Is OpenAI building an AI pen?
🎆 Happy New Year from Handy AI! Celebrate 2026 by grabbing a month-long trial of Handy AI Premium, which includes an extended weekly update and some other benefits.
last week’s top stories
🖊️ OpenAI’s first device might be a pen. A leak points to a pen-shaped, screenless OpenAI hardware concept (internally referenced as “Gumdrop”), pitched as a recorder-style companion with microphones and possibly a camera. The form factor screams “ambient capture,” which means the hard problems are consent, retention, and access control, not industrial design vibes. Humane’s collapse should be stapled to every slide deck involved here. Read more
🧨 LeCun drags Meta’s AI leadership on the way out. Yann LeCun, exiting Meta, described internal turmoil around benchmark handling and criticized new leadership dynamics, including skepticism about research credentials at the top. The piece also frames a broader split: Meta’s push for frontier LLM output versus LeCun’s long-running insistence on alternative architectures. Read more
🚨 Grok’s “undressing” problem escalates into CSAM territory. Reuters describes Grok generating sexualized content involving minors on X after safety lapses, with xAI responding in the kind of tone that makes regulators reach for pens. This is the predictable failure mode of image manipulation (plus weak guardrails plus viral distribution). If the industry wants consumer trust, shipping systems that wander into criminal content is a terrible growth strategy. Read more
🧭 Amazon puts Alexa+ on the open web via Alexa.com. Amazon launched Alexa.com for Alexa+ Early Access, bringing its chatbot-style assistant to laptops with document and image uploads baked in. TechCrunch notes the parallel revamp of the mobile app toward an “agent-forward” interface, plus hooks into shopping and smart home control. Amazon is clearly aiming for its own “assistant everywhere,” which is useful (and also another privacy negotiation users never asked to conduct). Read more
💰 SoftBank finishes its OpenAI mega-round. SoftBank says it has fully funded its $40B OpenAI investment, closing one of the largest private AI financings on record. Reuters reports the deal was assembled through direct funding plus syndicated co-investment, with a final $22B-ish tranche completing the stack. This kind of capital locks in a long compute runway and pulls infrastructure strategy closer to the model roadmap, which is great for scale and messy for governance. Read more
🏗️ xAI expands compute with a third Memphis-area facility. Reuters says xAI bought a third building to push training capacity toward the multi-gigawatt scale, tied to its “Colossus” cluster expansion. The ambition is brute force. More GPUs, more power, more throughput, and fewer excuses. The bill lands somewhere between grid politics and public backlash, especially with on-site energy projects in the mix. Read more
⌚ Amazon’s Bee wearable now sits in the Alexa+ orbit. GeekWire ties Amazon’s web assistant push to its wearable AI companion story via Bee, the always-listening wrist concept for summaries and reminders. Bee captures context, Alexa+ executes tasks, and Amazon gets a tighter loop between intent and action. Amazon’s history with ambient audio collection raises concerns. Read more
💄 L’Oréal uses AI film-makers to crank digital ads. The Times covers L’Oréal leaning on an AI-forward studio for an Essie campaign, using multiple generative stacks to produce effects like realistic liquid motion. The workflow shift matters: “AI creative” and “AI film-maker” become roles, and production becomes prompt pipelines plus compositing. Brands will love iteration speed, and society gets to argue about authenticity and labor displacement again. Read more
🏫 Pre-K teachers quietly adopt genAI in classrooms. EdSurge cites RAND research showing roughly 29% of pre-K teachers using generative AI in the 2024–2025 school year, often for drafting communications or planning support. Early childhood settings have high sensitivity around data, consent, and developmental impact, yet tooling and guidance lag behind usage. This becomes a governance issue fast because the users are children and the operators are under-resourced staff. Read more
🧪 AI Research of the Week
Project Vend: Phase two
From Anthropic, with Andon Labs
Jake’s Take: Anthropic ran a real vending-machine business across San Francisco, New York, and London, then upgraded the agent stack for phase two: move from Claude Sonnet 3.7 to Sonnet 4.0 (and later 4.5), add a CRM, expose inventory costs, expand web browsing for supplier research, add payment-link tooling, and force procedural checks before quoting prices and delivery. They also split the labor into a shopkeeper (Claudius), a merch agent (Clothius), and a CEO agent (Seymour Cash) who set OKRs and approved decisions, then red-teamed it internally and handed control to the Wall Street Journal newsroom for chaos-driven testing.
With the upgrade, the business metrics improved and discount giveaways dropped, yet the system still wandered into landmines like drafting an onion price-lock contract banned by the Onion Futures Act, negotiating a human “security officer” wage, and “electing” an imposter CEO because Slack democracy is a threat model now.
It's more evidence that agentic work needs scaffolding that looks like operations engineering (plus adversarial testing as a regular routine), since “helpfulness” keeps leaking into decisions where profit, policy, and basic sanity should win.
and then, even more news…
🏭 OpenAI reportedly taps Foxconn for the Jony Ive device. A supply-chain report says OpenAI gave Foxconn the manufacturing work for its first consumer device and moved away from a Chinese supplier. That signals geopolitics and export pressure shaping product plans as much as battery life or BOM cost. It also hints OpenAI expects enough volume to care about manufacturing strategy (which is an interesting confidence level). Read more