PastAGI

Practical AI guides, honest tool reviews, engineering deep dives, real-world use cases, and sharp analysis that cuts through the hype.

Recent

Text-to-SQL evaluation tests whether AI-generated SQL returns defensible answers on real warehouse data instead of just matching public benchmark scores.

Engineering • Aug 27, 2026 •12 min read

Text-to-SQL Evaluation That Catches Silent Wrong Answers

Public benchmarks say 89%, your warehouse says otherwise. Build a text-to-SQL evaluation with schema-specific oracles that catches silent wrong answers.

#text-to-sql #nl2sql +4

A GPU cluster in a data center represents the hardware baseline behind MoE serving cost, where residency, communication, and utilization set the real invoice for a sparse model.

Engineering • Aug 26, 2026 •12 min read

MoE Serving Cost Math for 6 of 125B Active Parameters

MoE serving cost for a 6-of-125B model is not 6B per token. All 125B stay in VRAM, so run the builder math on residency, routing, and break-even.

#moe #inference-cost +4

AI agent memory depends on layered persistent stores rather than raw context windows to retain recruiter preferences and candidate history across sessions.

Use Cases • Aug 25, 2026 •12 min read

AI Agent Memory Lessons From LinkedIn's Hiring Assistant

AI agent memory lessons from LinkedIn's hiring assistant. This four-layer teardown covers token payback math, decay rules, and privacy classes to copy.

#agent-memory #ai-agents +3

Speculative decoding turns idle CPU compute into nearly four times faster token generation for large language model inference.

Engineering • Aug 24, 2026 •11 min read

Why Speculative Decoding Pays Nearly 4x on CPUs

Speculative decoding turns idle CPU cores into 4x faster LLM generation. Learn why it works, when gains collapse, and when CPU beats GPU or API.

#speculative-decoding #cpu-inference +3

AI agent token usage overtaking human-driven traffic on LLM API gateways and reshaping how platforms meter, price, and serve machine callers.

News • Aug 23, 2026 •11 min read

AI Agent Token Usage Overtook Humans on OpenRouter

OpenRouter data shows AI agent token usage passed human traffic on February 6, 2025, with 14x growth and ~70 percent cached. Here is how to audit your mix.

#ai-agents #token-usage +4

WhatsApp Scam Alert flags scam patterns on the phone itself while message content remains unreadable to servers under end-to-end encryption.

Use Cases • Aug 22, 2026 •12 min read

How WhatsApp Scam Alert Detects Scams It Cannot Read

WhatsApp Scam Alert flags scams without Meta reading your messages. See how on-device AI works under end-to-end encryption and how to copy the pattern.

#whatsapp #on-device-ai +4

A teardown of Stampli's AI-assisted product launch claim, examining what the 68% launch-hour reduction measures and the conditions that made it possible.

Use Cases • Aug 21, 2026 •10 min read

AI-Assisted Product Launch Teardown of Stampli's 68% Claim

A builder's teardown of Stampli's 68% launch hour cut with Codex and ChatGPT, and the AI-assisted product launch workflow your team can copy.

#codex #chatgpt +4

An AI text detector built on a locally fine-tuned language model scores writing samples to separate machine-generated text from human writing.

Guides • Aug 20, 2026 •11 min read

Build an AI Text Detector, Then Test If It Can Ship

Build an AI text detector with small local models, stress-test it on short and human-edited text, and use the error rates to decide if it ships.

#ai-text-detection #local-models +3

Deep learning GPU training where gradient accumulation splits an effective batch into micro-batches that change wall-clock time on rented hardware.

Engineering • Aug 19, 2026 •12 min read

The Real Cost of Gradient Accumulation on T4 and L4

Gradient accumulation can make identical batches train at different speeds. Learn why micro-batch shape drives T4 vs L4 wall-clock time and throughput.

#gradient-accumulation #gpu-benchmarks +4

LLM context window management keeps a growing agent conversation within the limited space a model can actually read and reason over.

Guides • Aug 18, 2026 •11 min read

LLM Context Window Management With a Token Budget

Learn LLM context window management with a token budget ledger, a stepwise compression ladder, and the prompt cache trap that punishes trimming.

#context-windows #token-budget +4

OpenAI Ultrafast mode accelerates GPT-5.6 Sol decode on dedicated inference hardware, with real-world gains set by each workload's decode share.

Tools • Aug 15, 2026 •9 min read

The Real OpenAI Ultrafast Mode Speedup, Workload by Workload

OpenAI Ultrafast mode promises up to 14x speed, but decode is only part of request latency. Here is the math for chat, RAG, tool loops, and voice agents.

#openai-ultrafast #gpt-56-sol +4

LangChain versus LangGraph comparison for orchestrating stateful AI agents with fundamentally different execution models.

Tools • Aug 13, 2026 •11 min read

LangChain vs LangGraph for Stateful AI Agent Orchestration

The langchain vs langgraph decision is a shift from stateless DAGs to cyclic state machines for building stateful autonomous AI agents in production.

#langgraph #langchain +3