Back of Envelope

They removed the default model setting. The whole thing is a router now and the reason is compute cost, everything that looks cheap gets shunted to Instant no matter what you pick. There is no setting to fix it because there is no setting any more.

Take your own observation seriously: you said you cannot tell the answers apart on half your prompts. That is not a failure to perceive quality, that is a measurement, and it says the cheap tier is the correct choice for that half.

The useful version of this question is not "which model is better", it is "on which of my prompts can I detect a difference blind". Everything you cannot detect a difference on should run on whatever is fastest. That is a decision you can make today without waiting for anyone's leaderboard.

Treating the filter as a once-a-season purchase is the standard mistake. Mine was audibly struggling by December and I blamed the motor.

The boring answer is that you keep your own ledger and stop reading the provider dashboard for anything except reconciliation.

Every model call returns a usage object. Write one row per call before you return the response:

request_id, user_id, feature, model,
input_tokens, cached_input_tokens, output_tokens,
attempt_no, ts

Cost is then a SQL query with a price table you own, not a dashboard. Two useful things fall out immediately: cost per user, and cost per feature. When I did this the first query showed 4% of accounts were 61% of spend, all of them on one code path.

Second thing: attempt_no in that table is the column that would have caught this on day one. Alert on p99 attempts, not on dollars. Dollars are a lagging indicator of a loop that is already broken.

Also, if you are re-sending the whole file every attempt, put the static part of the prompt first and use prompt caching. Cached input tokens are dramatically cheaper than fresh ones, and a retry loop is the single best-shaped workload for it because everything except the last message is identical.

We tried exactly this and it got worse, not better. The small model failed the tool schema more often, so it looped more often, so total tokens went up even though price per token went down. Fix the loop first, then price-shop. Otherwise you're just buying a cheaper way to run forever.

Export to a plain PDF from a single-column document, no text boxes, no tables, no header or footer, and you have addressed ninety percent of what any of these tools can legitimately flag. That took me four minutes and I have never opened one since.

The band everyone quotes is 300 to 500 miles, which is roughly 500 to 800 km, and my own logs sit near the middle of that for firm daily trainers. 620 km puts you inside the window where it's completely reasonable for them to be done. On cost, that band works out at something like 15 to 25 pence per kilometre for a £120 shoe, which is cheap by the standards of the rest of the hobby. If your legs are complaining and you're past 600, this isn't you inventing an excuse.

Mostly. Radicals are worth a couple of hours as a separate exercise so you can tell visually similar characters apart, and then never again in isolation.

The 29 inch convention is about 73.7 cm and it's a writing-desk height inherited from an era of paper, and it broadly suits people somewhere around 5'10" and up. Typing surfaces want to be lower than writing surfaces, which is why the whole population under about 5'7" ends up shrugging. Two numbers I'd hold onto: keyboard at seated elbow height, and at least 60 cm of depth so a 24 inch monitor can sit roughly an arm's length away: 75 to 80 cm of depth is noticeably more comfortable if the room allows it. Anything shallower and you'll end up leaning back to focus, which puts the strain in your neck instead of your shoulders.

First time I reverse seared I got a grey chop with no crust at all, because I took it straight from a humid oven into the pan with a wet surface and it steamed. Now I dry brine, salt them and leave them uncovered on a rack in the fridge for at least 12 hours, and the difference in the crust is not subtle. If you only change one thing, change that.

That is what I do with anything questionable. Nothing bad happens from using it, you just cannot count on the number on the label, so put it where the stakes are low.

Worth adding that the protection is only as good as the trader still being in business in eight months, which on very small independents is not guaranteed.

Right first step. Measure before you rebuild. Half the problems people solve in portfolios turn out to be two positions.

Get bloods before you decide either way. I took a multi for six years and it turned out my ferritin had been on the floor the entire time, because the iron in a general multivitamin is a token amount. A test told me one specific thing to fix. The multi was quietly masking the question I should have been asking.

Setting aside who said it, I think the claim is wrong as stated and it does damage in rooms like the one you were training. Simplification isn't a solo act, it depends on what your listener already has to build on. I can explain what a Fourier transform is for in two sentences to a musician, because octaves and overtones give me the scaffolding, and I cannot do it in twenty minutes for someone with no reference points at all. That isn't me not understanding it. The quote gets used to shame specialists into producing analogies that are simple and quietly false, which is worse than an honest 'this one takes an hour'.

Track it or you will never know. Date, company, channel, tailored yes or no, outcome. Ten minutes a week and after two months you have your own answer to this question rather than mine.

1.5 hours × 5 days × 50 weeks = 375 hours a year, so 750 over two. Worth noting half of those hours should be listening if the goal is following a fast table, table conversation is a listening problem long before it's a speaking one.

Re-embedding on every save is the whole $402 and it's the easiest thing on your list to fix.

Chunk first, hash each chunk, store the hash next to the vector. On save, re-chunk, and only send chunks whose hash changed. Someone editing paragraph 4 of an 11-paragraph note re-embeds one chunk, not eleven. Our re-embed volume dropped 94% and the quality is identical because the unchanged text produces the identical vector anyway.

Second, debounce. There is no reason to embed on keystroke-save. Queue it, coalesce, run 30 seconds after the last edit. Combined with hashing that took us from ~180k embed calls a day to about 6k.

Third, check whether your embedding model supports shortened output dimensions. If it does, going from 1536 to 768 halves your index size and your index build time with a small retrieval quality hit that is usually invisible for a notes product. Measure it on your own data before you commit, obviously.

Same experience. The useful session was almost entirely about how I sit, where the bench was, and how much my shoulders were doing. Nobody had ever looked at that.

Politely disagreeing with taking wheel truing off the list entirely. A wheel that is a couple of millimetres out is a fifteen minute job with a spoke key and a zip tie on the frame as a gauge, and it costs nothing to try. Building a wheel from scratch or rescuing one with a serious buckle, absolutely, that is a shop job. But the small correction is one of the highest value things a home mechanic can learn, because it stops a small problem becoming a shop visit.

Spend the GBP 50 on a password manager subscription and a hardware key instead. Both of those measurably reduce the odds of the thing that actually happens to normal people, which is credential reuse and phishing, and neither requires you to trust a company with all of your traffic.

Ask one boring question before deciding: does the TA line come with a tuition waiver and health cover that the internship does not? People compare gross pay and forget that the assistantship is often quietly worth another 15 to 20 thousand in things you would otherwise pay for. If the internship is cash only and you lose enrolment status for the summer, the numbers can flip completely.

"Renting an IP address in another country" is the clearest one-line description of the product I have seen, and it survives contact with every marketing page.

It works for concepts and it fails for procedures, and I suspect your six weeks contains both. If the exam asks you to explain why something happens, then write and check. If it asks you to actually do something: integrate by parts, balance an equation, work through a proof - no amount of explaining will get you there, because that's a motor skill dressed as knowledge. For those, closed-book problems under time pressure is the only thing that transfers, and the explanation technique will fool you badly because you can describe the method perfectly and still not execute it.

In taxable, a dividend is a realisation you did not choose. In sheltered, it is a non event. Decide which account the money lives in first and most of the argument answers itself.