Yajur

@kimmonismus Typing in /compact brings back the sliding widget with the progress ring

Codex Desktop no longer shows visible context/token usage indicator? Bug or did they delete it?

@harshmadhusudan Figure in oil shocks. It breaks these neat convergence argument.

If people who are not supposed to be experts say something like this, that is fine. It is not their core area. But for people who are supposedly professional investors, this is an astonishing level of ignorance. I do not mean to pick on this gentleman as there are many more such folks. Point is to teach and to learn. India's PPP conversion rate is 20, quibbles aside. The multiplier to get to 96 is 4.8. Theoretically, if India and US had same incomes the multiplier would be around 1. The reduction of the multiplier as a country grows is called the Balassa Samuelson effect. Some call it the Penn effect. I would also like to acknowledge Jagdish Bhagwati's work in this context. Now, PPP conversion rate drifts upwards by relative inflation difference and some adjustments. While India is likely to bring down its 4% target inflation before 2047, let us say it remains 4% and US at 2% (though they have been exceeding it.) That takes 2047 PPP rate to 30 from 20. Now a 240 USDINR then would mean a 8x PPP/MER. Such a number is flat out impossible for an economy-pair like US-India, even less so that by 2047 India will have 35-40% of US incomes adjusted for cost of living and population (like say a Chile today) at only 5% faster growth (RLCU.) Currently India has 13-14% of US incomes. Now what is an ideal multiple for 35-40% of US income. It depends on global cycle, country scale, deficits etc. But broadly it should be in the 1.5-2.5x region. Which means the rupee should be in the 45-75 region as per this analysis. The midpoint is 60, which is 4 times stronger than the quoted analysis has suggested. So unless one believes that India will have double digit inflation, US low inflation, India low growth, US high growth and all four together - the quoted number is almost impossible to materialise. I understand this is boring, step-by-step analysis but this is the only way to understand this phenomenon. Hope it helps.

@asmah2107 Or maybe frontier LLMs distill better into lower dim manifolds retaining similar (reasoning) abilities.

Prediction : By 2026 end, the entire industry might switch. We’ll likely stop spending money on frontier models for plumbing jobs. Instead we’ll use cheap, fast, task-specific models for the boring stuff and save the expensive frontier models only for the hard reasoning parts. Result? Agent costs drop dramatically (5-10× cheaper in many cases), agents feel snappier, and small teams can actually compete.

@thsottiaux Are we allowed to code own harness without violating ToS?

A little secret. About 5% of our production traffic is on the Pi harness, about another 5% is on OpenCode. Reminder you can use your ChatGPT account in a flourishing set of other tools. We’ll continue to make Codex awesome, but you have options.

@thsottiaux Vote yes. Very useful for bootstrapping complex experiments overnight.

Should we bring batch compute to codex? Aka /slow mode

@_anshulr Observed instances where the latest Gemeni Flash model's reasoning tokens often output incoherent reasoning tokens such as "Let's Go!" in an infinite loop. Also, reasoning tokens should be tighter for ex - "..I am hyperfocused on the result.." is a bit oddish and adds to cost.

3x limit increase forever

@_mohansolo Observed instances where the latest Gemeni Flash model's reasoning tokens often output incoherent reasoning tokens such as "Let's Go!" in an infinite loop. Also, reasoning tokens should be tighter for ex - "..I am hyperfocused on the result.." is a bit oddish and adds to cost.

An update: we’re 3xing the rate limits for Gemini models across all paid tiers in Antigravity and resetting everyone’s Gemini quota for the week. We understand some people hit their rate limits quickly and wanted to respond fast. Lots more to come and enjoy building!