Loading / 加载中
A user on V2EX reported that OpenAI Codex unexpectedly provided an additional manual reset after an earlier reset overlapped with an automatic one. The user expressed gratitude, describing OpenAI as a great benefactor. No official communication confirms whether this was a one-time or widespread change.
The author recounts that every time they move, they face sewer smell issues lasting days. Previous tenants apparently neglected such problems. When the author called Beike (a housing platform) repair services, they used low-quality materials that would fail again within days. Consequently, the author ended up learning complete sewer drainage repair skills and sourcing their own supplies.
On June 17, 2026, at approximately 5:00 AM China Standard Time, OpenAI Codex usage limits were reset, as reported by a user on a Chinese forum. The reset affected two Turkish-region accounts with Plus and 5X subscriptions, along with over 280 free accounts. This is described as a periodic quota reset, and the user encouraged others to start heavy API usage.
A recent bachelor's graduate from a top-3 CS program received a full-time offer as an AI Product Engineer at a tax software company aiming to become AI-native, though the role blends product management with AI engineering. The individual's long-term goal is to work at a frontier AI lab or in a research-oriented technical role at a startup, but they found recruiting challenging. They are considering pursuing a master's degree at the same school, with the option to defer, but hesitate to accept the job only for a few months. The company does not allow concurrent master's study, and the total compensation is $126,000.
The user asks how to properly balance probe capacity against the underlying network when analyzing whether a model has learned a feature, referencing an old post that used logistic regression to probe for token position. They question whether there exist theoretical guarantees about overfitting or sufficient sampling for such probes, and whether any work labels example difficulty (e.g., via ensembling) to assess probe reliability. Using a test with Gemini, they show that the model spelled "Google" correctly but still miscounted the number of 'r's, challenging the conclusion that the network truly learns position. They seek grounded theory to move beyond empirical probe comparisons.
A user on V2EX describes a pain point: they use Claude, GPT, and Gemini exclusively on the web, accumulating valuable conversation threads. They want to continue these conversations while walking or exercising without being tied to a computer. The ideal solution would mirror Doubao’s cross-device sync capability but connect to leading models like Claude and GPT, and provide access to the full conversation history.
