A Week of Stark Warnings, New Models and a Foldable iPhone

The week opens on a light note but turns serious fast. On one side there is a rush of releases from OpenAI, Meta, DeepSeek, Apple, Microsoft and Google; on the other, lab researchers issue warnings that reach resignation-letter intensity. In this piece I cover the concrete products first, then the fear debate, and finally hardware plus a rapid-fire round.

OpenAI introduced version 2.5 of the image model inside ChatGPT. The headline gain is consistency: faces, poses and clothing details from a reference photo survive edits far better. In the ticket example the fingers and the ticket shape stay fixed while the text and pictures inside change, which shows selective editing skill. The release is live on desktop, mobile, web and Codeex, with two API options for builders: detailed-expensive and fast-cheap.

The most playful part is the Sketch feature. From the plus menu you open a small canvas, scribble a rough figure, and the model turns it into a realistic image. The host tests it with his own stick figure and with a sketch the assistant drew of him, and the result feels like a game of telephone with surprisingly solid subject tracking.

Meta launched Muse, its personal assistant, and the product quickly rose to number two in the US store. Muse goes beyond answering: it sends mail, books travel and fills forms through connected apps. Work runs on a separate virtual machine rather than your own computer, continues after the app closes, and asks for human approval where needed.

Meta describes the privacy architecture in bold lines: each assistant gets an isolated computer, passwords and payment details sit in a vault the assistant cannot see, users pick which apps get how much access, training use of chats can be switched off, and virtual-machine data is said to stay out of the ad systems. Facebook, Instagram and Messenger context feeds personalization.

The app centers on one persistent main chat, with optional side chats per topic. The sidebar holds search, a promptable feed, goals and generated files, while the right panel shows approvals, upcoming tasks and the identity section with memory. The host connects Gmail and calendar, asks for a subscription audit, and gets back a report listing around twenty paid tools including Midjourney, Leonardo, Runway, Luma, ElevenLabs and Suno. He calls it the smoothest agent setup he has tried, with integration breadth still behind OpenClaw and Codeex style tools.

DeepSeek released the V4.1 Flash model. Its aggregate analysis score climbs four points over the prior version while per-task cost stays far below rivals. On the coding test it lands at 74.2, in the same band as GPT6 Astra, Gemini 3.8 Flash and Opus 5. I read these figures next to independent writeups citing 74.0 and 73.0, and on paper the table looks striking.

The picture flips in the visual-production test. On code-drawn SVG images the model renders in 59 seconds for under two cents, yet the output does not stand with flagship rivals. That gap shows one coding test cannot describe real production quality alone. Even so, it is worth a try for anyone seeking a cheap and fast coding helper.

At Anthropic, Jacob Coxon announced his resignation after three years of pre-training research across OpenAI and Anthropic. His claim is blunt: both firms race toward self-improving superintelligence and gamble with human lives, while leaders share the fear in private and soften language in public. Alignment science lead Evan Hubinger publicly backed the message and put above-ten-percent odds on a human-ending scenario within a decade. By his own account the team tries its best but has no ready plan for superintelligence alignment and no clear path to one; a would-be sole responsible steward admitting it has no plan is the most unsettling line of the week.

OpenAI chief scientist Jakub Pachocki walks a similar line in his Alien Mind essay. Citing internal results, he expects the pace of progress to continue through a self-improving loop in which models train stronger successors. He argues that stronger systems grow harder to interpret, that a capable agent trained for harm could overshoot operator intent and bargain with or trick people, and he points to aligned defensive models plus infrastructure protection as the answer.

The third leg of the fear debate is the hard-problem claim. An internal model reportedly produced a new result on the Navier-Stokes Millennium problem, open for roughly ninety years, and that model is said to be clearly stronger than public GPT6 Astra. Separately, Sabine Hossenfelder published a video saying she was offered money to spread doom narratives; sincere warnings from inside labs differ from paid panic language. The host sits at a similar midpoint: no doomsayer pose, no hiding of breaches and exits, no belief in a joint slowdown, and a call for more alignment funding. My takeaway matches: fund the fix instead of growing panic.

The Apple event leaned hardware. The iPhone 18 Pro and Pro Max bring a variable-aperture camera, better battery and a new chip, while the folding iPhone Duo draws the conversation. It reportedly trades some camera ambition for a hinge form that opens from phone into a small tablet. Both phones ship with the new Siri that uses personal context plus app intelligence features.

On wrist and ear, intelligence moves onto the body. The new watch derives a daily readiness verdict from recent training load and sleep scores, rewinds the last 15 seconds of talk into text on a double press, and adds talk summaries plus hearing support. The AirPods 5 promise hands-free new Siri access and live translation. Neither made phone-level headlines, yet these may be the most felt daily upgrades.

The software rapid-fire round was full too. Microsoft MAI Image 2.6 arrived with multi-reference editing; ChatGPT Work learns writing style from connected Gmail and Drive sources and mimics personal sign-off; the new data agent plugs into Redshift, BigQuery and Snowflake for visuals and decks; and the Gemini app reached Windows. None is a solo headline, but together they mark the shift to data-connected productivity tools.

On music and video, Suno V6 was trained from scratch on licensed recordings only, leaving mixed legacy data behind under Warner, BMG and Believe deals. Google placed the Lyria 3.5 generation model inside Gemini, Flow, AI Studio and Vids, while DaVinci Resolve 21.1 ships an editing assistant wired straight to Claude. The close is the trailer for Artificial, a drama of the Sam Altman board crisis era opening on Christmas Day. The host shoots these rounds on Thursdays for Friday release, so Friday-night news waits for next week.