This is helpful. Someone should create a similar benchmark for the BEAM. This is also a good reminder to continue working on snakepit [1] and snakebridge [2]. Plenty remains before they're suitable for prime time.
In many ways, we're past the point of no return. So-called ubiquitous technical surveillance is largely the norm, often encroaching by design beyond the boundaries of expected decency.
Informational terrorism, a dysphemism that describes the manner by which certain data is abused to "re-rank content" for a "personalized experience," is encoded into the DNA of certain large tech companies.
The ideal would have been a security-first (privacy-first) industry and supply chain. The ideal never was going to happen, anymore than the early educational ideals of the television industry.
Ergo we are not past the point of no return. That point never existed. We are right where we should expect to be, with most people victimised by the industry and the supply chain, and with a small percentage of people working in security/privacy education to mitigate unsafe practices.
Seatbelts and airbags exist. Smoking is banned in many public settings. It took a senseless amount of carnage to achieve these measures.
We just haven't achieved the requisite amount of privacy carnage. Yet.
Yes. The only question left is when does the terror begin? And it will--it will be our own governments clamping down on all of us. The digital norm globally will be China under the CCP. That is the future for all of us unless we turn it off, but we won't because humans are stupid.
This is largely the attitude that led to this in the first place. This is about failures of messaging, campaigning, and organising. It is a lack of democratic engagement that directly stems from the idea of individual choice being supreme over everything.
Incredibly helpful. This article helped me to identify significant fundamental shortcomings in my first attempt to design an agent/tool arbitration protocol:
I'm working on an ML platform in Elixir. Orchestrating CPU-intensive Python processes is central to the project, so I created a session based pooler called `snakepit` [1].
The latest revision is building a batteries-included gRPC Python bridge that enables streaming, bi-directional tool use, as well as an innovative variables feature for experimental ML inspired by ideas from DSPy's team.
One of the later project goals: A Python client that can manage the Elixir orchestrator that manages pools of Python, in a distributed environment. Maybe I'll call that submodule `snakepits`. In this embodiment, it will be an effective albeit much more sophisticated replacement for `asyncio` for some use cases.
https://hex.pm/packages/vllm