Tag
47 articles tagged #LLM.
P.01Mistral Large 4 is a 1-trillion-parameter model with 49B active, in API preview since Oct 6, weights due Oct 27. What is confirmed, what isn't, what to do.
P.02GPT-6.1 Sol launched September 29 near Astra's scores at about a fifth of the cost per task. What the numbers mean for agent loops, and what to test first.
P.03vLLM v0.28.0 moved bitsandbytes quantization out-of-tree, breaking upgrades for 4-bit and 8-bit inference. How to check if you're affected and fix it fast.
P.04Guardrails AI validates structured output, NeMo Guardrails controls dialog flow, and Llama Guard classifies safety. Here is which one fits your LLM app.
P.05Sakana AI shipped Fugu Ultra v2 on Sept 11, routing every request across a hidden pool of models instead of running one. What that buys you, and what it costs.
P.06The OpenAI Agents API runs the Codex harness for you: durable sessions, sandboxes and MCP tools. How to integrate it in Python, and when to pick something else.
P.07Abliteration.ai sells API access to open-weight AI models with safety refusals surgically removed, and the buyer inherits every future flaw.
P.08Google shipped Gemini 3.8 Flash on September 2, its fourth Flash release since May. Pricing didn't move and the base model didn't change. What did.
P.09What OpenAI, Google, and xAI actually charge per million tokens right now, tier by tier, so you can price a real workload before picking a provider.
P.10Unit tests don't work on a feature that answers differently every time. Evals do. How to build a practical eval harness for an LLM feature, with real code.
P.11Red teaming means attacking your own prompts, retrieval pipeline, tools, and guardrails before a stranger does. The methodology, and tools that automate it.
P.12LoRA and QLoRA fine-tune a multi-billion-parameter model on one consumer GPU by training small adapter weights. How each works, and when to pick which.
P.13DeepSeek retired deepseek-chat and deepseek-reasoner on July 24, 2026 and added peak-hour pricing. The real fix if your integration broke, not just a rename.
P.14GPT-5.6 Sol runs on Cerebras wafer-scale hardware at up to 750 tokens per second, roughly 10x typical GPU inference. Who that speed is actually for.
P.15AI engineer is not ML engineer. One trains models, the other builds products on them. The screen that tells them apart and finds people who actually ship.
P.16LLM bills grow faster than usage. Prompt caching, semantic dedup, tiered routing, and batch inference cut 40-80% without degrading output quality.
P.17Comparing LangSmith, Braintrust, and W&B Weave for LLM evaluation: what each does well, where each breaks down, and a minimum viable eval pipeline.
P.18Both frameworks can build RAG pipelines and agent systems, but they're designed with different priorities. Here's when to reach for each and when to skip both.
P.19Ollama runs Llama, Mistral, Phi-4, and dozens of open-weight models on your laptop with one command. Here's what actually works and when to use it.
P.20Workers AI runs open-weight models (Llama, Mistral, Whisper, embeddings) inside Cloudflare's network. What's useful, what the limits are, and when it fits.
P.21Getting a language model to return reliably structured data is not just about asking nicely. Here's the pattern that actually works at production scale.
P.22MCP is the standard for connecting AI models to external systems. How it works, how to implement a server, and what to lock down before production.
P.23Clients ask agencies to automate PDFs constantly. Here's how to actually build document extraction pipelines: OCR, vision models, and validation.
P.24RAG and fine-tuning solve different problems: one changes what a model knows, the other changes how it behaves. Here's how to decide which one you need.
P.25Three agent orchestration frameworks, three different mental models. Here's when each earns its place, what it costs in complexity, and the choice at 2am when it breaks.
P.26An agent that forgets everything when the session ends is a limited tool. Here are the practical patterns for building different kinds of memory into your agents.
P.27When your AI agent needs to run the code it writes, you can't let it touch your production servers. Here's how the main isolation options work and when to use each.
P.28AI SDK 7 replaces maxSteps with stopWhen: isStepCount(), renames system to instructions, and moves structured output into generateText. Working code for each.
P.29RAG, constrained generation, output validation, and self-consistency reduce LLM hallucination in production. Here is what each layer costs and when to use it.
P.30LiteLLM and OpenRouter add failover, cost caps, and model fallback to your LLM calls. Here is how each works and when a routing layer is worth building.
P.31Add similarity search to your existing Postgres database using pgvector. Real setup, indexing strategies, and when you actually need a dedicated vector database.
P.32LLM observability means tracking traces, token costs, latency, and output quality to debug production failures instead of guessing. Covers Langfuse and Helicone.
P.33Unit tests confirm your code runs. They don't confirm your AI feature gives good answers. Here's how to build an eval pipeline that catches real failures.
P.34LLM calls are slow and expensive, so caching is the obvious fix. Here's when it backfires and how to implement exact-match and semantic caching.
P.35Feature flags let you ship AI features to a small slice of users first, watch for quality problems, and kill the feature in under a minute if something breaks.
P.36Getting valid, schema-conforming JSON from an LLM takes more than a JSON mode flag. Here is what works in production, from native APIs to instructor.
P.37Every AI budget starts with API costs and ends in surprises. What production AI features really cost once evaluation, observability and prompt rot are counted.
P.38Prompt injection is the SQL injection of the AI era. Here's what the attack looks like, why it can't be patched, and how to actually defend against it.
P.39ZeroDayBench tested GPT-5.2, Claude Sonnet 4.5, and Grok 4.1 against 22 novel zero-day vulnerabilities. Here's why they failed to patch 73% of them, and what it means.
P.40EditorPickContext engineering means choosing what enters an LLM's context window: instructions, retrieved docs, tool output, memory and history. Eight techniques.
P.41Start with hybrid search: dense vectors plus BM25, fused with RRF and reranked. How dense, BM25, SPLADE and ColBERT compare, and when long context wins.
P.42Run Llama 4, Qwen3, Phi-4, and Mistral on consumer GPUs like the RTX 4090 and 5090. Covers quantization, inference engines, VRAM needs, and local vs. API costs.
P.43Use reasoning models for multi-step code, math and planning, not lookups or chat. How effort and thinking levels work in Claude, OpenAI, Gemini and DeepSeek.
P.44EditorPickBuild RAG that works: chunk on structure, add context to each chunk, index it in vectors and BM25, then rerank and check citations. Code for every step.
P.45EditorPickAI-first web agencies build apps with built-in intelligence, like chatbots and predictive features, as one product instead of two disconnected teams.
P.46EditorPickPrompt engineering shapes behavior, RAG adds knowledge, fine-tuning changes reasoning. Here's the cost and performance comparison to pick the right one.
P.47Traditional test suites break on non-deterministic AI outputs. Here's how we QA LLM features: constraint checks, golden datasets, and LLM-as-judge evaluation.