Isn't this solving a subproblem of the overall issue of uncompressed tool call polluting context?

Furthermore, this could be compressed even further with a dynamic legend of every UUID in the context. So UUID@Bravo and UUID@Delta would be the actual symbols in the context but dynamically replaced when calling tools.

How much does your data privacy cost?

As stated in the analysis, thousands of dollars. That said, the smart thing to do is target smaller models (few billion parameters) and then use larger models for non-privacy tasks.

Regarding AVX: could rust be compiled with different symbols that target different x64 instruction sets, then at runtime choose the symbol set that is the more performant for that architecture?

I'm not sure how that works. You either let the compiler compile your whole program with AVX (which duplicates the binary) or you manually use AVX with runtime detection on selected places (which requires writing manual vectorization).

> First Revision - Completed with minimal human contribution (auto-accept mode) > > Second Revision - Planned with increased human-in-the-loop involvement

At least 'minimal human contribution' in the first revision. So scan read it as it revised the book.

I do find it amusing to think that people might just ask an AI to summarize the book instead of actually reading it. Maybe even 'review' it as well.

The future will have authenticity in short supply

> Compilers are not yet rewriting data structures and algorithms based on microarchitecture.

Afaik mojo allows for this with the autotuning capability and metaprogramming

What data type/precision? How much parallelism versus iterative? Is this fully in fabric or part of a SoC? Easy to get mired in many details.

HW design has so many degrees of freedom which makes it fun but challenging.

What specific topics are you trying to address with this?

Showing the tradeoffs between complexity & throughput of HLS and HDL implementations is the goal. I assume, uint8 should be sufficient to make a point?

Newton Raphson could be used to calculate a reciprocal. (in a bit width larger than 8). If starting from a good reciprocal approximation, convergence to bit accurate reciprocal should not take many iterations. Then multiply the reciprocal with the numerator to perform the divide

I worked on an automotive FMCW LiDAR that didn't quite make it to market. Cool technology but it was difficult to scale the cost down, which is pretty important for automotive. Margins are very low in that market

Does anyone have a good summary of the Buck(2) track?