that used to stall due to cost and throughput can turn into a data problem only, so researchers can search for signals across many cancer types before spending lab dollars. Worth nothing though that these outputs are model predictions, so they should only be driving hypotheses and triage what to test next (not settle clinical decisions without real lab confirmation).
and then, even more news…
💬 Anthropic pushed Claude Code into Slack. Anthropic added Claude Code capabilities inside Slack so teams can invoke coding help directly from threads, pulling context from conversations and connected repos. This move targets the real bottleneck: handoffs between chat, IDE, CI, and code review. Read me