Why do you hate datasheets so much?

"Not having access to ASML's lithography [wikipedia.org] puts any of these efforts at a pretty serious disadvantage."

From the summary: "The chip is designed for inference"

Inference is more constrained by memory latency and capacity than by compute. You don't need EUV tools for DRAM (yet) and CXMT and YMTC have capacity. You can get pretty far on a mature logic node if you can pile on the memory.

"They're 3x higher than Starlink's satellites so you can expect (a minimum of) 3x latency."

Even without knowing the specific values, you can reason that 3x orbital radius would mean a *maximum* of 3x latency because both a) the baseline from each ground station to the point under the satellite contacted by the base station* and b) the delay through the processing circuitry are not zero.

*: any satellite to satellite latency would increase with a higher orbit.

" It feels outright insulting to asked to read LLM output in a context where it is *supposed* to be human feedback."

More insulting or less insulting than having your human feedback rejected because someone claims it is LLM output?

What's this criterion does is provide non-falsifiable cover for rejecting anything.

"Supersonic flight is incredibly noisy"

A ban by Mach number permits noisy planes as long as they fly slow... which means that you are exposed to the noise for longer. (Which is why actual airport noise limits are time-weighted.)

The actual problem is the noise, so regulate the noise.

a) Supply curves slope upwards.
b) Datacenters are paying with Other People's Money; schools are paying with taxpayer money.

"the prediction is actually quite close to a situation that occurred 23 years ago"

If by "quite close" you mean significantly different, which you don't and why the snark was justified.

" is that a signal that we've about hit the peak of the memory demand / high-price cycle"

There's two ways the cycle peaks:
a) there's a bubble and it bursts. In this case this is just a money grab.
b) There's a secular trend and suppliers increase their growth rate to meet it. Growth in futures contracts becomes the justification for growth in capex.

The data supports the presence of a secular trend on top of the cycle, but memory suppliers keep getting burnt.

"I'd wager a guess that this information includes the accounts that the customer is using LastPass to store passwords for."

On what basis would you make that wager? Do you have information that was not released?

-1, economically illiterate

The increased prices for new items will increase demand for refurbished items. In the short term the supply of refurbished items is constant so the price goes up. In the long term, this creates an incentive to refurbish more marginal items that would require more parts and labor than they could have previously recovered.

If you refuse to pick winners, all you're left with are the losers.

The market can become rational faster than the individual investor can liquidate.

A day-over-day change of less than 5% in an index like the NASDAQ or S&P500 is not a big number. The median investor has no reason to be looking at day-over-day changes at all. If you can't stomach a single digit percentage drop, first stop looking at the day-over-day changes and second and then get out of stocks.

"Canonical's announcement doesn't use the term 'AI' anywhere in it."

STTs are not LLMs, but they are AI that use CNNs and RNNs.

The models Canonical mentions by name are Whisper, Parakeet, Nemotron, and Qwen3-ASR (https://github.com/canonical/myna/blob/main/docs/architecture/Myna%20-%20System%20Architecture.png).

"If you have a good keyboard, you can be far more efficient with it than with voice input."

-1, provably false
https://www.medrxiv.org/conten...