llm

9 articles · newest first

Hand-drawn editorial illustration of a small robot reading a long paper letter, while a tiny coral hand reaches out of the page to steer the robot's hand, on a near-black background

Your AI agent just read a web page. Who is it obeying now? Prompt injection, explained

Prompt injection is how text your AI reads turns into orders it follows. What direct and indirect injection are, why tools raise the stakes, Simon Willison's lethal trifecta, the defenses OWASP, Anthropic and OpenAI recommend, and their own admission that none of it is a full fix.

Aliteq Team · Oct 16, 2026 · 11 min

Hand-drawn editorial illustration of a single paper ribbon flowing one way out of a speaker cone toward an open hand, in violet and coral on a near-black background

Software

Your AI chat types its answer word by word. That's SSE, not WebSockets. When you need which

Claude's and OpenAI's APIs stream tokens over server-sent events, a plain HTTP response that never hangs up. Here's when that's all you need, when you really need WebSockets, why polling is fine more often than you'd think, and the proxy and serverless clocks that cut long connections.

Aliteq Team · Oct 16, 2026 · 13 min

Flat vector illustration of a person in a hoodie sitting cross-legged beside an open filing cabinet drawer of blank documents joined by thin coral lines, on an indigo-violet background

Google's Open Knowledge Format (OKF), Explained: Markdown Files That Agents Can Trust

Google Cloud's new spec for agent knowledge is a folder of markdown files with a few lines of YAML on top. Here is what is in the file, what v0.2 added, and where it differs from a vector index.

Aliteq Team · Oct 3, 2026 · 8 min

Hand-drawn editorial illustration of a small chat-bubble robot with a very long questionnaire scroll unrolling from it across a dark page

Security

The AI Questions Enterprise Buyers Ask About Your LLM, With Answer Templates

The vendor review now has an AI tab. The eight things it asks about an app that calls OpenAI or Claude, where each question comes from (SIG, CSA's AI-CAIQ, NIST, ISO 42001, OWASP, the EU AI Act), a fill-in-the-blanks answer for each, and the evidence to attach.

Aliteq Team · Oct 2, 2026 · 14 min

Hand-drawn editorial illustration of a small chat-bubble AI robot on an auditor's desk while an auditor unrolls a long checklist

Security

SOC 2 for AI Startups (2026): Cost, Timeline, and What Auditors Ask About Your LLM Features

Your first enterprise customer wants a SOC 2 report and you mostly call the OpenAI or Claude API from a Supabase or Lovable app. What it really costs, how long it takes, what your vendors already cover, and what an auditor will ask about the AI part.

Aliteq Team · Updated Oct 3, 2026 · 14 min

Mark Zuckerberg speaking about Meta's AI strategy

meta just promised to open-source the AI model that actually rivals GPT and Claude

Muse Spark 1.2 isn't the laptop-sized model Meta gave away last week — it's the one built to fight Claude Opus 5 and GPT-5.6 head-on, and Zuckerberg just said it's going free too.

Aliteq Team · Aug 13, 2026 · 6 min

Line drawing of a graphics card next to the headline What fits 8GB? Local AI models that run on an RTX 4060 or 5060, with a card reading Qwen3.5 9B, Q4_K_M, 32K context, 7.2 GB, our VRAM engine, Nov 3, 2026

Which local AI models fit on an 8GB GPU? Qwen3.5 9B, Gemma 4 and what doesn't

Qwen3.5 9B and Gemma 4 E4B fit an 8GB card with room for real context. Gemma 4 12B only just fits. Qwen3.8 27B and gpt-oss-20b don't, and Qwen3.8 Flash Next needs 48GB or more of system RAM. Every number below, model by model, re-checked Nov 3, 2026.

Aliteq Team · Updated Nov 3, 2026 · 8 min

rows of server infrastructure representing large-scale cloud computing

Hardware

Qwen3.8-Max open weights (2.4T-A95B): what it takes to run, and the 27B you can actually run

Alibaba put its Max-class model on Hugging Face as Qwen3.8-2.4T-A95B. The download is 2.5 TB in FP8 and the smallest community quant is still 397 GB. Here is the file-size math, the license, and the sibling that fits a 24 GB card.

Aliteq Team · Updated Oct 29, 2026 · 8 min

Thumbnail: Fine-tune an LLM, rent the smallest card that fits. QLoRA needs 6 GB for 8B and 41 GB for 70B. A card reads RTX 4090, Vast median, $0.47/hr, read Nov 3, 2026, with the Vast.ai and Runpod wordmarks.

Which cloud GPU should you rent to fine-tune an LLM, and what does a run cost?

Pick the card by the VRAM your fine-tune needs, not by the biggest name. On Nov 3, 2026 a 24GB RTX 4090 rented for a $0.47/hr median on Vast.ai ($0.89 on Runpod), enough for QLoRA up to about 14B; a 70B QLoRA needs an 80GB A100 from about $0.95/hr. Here is the card for each job and the math for a run.

Aliteq Team · Updated Nov 3, 2026 · 9 min