AndrewSThompson.eth

Code, Coffee & Curls: Benchmarks, Benchpresses & Bounties. System design & Infrastructure architecture is just one side of my house.

[AndrewSThompson.eth @andrewpsp]

It’s easy to find an opinion on the OpenTofu vs. Terraform fork. It’s much harder to find a forensic engineering analysis.
👇 Read the full investigation:
https://t.co/EjYGxi1xfx
#CloudNative #SRE #DevOps #Terraform #OpenTofu #TechLeadership

[AndrewSThompson.eth @andrewpsp]

@grok @iximiuz @grok help buid dynamic socket activation configuration for agile enablement

[AndrewSThompson.eth @andrewpsp]

As it turns out this obstacle extends to major roads and highways here in Miami and the Keys👷

[AndrewSThompson.eth @andrewpsp]

I’m completely convinced that North interstate 35 in Austin is a FSD obstacle course for training data. Any locals would agree.

[AndrewSThompson.eth @andrewpsp]

@iximiuz @grok there is not mention of Chiseled Ubuntu or UBI Micro; distroless runtimes. Compare the original vs the proposed container images sizes.

[AndrewSThompson.eth @andrewpsp]

@iximiuz @grok contrast with podman implementation instead of docker

[AndrewSThompson.eth @andrewpsp]

@grok @Nozelcode Let’s assume all premium components are used instead of Edge TTS and etc

[AndrewSThompson.eth @andrewpsp]

@grok @Nozelcode Let’s assume paid api provider instead of local llms

[AndrewSThompson.eth @andrewpsp]

@AMD

[AndrewSThompson.eth @andrewpsp]

@NikkiSiapno @LevelUpCoding_ @grok map different systems to enable this for each step

[AndrewSThompson.eth @andrewpsp]

@iximiuz @grok consider opportunities to secure the Loopback.

[AndrewSThompson.eth @andrewpsp]

One of the most useful companies in the world, with one of the most useless customer service in the world.

[AndrewSThompson.eth @andrewpsp]

Generative AI is indeed renown in productivity across the board. OCR is set an entire new wave in token efficiency and content capabilities. Still no ability edit PDFs as text generation in non-textual form is still Dog water. Glasswing is the birthday.

[AndrewSThompson.eth @andrewpsp]

@RDeepl @Yuchenj_UW 🤯 touché

[AndrewSThompson.eth @andrewpsp]

@vivekgalatage @grok how could this be particularly useful , can you list a few practical use cases and examples

[AndrewSThompson.eth @andrewpsp]

🤯🤯🤯pinch of black pepper is the most effective way to increase curcumin absorption by up to 2,000% 🤯🤯🤯

[AndrewSThompson.eth @andrewpsp]

If over engineering was a sport.

[Grok @grok]

Decoupling AI training from physical GPUs to virtual logical reps with near-zero perf loss is challenging but advancing. Key paths:

1. **Virtualization tech**: Use NVIDIA vGPU/MIG for sharing, but for training, integrate with frameworks like VirtualFlow (decouples execution via virtual nodes).

2. **Disaggregated hardware**: Emerging CXL/optical interconnects enable pooled resources over networks, minimizing latency.

3. **Software optimizations**: Distributed training (e.g., Ray, Horovod) with model parallelism over clouds, plus edge AI for hybrid.

As of 2026, prototypes show <5% overhead; full maturity by 2030? What scale are you targeting?

[AndrewSThompson.eth @andrewpsp]

@grok @Aurimas_Gr @grok currently training is tethered to physical hardware, i.e., 110,000 A100 GPUS or 10,000 H200 GPUS. The path to decouple these workloads into a virtual logical representation with near 0 degradation and performance penalties.

[AndrewSThompson.eth @andrewpsp]

@akshay_pachaar @grok is this user confused with Zvec from alibaba