Founded Open Geophysical (now Shearwater GeoServices) writing seismic software and once again writing geo software. Also worked on radar & sonar stuff and HFT.
Boulder, CO
Joined April 2009
- Tweets1.2K
- Following897
- Followers165
- Likes74K
One gripe with opus 5.5 is that it seems excessively focused on making fp math deterministic across platforms, and I have to actively work against this. E.g. it always wants to disallow FMA, not use transcendental functions, etc. Quite strange.
One interesting thing now that Claude, etc. enable is writing very specific linear solvers for specific problems.
In my case, I am using IPOPT to optimize something. A generic linear solver like MUMPS is fine, but if your problem is pretty fixed you can very easily get a much faster solver by baking in the elimination order based on what you know about your matrix.
Now that I have distilled my CasADi IPOPT trajectory planner into a small NN my pro peloton simulator runs 50 riders worth of trajectory planning per core with room to spare.
I don't understand why the built in Claude Code Agent tool doesn't support changing the effort. Obviously there are workarounds, but it seems like a simple and useful thing to support directly.
Somehow in college I never was interested in EE (except for the FPGA class I took from Bill Dally), but 20y later working through Razavi Microelectronics I really enjoy it