barbegal

This simulates the chip but do I still have to write the simulation of my actual board which often has equally complex ICs on it and mixed analog and digital?

Without a human to benchmark against it's really tough to gauge how good these models are vs how good the task definitions and existing codebases are.

My intuition from the example full instructions are that the tasks are poorly specified which results in ~60% failures due to bad assumptions and missing requirements.

18 O enriched water is available to buy but the pricing suggests it's quite expensive to separate 16 O from 18 O. And you get much less variability in the ratio of isotopes from mixing compared to separating a single isotope as much as possible.

Interesting but I can't deal with the AI generated prose. Sentences like "This was not a garment for walking between a carriage and a doorway. It was ..." show me no human has proof read or edited the text.

Yeah I couldn't get past so many issues in the AI generated illustrations. Not useful at all when they are completely wrong.

Yes I would imagine lots of those type of services would be vulnerable if they hadn't updated to the latest kernel versions.

Yes you can pin it to core 1 whilst pinning all other tasks to core 0. Then will never be interrupted or preempted (except by interrupts created on core 1)

I don't understand why people get so hung up on Chrome using so much memory. A lot of this memory is "discardable" so will get dropped when the system is under memory pressure and the amount of memory allocated for this type of usage will depend on how much memory your system has available. If Chrome is using lots of memory then it's almost always because your system has lots of available memory. It allows the browser to cache large images and video assets that would otherwise have to be re-downloaded over the internet.

Does the KV cache really grow to use more memory than the model weights? The reduction in overall RAM relies on the KV cache being a substantial proportion of the memory usage but with very large models I can't see how that holds true.

Total receipts were over 11,000 so more like 100 hours or around $2000 so a similar price to the LLM.

This is good work. I wish branch predictor were better reverse engineered so CPU simulation could be improved. It would be much better to be able to accurately predict how software will work on other processors in software simulation rather than having to go out and buy hardware to test on (which is the way we still have to do things in 2026)

The memory overhead is fairly significant it uses between 1.5 and 3 times the space of the data stored.