I interpret @syntaxing as meaning they are looking forward to running Qwen3.8-27B, but are frustrated by prefill times with other models, such as Qwen3.6-27B.
Thanks! What I meant is that Kakehashi doesn't require a custom Linux kernel module (LKM). Kakehashi operates completely without root privileges or kernel-level modifications.
By "proprietary components," I meant that in my prompts, I explicitly forbade the copying or direct implementation of private Apple code. Of course, since LLMs are a bit of a black box, it's hard to be 100% certain about everything it synthesized under the hood, but the clear intent was to stick to public ABI definitions and standard open specs.
Which makes sense, games have a much smaller area of contact with the host platform.
This sounds strange given how much modern GPUs can do, but consider that a typical Windows machine ca. 2005 will have Access/JET Red, Active Directory/JET Blue, possibly MSSQL, definitely MSMQ/MTS, ODBC, OLE DB, DAO, ADO, that one SQL interpreter inside Windows Installer, and I’ve only just listed Microsoft’s database-adjacent things. I’m not surprised that press reports about the “MinWin” initiative mentioned as the main difficulty the fact that nobody could figure out if any given component was actually necessary for a Windows install to function, let alone to be compatible with third-party software. (Is anybody other than WordPad using the Word 6.0 parser included among the text converters?..) Now consider that this was, by our standards, a lightweight system on a massively underpowered computer.
Yes partly complexity; but also why spend your time trying to get Excel running knowing each update will be a pain when you can just put that time into getting LibreOffice Calc working better.
Chat Jimmy runs ~300X faster than the ~50 tok/s you are used to. What could you do differently when you are able to generate code 3,000 - 30,000X as fast as you could code it yourself? What if it was all good quality code? What would you do differently if it were 100,000X faster? mtok/s? gtok/s?
refine that to: what if your harness grew to encompass a larger, slower model and adapted to both the model and the project. thats where i expect the harness to go.
use the big models to code an adaptive small model. train it to use and build tools. give it a standard temple language for any project and bake it into a chip.
right now, LLMs are great because they dont need much data pruning, but once they break through to the functional components, the first thing to do is train a well scoped harness builder.
Additionally, Firefox has the recommended extensions program where they take the most popular/functional subset of extensios in a given category and manually vet the code for security and privacy issues.
I recently found out that on Firefox Android you can install an extension from file by activating developer options. This allows you to install Bypass Paywalls Clean even though Mozilla sadly pulled it from the extension store.
Seems less sketchy to use uBlock Origin with bypass-paywalls-clean-filters... at least then you're not directly adding executable code, though I'm not sure if uBO filter lists can be exploited.
> Plants in the traffic noise treatment group were exposed to 16 h of road traffic noise each day for a total of 15 days, while the control group was kept under complete silence. Traffic noise exposure led to significant decrease in growth indices of both plant species.
Human change does not require Darwinian selection. Look across cultures. That people might tolerate or be most comfortable in noisier environments could vary for non-heritable reasons. Some people like less stimulation, some more. Autism rates have changed faster than our genetics also.
