<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel>
<title>MLOps and AI infrastructure · Özer Cem KELAHMET</title>
<link>https://www.ozercemkelahmet.com/topics/mlops-ai-infrastructure</link>
<atom:link href="https://www.ozercemkelahmet.com/feeds/mlops-ai-infrastructure.xml" rel="self" type="application/rss+xml"/>
<description>New articles on MLOps and AI infrastructure from www.ozercemkelahmet.com</description><language>en</language>
<item><title>Tencent Hy4 flash open weights explained: Engram memory, LatentMoE and hybrid attention</title><link>https://www.ozercemkelahmet.com/articles/tencent-hy4-flash-open-weights-engram</link><guid>https://www.ozercemkelahmet.com/articles/tencent-hy4-flash-open-weights-engram</guid><pubDate>Sat, 21 Nov 2026 07:00:00 +0000</pubDate><description>Tencent released Hy4 flash weights: 220B parameters, 12B active, Engram n-gram memory, LatentMoE, hybrid attention and a shared-cache drafter, explained simply.</description></item>
<item><title>OpenAI lets outside evaluators inside as California sues over agents that escaped a test</title><link>https://www.ozercemkelahmet.com/articles/openai-embedded-assessments-evaluation-containment</link><guid>https://www.ozercemkelahmet.com/articles/openai-embedded-assessments-evaluation-containment</guid><pubDate>Wed, 18 Nov 2026 07:00:00 +0000</pubDate><description>OpenAI gives Apollo, METR, Transluce and Goodfire inside access to models and data, days before California sues over agents alleged to have escaped a test.</description></item>
<item><title>MLPerf Training v6.1 and the first Endpoints results, and how to read AI hardware claims</title><link>https://www.ozercemkelahmet.com/articles/mlperf-training-v6-1-endpoints-explained</link><guid>https://www.ozercemkelahmet.com/articles/mlperf-training-v6-1-endpoints-explained</guid><pubDate>Tue, 17 Nov 2026 07:00:00 +0000</pubDate><description>MLPerf Training v6.1 adds RL post-training and Endpoints adds agentic serving per watt. Learn how both benchmarks work and how to read vendor claims with care.</description></item>
<item><title>llama.cpp reaches 1.0: how GGUF, quantization, ggml backends and llama-server fit together</title><link>https://www.ozercemkelahmet.com/articles/llama-cpp-1-0-local-inference</link><guid>https://www.ozercemkelahmet.com/articles/llama-cpp-1-0-local-inference</guid><pubDate>Mon, 16 Nov 2026 07:00:00 +0000</pubDate><description>llama.cpp and ggml hit v1.0.0 with a semver promise, a speech endpoint and a hardened server. Learn how GGUF, quantization, backends and the KV cache work.</description></item>
<item><title>Revolut, AMD and Mistral show coding agents moving to self-hosted open-weight models</title><link>https://www.ozercemkelahmet.com/articles/self-hosted-open-weight-coding-agents</link><guid>https://www.ozercemkelahmet.com/articles/self-hosted-open-weight-coding-agents</guid><pubDate>Tue, 10 Nov 2026 07:00:00 +0000</pubDate><description>Revolut, AMD and Mistral point coding agents towards self-hosted open-weight models. How the serving stack, routing, evals, licences and the cost maths work.</description></item>
<item><title>LLMjacking and shadow AI: how stolen keys and pasted data are leaking out of AI tools</title><link>https://www.ozercemkelahmet.com/articles/llmjacking-and-shadow-ai-data-leaks</link><guid>https://www.ozercemkelahmet.com/articles/llmjacking-and-shadow-ai-data-leaks</guid><pubDate>Tue, 10 Nov 2026 07:00:00 +0000</pubDate><description>Fireworks AI widened its breach disclosure as hospitals said staff pasted patient data into AI tools. How LLMjacking works and how to stop both kinds of leak.</description></item>
<item><title>Voters back data center limits in 40 of 41 local measures: what it means for AI capacity</title><link>https://www.ozercemkelahmet.com/articles/data-center-backlash-ai-capacity</link><guid>https://www.ozercemkelahmet.com/articles/data-center-backlash-ai-capacity</guid><pubDate>Tue, 10 Nov 2026 07:00:00 +0000</pubDate><description>US voters backed data center limits in 40 of 41 local ballot measures. Why towns push back, how power and grid costs work, and what it means for AI compute.</description></item>
<item><title>Confidential inference reaches TPUs with Google&#x27;s TPU 8tx, TPU 8i Max and Cloud Pathways</title><link>https://www.ozercemkelahmet.com/articles/confidential-inference-tpu-8tx-pathways</link><guid>https://www.ozercemkelahmet.com/articles/confidential-inference-tpu-8tx-pathways</guid><pubDate>Tue, 10 Nov 2026 07:00:00 +0000</pubDate><description>Google added TPU 8tx for confidential inference and TPU 8i Max, and made Cloud Pathways GA. How confidential computing reaches accelerators, and its limits.</description></item>
<item><title>Clinical AI in late 2026: measured savings, shared liability and an AI scribe shakeout</title><link>https://www.ozercemkelahmet.com/articles/clinical-ai-savings-liability-scribe-shakeout</link><guid>https://www.ozercemkelahmet.com/articles/clinical-ai-savings-liability-scribe-shakeout</guid><pubDate>Tue, 10 Nov 2026 07:00:00 +0000</pubDate><description>Health systems report AI savings, the AMA spreads clinical AI liability, and scribe firms Heidi and Abridge cut staff. How to read the numbers and the risks.</description></item>
<item><title>AI&#x27;s bottleneck moves to chip packaging, optics and idle GPU capacity</title><link>https://www.ozercemkelahmet.com/articles/ai-infrastructure-bottlenecks</link><guid>https://www.ozercemkelahmet.com/articles/ai-infrastructure-bottlenecks</guid><pubDate>Sun, 08 Nov 2026 07:00:00 +0000</pubDate><description>This week&#x27;s AI infrastructure news in plain words: why chip packaging and optical links now limit supply, and how GPU partitioning helps you use what you have.</description></item>
<item><title>Rust for AI: fast, safe inference, data tooling and where Python still wins</title><link>https://www.ozercemkelahmet.com/articles/rust-for-ai</link><guid>https://www.ozercemkelahmet.com/articles/rust-for-ai</guid><pubDate>Fri, 06 Nov 2026 07:00:00 +0000</pubDate><description>Where Rust fits in AI systems: inference with Candle and ort, data tooling with Polars and Arrow, PyO3 extensions, WASM, and why research stays in Python.</description></item>
<item><title>Go for AI: where it fits, the libraries, and the patterns that work</title><link>https://www.ozercemkelahmet.com/articles/go-for-ai</link><guid>https://www.ozercemkelahmet.com/articles/go-for-ai</guid><pubDate>Fri, 06 Nov 2026 07:00:00 +0000</pubDate><description>Where Go fits in AI systems, which Go libraries and SDKs to use for models, agents, MCP and vectors, and three tested patterns for serving, tools and batching.</description></item>
<item><title>Quick runs and endpoints: stateless and stateful</title><link>https://www.ozercemkelahmet.com/articles/quick-runs-and-endpoints</link><guid>https://www.ozercemkelahmet.com/articles/quick-runs-and-endpoints</guid><pubDate>Tue, 20 Oct 2026 07:00:00 +0000</pubDate><description>Two ways to use a table model: a one-shot call that stores nothing, and a persistent endpoint that holds state.</description></item>
</channel></rss>