<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom"><channel>
<title>Generative AI and LLMs · Özer Cem KELAHMET</title>
<link>https://www.ozercemkelahmet.com/topics/generative-ai-llms</link>
<atom:link href="https://www.ozercemkelahmet.com/feeds/generative-ai-llms.xml" rel="self" type="application/rss+xml"/>
<description>New articles on Generative AI and LLMs from www.ozercemkelahmet.com</description><language>en</language>
<item><title>DaisyRose-27B and self-rewriting weights, two new ways to improve an open model after training</title><link>https://www.ozercemkelahmet.com/articles/daisyrose-rl-self-rewriting-weights</link><guid>https://www.ozercemkelahmet.com/articles/daisyrose-rl-self-rewriting-weights</guid><pubDate>Sat, 21 Nov 2026 07:00:00 +0000</pubDate><description>RadixArk post-trains a 27B model with asynchronous agentic RL, and a new paper lets a 27B model edit its own weights without gradients. How each one works.</description></item>
<item><title>SenseNova Loop-9B explained: how a looped transformer runs 32 layers as 56</title><link>https://www.ozercemkelahmet.com/articles/sensenova-loop-9b-looped-transformer</link><guid>https://www.ozercemkelahmet.com/articles/sensenova-loop-9b-looped-transformer</guid><pubDate>Fri, 20 Nov 2026 07:00:00 +0000</pubDate><description>SenseTime&#x27;s open-weight Loop-9B reuses 24 of its 32 layers twice per pass. Learn how looped transformers trade compute for parameters, and what it costs.</description></item>
<item><title>How to Read Microsoft&#x27;s MAI Code 2 Cost Claims and What Licensed Weights Mean</title><link>https://www.ozercemkelahmet.com/articles/mai-code-2-licensed-weights</link><guid>https://www.ozercemkelahmet.com/articles/mai-code-2-licensed-weights</guid><pubDate>Fri, 20 Nov 2026 07:00:00 +0000</pubDate><description>MAI Code 2 claims 98% of GPT-6.1 Sol at 40% of the cost. Learn what averaged benchmarks hide, how Flash routing works and what licensed weights mean.</description></item>
<item><title>Gemini 4 Neon explained: model tiers, thinking levels, cached input and orchestrator agents</title><link>https://www.ozercemkelahmet.com/articles/gemini-4-neon-thinking-levels-caching-subagents</link><guid>https://www.ozercemkelahmet.com/articles/gemini-4-neon-thinking-levels-caching-subagents</guid><pubDate>Fri, 20 Nov 2026 07:00:00 +0000</pubDate><description>Google&#x27;s Gemini 4 Neon sits between Carbon and Hydrogen. Learn how model tiers, thinking levels, 95% cached input and orchestrator subagents shape cost.</description></item>
<item><title>What RL on expert-built tasks teaches Nemotron 3.5 Super, one LoRA adapter at a time</title><link>https://www.ozercemkelahmet.com/articles/rl-expert-tasks-nemotron-lora-adapters</link><guid>https://www.ozercemkelahmet.com/articles/rl-expert-tasks-nemotron-lora-adapters</guid><pubDate>Thu, 19 Nov 2026 07:00:00 +0000</pubDate><description>A new paper trains five rank-32 LoRA adapters on Nemotron 3.5 Super with CISPO and no SFT. We explain the method, the 12 benchmark gains and the failure modes.</description></item>
<item><title>GPT-6.2 Astra and Sol: what changed, what it really costs and how to pick a tier</title><link>https://www.ozercemkelahmet.com/articles/gpt-6-2-astra-sol-pricing-routing</link><guid>https://www.ozercemkelahmet.com/articles/gpt-6-2-astra-sol-pricing-routing</guid><pubDate>Thu, 19 Nov 2026 07:00:00 +0000</pubDate><description>OpenAI&#x27;s GPT-6.2 Astra and Sol cut frontier prices and tokens per task. How to read the benchmarks, cached-input pricing, Ultrafast, and which tier to route to.</description></item>
<item><title>Context has a price: why trimming a coding agent&#x27;s prompt can double its bill</title><link>https://www.ozercemkelahmet.com/articles/context-price-coding-agents-prefix-caching</link><guid>https://www.ozercemkelahmet.com/articles/context-price-coding-agents-prefix-caching</guid><pubDate>Thu, 19 Nov 2026 07:00:00 +0000</pubDate><description>A 20,000-run study shows retrieved code and observation masking can raise coding agent cost under prefix caching. How the token bill works and how to measure.</description></item>
<item><title>RAG types explained: naive, hybrid, reranked, HyDE, corrective, agentic and GraphRAG</title><link>https://www.ozercemkelahmet.com/articles/rag-types-explained-naive-hybrid-rerank-hyde-corrective-agentic-graphrag</link><guid>https://www.ozercemkelahmet.com/articles/rag-types-explained-naive-hybrid-rerank-hyde-corrective-agentic-graphrag</guid><pubDate>Tue, 17 Nov 2026 07:00:00 +0000</pubDate><description>The main kinds of retrieval-augmented generation, what each fixes, what it costs and when to use it, with a live simulation that scores each on one corpus.</description></item>
<item><title>Memorilla packs a whole document store into 16 memory tokens for a frozen LLM</title><link>https://www.ozercemkelahmet.com/articles/memorilla-latent-memory-tokens</link><guid>https://www.ozercemkelahmet.com/articles/memorilla-latent-memory-tokens</guid><pubDate>Mon, 16 Nov 2026 07:00:00 +0000</pubDate><description>How Memorilla, an open memory module from a Stanford-led team, turns a document store into 16 soft tokens for a frozen Qwen3-8B, and how it compares with RAG.</description></item>
<item><title>The new LLM serving stack: cache-aware routing, continuous batching and expert pruning</title><link>https://www.ozercemkelahmet.com/articles/llm-inference-serving-stack-cache-routing</link><guid>https://www.ozercemkelahmet.com/articles/llm-inference-serving-stack-cache-routing</guid><pubDate>Tue, 10 Nov 2026 07:00:00 +0000</pubDate><description>Buoyant&#x27;s Kinference, Fireworks Forge and AWS research on MoE expert pruning show where LLM serving costs now go. How KV caches, routing and batching work.</description></item>
<item><title>Kimi K3.1 Pro and ERNIE 5.5 Flash Base: sizing giant open MoE models</title><link>https://www.ozercemkelahmet.com/articles/open-weight-giant-moe-models</link><guid>https://www.ozercemkelahmet.com/articles/open-weight-giant-moe-models</guid><pubDate>Sun, 08 Nov 2026 07:00:00 +0000</pubDate><description>Moonshot and Baidu published very large mixture-of-experts weights this week. Here is how MoE works, what the licences allow, and how much memory hosting needs.</description></item>
</channel></rss>