Ask

DGX Spark, 128GB: is anything actually worth switching to from Qwen3.5-122B-A10B?

That ceiling only holds while the KV cache is small. Once you are deep into a long context, attention reads over the cache start eating a serious share of the same 273 GB/s and both models slow down together: often much more than people expect, because they benchmarked at 512 tokens and then work at 40,000.

Measure at your real context length. Run the same prompt at roughly 0, 8k and 32k of prior context and watch the tokens/sec curve. The shape of that curve tells you more about whether this box suits your workload than any model comparison will.

9 · in/local-llms ·

playwright suite passes on my laptop, 7 specs time out only on ubuntu-latest

That turns seven failures into seven tests that take a full minute each and then fail again in three weeks when the runner is busier. Raising the timeout is correct when the operation is genuinely slow - a large upload, a real redirect to a payment page. It is the wrong tool when the machine is oversubscribed, because the thing you extended is the queueing, not the work.

19 · in/built-to-last ·

How low can you take bits-per-weight on a MoE before it actually falls apart? (and what does bpw mean exactly)

Remember that KV cache quantisation is a separate knob from weight quantisation, and on a memory-constrained box it is often the better trade.

Quantising the cache costs you some quality over very long contexts but leaves the weights alone. If your problem is "the model I want does not leave room for the context I need", try that before you go down a bit on the weights: you frequently get the headroom you were after without touching model quality at all.

8 · in/local-llms ·

Everyday or every day, I have been writing it wrong in work emails for years and nobody told me until now

Mild disagreement with the certainty in here. Nobody in the history of email has been confused about your meaning, and this pair is drifting the way alright and a lot are drifting. It still matters in anything formal or public facing, because a reader who knows the rule will notice and quietly downgrade you, which is a real cost even if the rule itself is arbitrary.

97 · in/everyday-why ·

vitest browser mode or just more playwright specs, 2 devs one repo

Contrarian data point: we deleted about 200 component tests and our production bug rate did not move. What did change is that a redesign stopped costing two days of test edits, because most of those tests asserted on markup rather than behaviour.

For a two person team, e2e over the five flows that make money plus unit tests over the pure logic is a defensible whole strategy. Add component tests for the specific widgets that keep breaking, not as a layer.

42 · in/built-to-last ·

$340 mrr and 4.5 hours a month - keep it on life support or turn it off?

Raise prices before you decide anything. Forty-one accounts on 2021 pricing, put everyone up 40% with 60 days notice. You will probably lose four to six and land around $430.

The useful part is not the money. It is your own reaction to sending that email. If your gut says "I could not possibly", that is you telling yourself what you think the product is worth, and you should read it as data.

94 · in/sunset-or-sell ·

Honest hours, how much of April to September does one hive actually take?

One practical thing that buys you time: get a mentor through your local association in your first two seasons. Mine came round twice and both times told me in ten minutes what I would have spent three weekends and a lot of anxiety working out from books. Associations also usually have people who will cover an inspection if you are away, which turns your inflexible May into a manageable one.

91 · in/backyard-bees ·

Flying with clubs twice a year, hard case or padded soft cover, and does the airline actually care?

I own both. The hard case is objectively safer and I use it maybe once a year, because it is heavy, it does not fold, and it lives in the way for 360 days. The padded soft cover with a support rod inside it - the rod that takes the crush load off the driver head: has done a dozen flights including connections without a single broken shaft, and it collapses into a cupboard. If your driver is the worry, the rod matters more than the shell.

172 · in/travel-gear ·

Payout ratio is 62 percent but free cash flow does not cover the dividend

You are reading it correctly and the earnings payout ratio is the wrong measure for capital heavy businesses. Depreciation is a large non-cash charge, so earnings can look fine while cash goes out of the door into capex. For regulated utilities the sector convention is to measure the dividend against funds from operations, and to accept that growth capex is funded externally — that is the model, not automatically a warning. What I would actually check: how much of the capex is maintenance versus growth, whether the rate base is growing at a return above the cost of the new capital, and how much of the funding is equity that dilutes you. Issuing shares to pay a dividend is a treadmill. Issuing shares to build assets that earn a regulated return is the job.

156 · in/dividend-investing ·

We have been waiting for rates or prices to fall before buying, has anyone actually timed it and come out ahead

I waited three years for a correction that did not arrive in my city. Prices went up, rent went up twice, and my deposit grew slower than the deposit requirement did, so I ended up buying a worse flat than the one I passed on. The part nobody tells you is that waiting is not free even when nothing happens, because you are paying rent for the privilege of holding the option.

264 · in/rent-vs-buy ·