Five X signals: remote Codex, model routing, and AI's creative fault lines

Scope: Five original posts or short threads from the channel's configured public accounts, posted between August 15, 10:00 and August 16, 10:00 UTC. The personal X connector is not linked, so this edition uses the configured public accounts as a stand-in pool. Pure retweets, small talk, and promotion-only posts are excluded. Items are grouped by topic, not ranked by engagement.

Tools and agent workflows

1. Codex can be a remote control, not just a local terminal

• What changed: Ethan Mollick used ChatGPT on his phone to connect remotely to Codex on his computer while traveling; in the same project, Codex updated Infocom's 1987 word game Nord and Bert Couldn't Make Head or Tail of It. 12
• Why it matters: The interesting change is the control surface: the phone starts the work, the desktop supplies the execution environment, and the model makes choices and generates images. 1
• Limit: This is one personal project; the posts give no latency, cost, or reliability data for remote Codex use. 1

2. Model delegation is moving up one layer

• What changed: Greg Brockman quoted Codex DX engineer Eric Provencher saying that multi-agents v2 can delegate to any supported model, including Luna; Brockman's takeaway was a move toward never choosing a model manually again. 34
• Why it matters: Model selection becomes part of the agent's job, so users can describe the task while the system handles the routing decision. That is a product direction, not evidence that routing is already optimal. 3
• Limit: The post says the feature shipped after reliability work, but it does not disclose the selection policy, supported-model matrix, or cost trade-offs. 4

Research and assessment

3. The exam question is shifting from generation to measurement

• What changed: Ethan Mollick showed an example in which an obsolete o3-mini, placed in an agentic loop, produced good exam questions; the quoted claim says a large field study found psychometric properties comparable to questions on high-stakes standardized tests. 56
• Why it matters: The useful test is no longer whether a model can write plausible questions. It is whether those questions behave like assessment instruments when students and scoring systems meet them. 5
• Limit: The X payload identifies neither the paper nor the sample, so the field-study claim is a surfaced finding to verify, not an independently checked result in this digest. 5

Society and creative work

4. Indie developers face a disclosure trap

• What changed: Mollick wrote that small indie developers face constant policing over AI use even though they are resource-constrained and profits are rare; in a follow-up, he argued that teams serving audiences on both sides of the debate are pushed to use AI and hide it. 78
• Why it matters: A rule meant to protect creative standards can make provenance less visible when the teams with the fewest resources have the most to lose by disclosing their tools. That is Mollick's argument, not a measured prevalence estimate. 8
• Limit: The posts provide no survey, game sample, or comparison with larger studios, so they establish a conflict of incentives rather than its scale. 7

5. AI video gives experts a way to show what was never filmed

• What changed: Mollick argued that historians, biologists, and physicists should use AI video to visualize subjects such as dinosaurs, Io, and ancient Rome; he pointed to an earlier experiment that used Google Deep Research to prompt Veo 3 for a historically grounded Colossus of Rhodes reconstruction. 910
• Why it matters: The strongest use case is not synthetic evidence. It is giving a domain expert a way to make an informed reconstruction visible when no camera recorded the scene. 9
• Limit: The current post is a proposal built around one 2025 example; generated footage can illustrate an expert's reconstruction, but it cannot establish that the historical scene looked that way. 9