Overview
Today feels like a study in acceleration. DeepSeek’s new models set off a fresh round of score-watching and debate, Apple drops an open-source image-and-video generator, and Gemini 3 shows off hands-on coding tricks for 3D on the web. SpaceX keeps stacking orbital wins, while posts about inflation and governance frame a mood that is both ambitious and wary.
The big picture
SpaceX deploys 29 Starlink satellites, nails another droneship landing
SpaceX confirmed 29 Starlink V2 Minis reached low-Earth orbit from the Group 6-86 mission, with Falcon 9 lifting off from Kennedy at 2:44 a.m. EST. The booster, on its fourth flight, landed on Just Read the Instructions about 8.5 minutes later, marking the company’s 541st booster recovery. It is the 111th Starlink mission of 2025, bringing the year’s total deployed to 2,871, supporting over 8 million subscribers worldwide.
DeepSeek V3.2 claims top marks on reasoning benchmarks
Victor Mustar flags DeepSeek-V3.2, a 685B-parameter model that he says tops key 2025 benchmarks like IMO, IOI and ICPC, beating GPT-5 in select reasoning tasks. The model touts 128k contexts with DeepSeek Sparse Attention at lower cost and faster speed, plus heavy RL post-training focused on agent tool-use across thousands of synthetic environments.
Under the hood of V3.2 and V3.2-Speciale
Tanishq Mathew Abraham highlights the tech notes: V3.2 targets GPT-5-level performance on math-heavy sets like AIME and HMMT using sparse attention. The V3.2-Speciale variant leans into competition-grade reasoning, with gold-level scores across IMO, CMO, ICPC and IOI 2025. GRPO tweaks such as unbiased KL estimates and off-policy sequence masking aim to boost generalisation without external tools.
Scoreboards and sceptics as the whale resurfaces
Ashutosh Shrivastava cheers DeepSeek’s returns with tables claiming wins over GPT-5 High, Claude 4.5 Sonnet and Gemini 3.0 Pro on math and coding. Chubby posts the same energy, calling V3.2 a daily driver and V3.2-Speciale a maxed-out reasoner, while replies fire back with accusations of distillation from Western models, echoing earlier IP rows.
A pointed US holiday jab tied to China’s AI sprint
Ex-xAI engineer @yacineMTB posts a sarcastic holiday greeting while quoting DeepSeek’s Speciale scores. Replies lean into the cost angle, noting reports that DeepSeek runs cheaply and uses non-NVIDIA Huawei chips, a reminder that the race is about price and supply chains as much as raw performance.
Apple puts STARFlow on Hugging Face
DailyPapers notes Apple’s STARFlow release, a 3B autoregressive flow model for 256x256 text-to-image with a deep-shallow Transformer design and competitive FID against diffusion peers. There is also STARFlow-V for up to 30 seconds of 480p text-to-video, using causal attention and WAN2.2-VAE latents. It is Apple’s first notable open-source generative release tied to a NeurIPS paper, and it revives normalising flows for exact likelihood training and faster sampling.
Gemini 3 writes interactive 3D scenes you can control with your hands
el.cine shows Gemini 3 generating a complete Three.js particle scene that responds to webcam hand gestures, with a simple prompt-to-HTML workflow in Google AI Studio. It is a neat demo of no-code web creation, from particle shapes to colours, and adds to the growing set of examples where large models scaffold real-time 3D in minutes.
DHH on Tailscale making multi-network life simple
David Heinemeier Hansson praises how easy it is to switch between nicknamed tailnets, backed by WireGuard. The screenshot shows listing and swapping accounts with plain commands. Replies trade notes on use cases from personal backups to industrial controls, with a few calls for open-source alternatives like Netbird.
Millionaire inflation meme lands with a thud
Seamus Coughlin posts a Mr Incredible meme aimed at a UBS report that 24 million US adults now have a net worth above $1 million. Replies argue the milestone feels smaller in an inflationary world, and point out much of that wealth sits in homes and other illiquid assets, skewed towards older cohorts.
David Sacks on who is allowed to govern
Sacks argues the media frames business experience as a conflict, which leaves government to career bureaucrats and activists who share the press’s outlook. He cites data on limited private-sector experience among recent economic appointees, tapping into a broader push for operators in the next administration.
Why it matters
The model race is moving on two fronts at once. DeepSeek’s claims push open-source and low-cost competition, which could change who can build agents at scale, while also stirring fresh IP disputes. Apple’s STARFlow suggests big players see room beyond diffusion, with flows promising faster sampling and exact likelihoods that could matter for training stability and controllability. Gemini 3’s 3D demo points to a near future where people prototype games or visual tools by typing, not wiring up frameworks.
SpaceX’s drumbeat of launches keeps internet infrastructure expanding. At the same time, posts on millionaires and governance hint at a mood of scepticism about headline stats and who holds power. The through line is the same: scale and access. Whether satellites, models, networks or public roles, the questions are who gets them, who pays, and who sets the rules.
