GPT-Live keeps talking while Astra tackles 10 math problems — and Codex fixes Windows

The short read

Scope: This edition covers original posts published from Aug 3, 18:00 through Aug 4, 18:00 Asia/Shanghai. The X account connector is not linked, so the source pool is the configured public AI and tech accounts rather than a personal following list.

AI tools and developer ecosystem

GPT-Live keeps the conversation moving while deeper work happens elsewhere

• What happened: OpenAI says GPT-Live can listen while it speaks, with a full-duplex voice model that handles deeper reasoning and tool use on an asynchronous path. 1
• Why it matters: The important change is architectural, not just a faster voice model: a slow tool call should no longer stall the audio path, so conversation and computation do not have to take turns. 2
• Concrete detail: OpenAI says its new system's p95 frame-delivery behavior matches the previous system's p50, and a protocol change cuts media startup from six network round trips to one. Those are engineering claims from the company, not an independent benchmark. 2

Codex is useful when the problem is boring Windows glue

• What happened: Ethan Mollick says he used Codex to fix driver problems, game incompatibilities, and startup-program failures across his Windows machines. 3
• Why it matters: These are exactly the jobs that are annoying to investigate but have enough observable symptoms for an agent to try fixes across several systems. The value is less "write code" than "remove a long tail of small technical chores." 3
• Limit: Mollick writes "Hours saved" but gives no before-and-after log, exact bug count, or time total. Treat this as a practitioner report, not a measured productivity study. 3

Research

OpenAI reports ten results on long-standing math and computer-science problems

• What happened: OpenAI says an internal version of Astra produced ten results across areas including sphere packing, coding theory, group theory, quantum complexity, lattice cryptography, and extremal combinatorics. 45
• Why it matters: The post is a stronger research signal than a benchmark score because the company names the problems and publishes manuscripts, reasoning walkthroughs, and Lean certificates for inspection. Whether the arguments change their fields still depends on mathematicians checking and extending them. 5
• Key number: OpenAI says finding the solutions used roughly $2,000 worth of tokens at GPT-5.6 Sol API rates; the manuscripts were then prepared by humans and the arguments formalized in Lean. 45

Business and enterprise

Mollick's question: why did Microsoft and Google ship more boldly at the start?

• What happened: Ethan Mollick argues that Microsoft and Google took larger early risks with AI: he points to Microsoft's GPT-4 rollout, its decision not to retreat after Sydney, and the quick arrival of Copilot; for Google, he cites early Deep Research and the rapid Bard pivot. 6
• Why it matters: His point is about organizational appetite, not model quality. Rough first products can create distribution and user feedback; once a market becomes legible, the fear of shipping a visibly imperfect product may become the larger constraint. 6
• Limit: This is a short retrospective thesis with no dates or linked evidence in the post. Read it as a strategy question worth testing against product histories, not as a verified chronology. 6