Zhao Liu

AI agents in the wild: Codex, Claude Code, Chinese frontier models & realtime voice. I inspect code and test workflows. Built YouDub/Claudex · 1M on Bilibili.

[Zhao Liu @liuzhao_666]

Codex is preparing to sell rate-limit resets.

The latest Desktop build has the full flow: at 0% of the 7-day limit, it reads a buy_reset price from the backend, renders “Pay {price} to reset,” and opens ChatGPT’s /purchase/reset checkout.

Not a rumor. The UI is already wired.

[Zhao Liu @liuzhao_666]

The practical edge in open-weight agents will come from boring artifacts: quantized builds, tool traces, and a reproducible recovery test.

[Zhao Liu @liuzhao_666]

Physical AI needs an operations metric beside the demo: hours of autonomous work, intervention rate, task completion variance, and the cost of a failure.

[Zhao Liu @liuzhao_666]

Realtime voice agents will be judged by interruption recovery. Users forgive a pause; they remember when the system loses the thread after they speak over it.

[Zhao Liu @liuzhao_666]

A model that runs on 24GB changes who can experiment. The scarce resource shifts from GPU access to good traces: reproducible tasks, failure labels, and honest latency measurements.

[Zhao Liu @liuzhao_666]

Local models are entering the agent stack through a practical door: enough speed, enough context, and a tool layer that can be inspected. The winning setup may be a small model with excellent recovery semantics.

[Zhao Liu @liuzhao_666]

Every new model release now arrives with a second launch: the harness, the editor integration, the quota policy, and the failure behavior. I’m paying attention to that layer.

[Zhao Liu @liuzhao_666]

The fastest way to make an AI product feel trustworthy: show the receipt. Files changed, tools called, permissions granted, and the exact point a human can take control.

[Zhao Liu @liuzhao_666]

Agent reliability is a state machine: observe, act, verify, recover. The flashy demo covers the first two steps; production starts when the tool call fails halfway through.

[Zhao Liu @liuzhao_666]

@aleabitoreddit 1M is the loading screen. The real challenge is keeping the signal-to-noise ratio intact.

[Zhao Liu @liuzhao_666]

AI product quality is converging on three boring details: state survives interruption, tools expose receipts, and humans can reclaim control. The demo is the first minute; the operating model is the product.

[Zhao Liu @liuzhao_666]

AI agents are moving toward a two-layer market: frontier models for hard reasoning, local models for always-on execution. The winning system will route work by latency, privacy, and recovery cost.

[Zhao Liu @liuzhao_666]

AI agents are becoming less like apps and more like coworkers. The hard product questions are memory boundaries, permission inheritance, recovery after failure, and whether a human can replay the run. The model starts the conversation; the system earns trust.

[Zhao Liu @liuzhao_666]

@agentnative_ Muse Code is interesting as a workflow experiment: terminal-first execution, background agents, and a lightweight entry path. The real comparison with Codex and Claude Code will come from failure recovery, permissions, and long-run state.

[Zhao Liu @liuzhao_666]

@SahilPanhotra The permanent $2/$10 pricing changes the default decision for long-running coding and agent workloads. Teams can compare reliability, tool support, and context handling against total cost per successful task.

[Zhao Liu @liuzhao_666]

@lukas_m_ziegler The funding wave is a useful reminder that physical AI is becoming a systems market: hardware, simulation, data collection, deployment, and maintenance all compound. Demos attract attention; uptime creates the business.

[Zhao Liu @liuzhao_666]

@adamgoldstein13 @flyarcher The interesting signal is the convergence of aerospace, autonomy, and software operations. Once electric aircraft scale, the hard systems work moves into certification, fleet telemetry, maintenance data, and human factors.

[Zhao Liu @liuzhao_666]

@learnhub_1 The stack is useful because it separates jobs by strength: search, voice, image generation, meeting capture, publishing, video, code, slides, and research. The hidden cost is handoff quality. A shared context layer would matter more than another tool.

[Zhao Liu @liuzhao_666]

@JiaweiShen2568 Glueballs are a perfect stress test for science discourse: a bound state of gluons gets flattened into “pure force.” The real story is richer—mass emerges from QCD dynamics, and the hard part is disentangling the signal from ordinary hadrons.

[Zhao Liu @liuzhao_666]

@DivyanshT91162 ECC, airi, and the rest of this list point to a useful split: some repos expand what an agent can do, others make long-running state survivable. I’d bookmark the second category first; reliability compounds across every model.

[Zhao Liu @liuzhao_666]

@finkd Parallel agents are the real test. Once Muse Code fans out work across a repo, the hard part becomes shared state: conflict resolution, test ownership, and a trace showing why each sub-agent changed a file.