Long runs, native stacks, and the credit crunch

00:00:04 Ethan Mollick handed GPT-5.6 Sol full computer control over Codex and tasked it with winning the daily challenge for Slay the Spire 2 — randomized factors included, so it couldn't cheat off a saved run. It held its ground for five hours, making complex moves whose payoff came dozens of turns later, while managing resources carefully.

00:00:28 Then it won. The model cleared forty-eight floors at Ascension 3. Because Slay the Spire 2 launched after the model's training cut-off, every decision had to be reasoned from scratch without leaning on preloaded game knowledge. The win is headline value, but the five uninterrupted hours of play—no hallucinations, no task-drift—is what actually crossed a line.

00:00:53 Jan Stevens put it plainly in her reply: "Five hours of not just playing, but not drifting off-task, feels like the real line crossed here." The boss fight might have been the easy part. Yasu0x.hl tracked the implication in the thread: holding for five hours instead of brute-forcing through a loop means the model is learning game theory inside a closed system.

00:01:19 That's where we move past prompt-solving and into actual task completion.

00:01:25 A similar insistence on holding course over long horizons is what Ant Group's LingBot-VA 2.0 was built to solve, and Elvis — Omar Sarmiento, formerly of Google DeepMind — broke down the architecture today. Most video-action robot models are content-focused video generators with an action head bolted on later.

00:01:48 The mismatch is fundamental: video generators use bidirectional attention over all frames, while control runs strictly forward in time. Retrofitting one into the other forgets the pretraining priors you paid for. LingBot-VA 2.0 trains the entire stack for control from day one.

00:02:08 It maps world states and latent actions into a single semantic space — a visual-action tokenizer aligned to a frozen vision foundation model. The system learns latent actions self-supervised from unlabeled web video, so the signal doesn't depend on scarce teleoperated demonstrations.

00:02:30 The capacity design runs on a sparse mixture-of-experts video stream with 128 experts—keeping eight active per token, plus one shared. That totals about 13 billion parameters, with just 1.9 billion firing per token. The action stream stays dense for low-latency work.

00:02:50 They got real-time performance by distilling those experts down to two steps per chunk, then running FP8 TensorRT with a paged KV-cache and runtime cleanup. That cut response time from 927 milliseconds to 142 per chunk, pushing peak asynchronous execution to 225 hertz.

00:03:10 The architecture uses what the team calls foresight reasoning: it drafts the next action chunk while the current one executes, then re-grounds every prediction on fresh observations. The loop predicts, corrects, and anchors on new data without letting the look-ahead drift.

00:03:30 It averages 93.6 on RoboTwin 2.0, leaving only a 0.6 gap between clean and randomized test conditions. It adapts from 10 to 15 demonstrations and transfers across different robot bodies. When the architecture's inductive bias matches the problem's directionality, you don't have to fight the pretraining later.

00:03:53 That architectural honesty matters when a system has to run for hours without drifting off-course.

00:04:01 The pressure doesn't stop at model parameters. It shows up sharply in the physical infrastructure scaling to support them. Microsoft, Amazon, and Google emitted 119 million metric tonnes of carbon dioxide equivalent last fiscal year — roughly a third of France's annual output.

00:04:20 That represents nearly a 20 percent jump from the previous year, when the three companies together came in around 101 million tonnes, roughly equivalent to Czechia's emissions. Microsoft led with a 25 percent increase, followed by Google at 18 percent and Amazon at 16 percent.

00:04:39 JLL expects about 1,200 new datacenters globally between now and 2030, with demand overwhelmingly driven by AI. The Uptime Institute estimates that last year's announced projects alone would consume 1.3 percent of the world's electricity usage. The companies still claim net-zero targets — Google and Microsoft aiming for 2030, Amazon for 2040 — but the infrastructure buildout makes those goals harder to hit with every quarterly report.

00:05:10 The sharper tension shows up in Microsoft's sustainability report, which points to shrinking availability of carbon credits on global markets. "While companies are actively investing in or purchasing carbon credits, the figure suggests a possible lack of credit supply in the carbon market to meet the technology companies' needs," noted UCR's Shaolei Ren.

00:05:35 Everyone is focused on physical shortages — power, land, water for cooling — but there may be an equal shortage of virtual goods. The offsetting narrative relies on markets that can absorb tens of millions of additional tonnes per year. If those markets don't have the supply to match demand, the gap between reported and actual emissions widens in ways that never show up on a press release.

00:06:02 Cecilia Rikap at UCL put it bluntly: claims about ecologically friendly clouds are just marketing strategy when governments are asked to let these same companies use AI solutions to address the ecological crisis.