EchoFox

“Until death, all defeat is psychological."

karpathy pulling up to the office for his first day on the research team

Damn

This is like one of those audio journals you find on a corpse in a sci-fi horror game

sorry bro i’ll make the bios interface using react next time

@kayleecodez hate to say it, but everyone that rejects kubernetes inevitably ends up recreating it from first principles lol

We finally know why LLMs hallucinate. It's not the model. It's the geometry. @OpenAI text-embedding-3-large: 91/3072 dimensions do real work. @GeminiApp gemini-embedding-001: 80/3072 dimensions do real work. ~97% of your vector database is mathematically empty. Your RAG system is retrieving from noise. @ashwingop and I present "The Geometry of Consolidation" - a proof that RAG compression has a hard floor no algorithm can beat, set by a single spectral number your embedding model cannot escape. Every hallucination your RAG pipeline produces? This is why. Paper + results: github.com/niashwin/geome…

adulting is basically arriving at the same truths as your father, but from first principles.

Kill Them With Your Success And Bury Them With Your Smile Vijay took it seriously 🫡

to be clear, NVIDIA is NOT a car

@sriniksv Trying to msg you on twitter regarding a suggestion and can't seem to do it. Could you msg me?

@livingdevops What do u think is a good course or resource for MLOps? I have been looking for them online.

Everyone tells you companies use Kubernetes for MLOps. Nobody shows you how. MLOps on Kubernetes follows the same pattern as everything else. You solve one problem, and then the next. You have a model that works on your laptop. > You need to train it on real data at scale. >Your laptop has 16GB RAM, and the dataset is 200GB. >Training locally is not an option. So you run training as a Kubernetes Job. > A Job spins up a pod, runs training to completion, and terminates. > You get GPU nodes for training and release them when done. > You are not paying for idle GPU capacity. But training one model takes hours. > You need to run 50 experiments with different hyperparameters. > Running them one by one means waiting days for results. So you run parallel Jobs. > 50 pods are training simultaneously. > Each has different parameters. > Results come back in hours, not days. But now you have 50 trained models and no idea which one performed best. > You have no record of what parameters produced what result. > Next week, nobody remembers what worked. So you add experiment tracking. > MLflow running on Kubernetes. Every training job automatically logs parameters, metrics, and artifacts. > You always know which model came from which experiment. But your best model is sitting in an S3 bucket doing nothing. > It needs to serve predictions to your application. Spinning up a Flask app manually on an EC2 machine is neither repeatable nor scalable. So you deploy the model as a Kubernetes Deployment behind a Service. > Your model server runs as a container. > It scales with HPA when prediction requests increase. > It restarts automatically when it crashes. But your model gets stale. > Real-world data drifts from training data over time. > Predictions start degrading, and nobody notices until users complain. So you add monitoring. > Your model server emits prediction metrics to Prometheus. > Grafana dashboards show prediction distribution over time. > Data drift triggers an alert before accuracy degrades in production. But fixing drift means retraining. > Retraining manually means someone has to remember how to do it all. > Pull fresh data, run the job, evaluate the model, and deploy it. > That is four steps where humans make mistakes. So you build a pipeline. > Kubeflow Pipelines or Argo Workflows on Kubernetes. > Fresh data arrives, retraining triggers automatically. > New model evaluated against the old one, better model promoted to production, bad model gets discarded automatically. > Nobody touches it manually. That full loop is what MLOps on Kubernetes actually means.

LOL!!

Ummm…pretty sure you can tho…

@ConsciousRide Any resource?

As an AI engineer. Please learn: - Python (deeply - it is still king in 2026) - Core ML/DL (transformers, attention, backprop, optimization, loss functions) - Frameworks (PyTorch 2.x / JAX - pick one deeply; understand both eventually) - Model architectures (LLMs, diffusion, multimodal, MoE basics) - Fine-tuning & PEFT (LoRA/QLoRA, adapters, full fine-tune trade-offs) - Data pipelines (cleaning, augmentation, tokenization, dataloaders, streaming) - Evaluation (benchmarks, perplexity, BLEU/ROUGE/BERTScore, human eval, RAGAS) - Serving & inference (vLLM, TGI, TorchServe, ONNX, TensorRT, quantization) - Prompt engineering + RAG + agents + tool calling patterns - MLOps (tracking experiments, versioning models/data, monitoring drift)

someone's getting started early!

They're doing drone strikes with a KANBAN board and story points

Nireekshana (1982)