Gemini 4 Argon, Muse Wonder, and Jensen Huang in Taipei: 7 videos from October 12–18

The short version

Google just went FRONTIER (Gemini 4 Argon)

• Argon went on sale on October 13 at an introductory $2 per million input tokens and $10 per million output tokens, with cached input at 95% off. The standard $4 and $20 apply from November 30. 1
• Google pitches Argon as a model for "deep work". It ranks first on the Vals Index, which scores models on finance, coding, legal and tax tasks, and Google ran its hardest internal engineering jobs through it. 1
• The builds are the argument. A bare prompt produced a neon cyberpunk city at 60 frames a second in the browser, and Berman says Argon settles on about 30 minutes of work when a prompt names no budget. 2
• His Fall Guys clone took several rounds of prompting, ended up with ten levels, and plays. A clone of Excel, produced in a single 20-minute run, calculates in the browser and breaks only on a typed SUM formula. 2
• On the chart he assembled, Argon scores 77.9 on DeepSWE, just under Muse Wonder Full's 79, and leads Vibe Code Bench at 91.9%. It trails Opus 5.5 on Terminal-Bench, 57 against 66.4, and he puts it at $2 per completed task against $4.52 for Muse Wonder Full and $7.63 for the Claude models. 2

Applying Gemini 4 for large-scale codebase maintenance in Chromium

• Chromium is twenty years old and about 40 million lines. Every feature on top of the core engine was added by a different team, so changes ripple into places nobody expects. 3
• Chen's measure of the change is throughput. In two weeks with early access he landed more changes than he would typically land in months. 3
• Argon holds 20 changes in flight at once. Each one runs a couple of hundred thousand tests and two human reviewers, then a couple of million tests across new and old devices. 3
• A typical engineer lands one change a day, so 20 in flight is 20 days of work. Chen runs the model without a harness: a memory file of notes on the existing tooling and a plain instruction to untangle the code. 3
• Google's post describes the target as a roughly 6 million line folder of legacy browser features, cut loose from core page handling. Months of refactoring finished in weeks, and the user payoff is pages moving between tabs, side panels and floating windows without reloading. 1

Teaching a humanoid robot to play cornhole with Gemini 4 Argon

• Google's researchers showed Argon one video of a human throwing a beanbag. The model wrote the MuJoCo simulation and the whole-body control needed for Apptronik's Apollo humanoid, and Apollo sank seven consecutive shots in the real world. 1
• The clip narrates the run in the first person. It watches the arm swing and release, maps that human motion onto the robot's joints, then measures the board with two eye cameras and models both bodies in physics. 4
• It practiced in simulation and widened its stance so its hand cleared its hip. On the real robot it moved the release point a few centimetres at a time until the bags dropped. 4
• Google calls this Video-to-Sim-to-Real and claims it compresses weeks of hand-built simulation modelling into a few hours, which turns ordinary video into robot training data. 4
• The supporting number is Argon's 87% on ERQA, a benchmark of physical and spatial reasoning in 3D environments. Google reports it as the highest among frontier models. 1

Meta just dropped something insane... (Muse Wonder)

• Meta released two models on October 12. Muse Wonder is available through the Meta Model API and Muse Code, and Muse Wonder Full only to Muse Code subscribers. 5
• Both carry Active Memory, which keeps context across conversations, and Active Focus, which keeps a model working through blockers on a long task. 6
• On the chart Berman reads, Muse Wonder Full sits second to Opus 5.5 on the Artificial Analysis Intelligence Index, 56 against 58. It leads DeepSWE at 79, ahead of Argon, while Grok 4.7's 19.6 on the Harvey legal benchmark outranks Meta's charted 13.1. 7
• The price he reports is $4 per million input tokens, $20 per million output and $0.20 per million cached, which lands it beside Opus 5.5. Muse Wonder Full costs him $4.52 per completed task, against 72 cents for GPT-6.1 Soul Max. 7
• Access was the disappointment. His $50 plan hit its quota in three minutes on repeated 429 errors, Meta's AI head Alexander Wang replied that the problem was known and being fixed, and the $200 plan ran out too. 7

Visiting Taiwan w/ Jensen Huang

• Fridman's framing: Taiwan is two-thirds mountains and makes more than 90% of the world's most advanced semiconductors, and every AI data center depends on it. 8
• Huang's Computex pitch is thermal and mechanical. The new racks are 100% liquid-cooled, against 80% for Grace Blackwell, and assembly dropped from two hours to five minutes. 8
• He rates the Vera CPU at 1.8 times x86. His economics rest on one line: "AI doesn't rent cores," it generates tokens, so the goal is producing tokens as fast as possible and selling them. 8
• The origin story is the best part. NVIDIA went looking for a foundry in 1996 and TSMC, then a $300 million company, offered terms with no strings attached. Huang wrote to Morris Chang, who called back and invited him to Taiwan; the partnership is thirty years old. 8
• From the flywheel he argues that export controls should be narrow and selective. In his numbers, about a thousand NVIDIA teams are local to Taiwan, a new building will hold 4,000 people, and he plans to hire 3,000 more. 8

Scalable Training of Diffusion Samplers via Fixed-point Score Matching

• The problem: you can evaluate a probability density, but you cannot draw from it or compute its normalising constant. The densities that matter — molecular, material, posterior — are multimodal, high-dimensional and expensive, with one DFT energy evaluation taking hours. 9
• Classical MCMC mixes slowly across modes and must be rerun for each new temperature. Diffusion samplers before this work were unstable and needed many density evaluations; Blessing notes that the benchmark literature up to 2025 concluded the methods did not work. 9
• Fixed-point score matching turns training into supervised regression. Generate samples from the current model, store them with the target forces, regress the model's score onto the target score, and repeat until the model's own field is a fixed point. 9
• For amorphous materials his model Atlas matched reference energies, structures and radial distribution functions. It reached them with 10^5 target energy evaluations where parallel tempering needed 10^8, and trained on 64 particles it stayed accurate at 512. 9
• He closes on what the method still cannot do. The score is local, so distant modes can be weighted wrongly; free-energy estimates from path weights are cheap and high-variance; and uniqueness and convergence of the fixed point remain unproven. 9

KITE: a Kinesthetic Interface for Tangible Exploration

• KITE sits on a monitor and runs in two modes at once. It slides freely like a mouse until it meets a virtual object, then pushes back with matching force and torque. 10
• The mechanism is three wheels in a circular chassis, which let the puck move in any direction without turning, driven by gearless motors with a custom driver. 10
• A camera pointed up at asymmetric LEDs on the ceiling localises the puck against the screen, and motor feedback fills the gaps between camera frames so contacts feel stiff. 10
• The demonstrations are physics lessons: lift on an airfoil, hydrogen bonds pulled apart to feel electronegativity, tidal forces near and far from the Moon, springs in series and parallel, and Coulomb against viscous friction. 10
• Microsoft's page credits Elizabeth Childs, a PhD candidate at Stanford University. 11