Jev's Boring Use Cases Are the Ones That Actually Work
Jev is a lightweight model that runs far faster and cheaper than larger LLMs, delivering competitive accuracy for tasks such as reranking, j...
Sep 27, 2026
![First Post](data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 100 100"><defs><linearGradient id="gFirst" x1="0%" y1="0%" x2="100%" y2="100%"><stop offset="0%" stop-color="%238b5cf6" /><stop offset="100%" stop-color="%23ec4899" /></linearGradient><filter id="fFirst"><feDropShadow dx="0" dy="2" stdDeviation="3" flood-opacity="0.3"/></filter></defs><circle cx="50" cy="50" r="45" fill="url(%23gFirst)" filter="url(%23fFirst)"/><circle cx="50" cy="50" r="38" fill="none" stroke="rgba(255,255,255,0.4)" stroke-width="2"/><path d="M60 30 C 50 30, 40 40, 35 55 L 30 70 L 45 65 C 60 60, 70 50, 70 40 C 70 35, 65 30, 60 30 Z" fill="%23ffffff" opacity="0.9"/><path d="M35 55 L 45 65" stroke="%23ec4899" stroke-width="2"/><circle cx="65" cy="25" r="3" fill="%23fef08a"/><path d="M65 15 L65 20 M65 30 L65 35 M55 25 L60 25 M70 25 L75 25" stroke="%23fef08a" stroke-width="2" stroke-linecap="round"/></svg>)
![Weekend Warrior](data:image/svg+xml;utf8,<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 100 100"><defs><linearGradient id="gWknd" x1="0%" y1="0%" x2="100%" y2="100%"><stop offset="0%" stop-color="%231d4ed8" /><stop offset="100%" stop-color="%2360a5fa" /></linearGradient><filter id="fWknd"><feDropShadow dx="0" dy="2" stdDeviation="3" flood-opacity="0.3"/></filter></defs><circle cx="50" cy="50" r="45" fill="url(%23gWknd)" filter="url(%23fWknd)"/><circle cx="50" cy="50" r="38" fill="none" stroke="rgba(255,255,255,0.4)" stroke-width="2"/><path d="M30 30 C 50 20, 50 20, 70 30 L 70 50 C 70 70, 50 80, 50 80 C 50 80, 30 70, 30 50 Z" fill="%23ffffff" opacity="0.9" filter="url(%23fWknd)"/><rect x="40" y="45" width="20" height="20" rx="2" fill="%231d4ed8"/><rect x="40" y="45" width="20" height="5" fill="%23bfdbfe"/><line x1="45" y1="42" x2="45" y2="48" stroke="%231e3a8a" stroke-width="2"/><line x1="55" y1="42" x2="55" y2="48" stroke="%231e3a8a" stroke-width="2"/></svg>)
3 min
Subscribe to Lê Đức Minh's Newsletter
Get new articles and engineering posts delivered to your inbox.
I Reproduced RLCD to Find Out What Its "Calibrated RL" Actually Does
Reproducing Laya’s RLCD shows that its “RL” term is just a noise‑smoothed estimate of the cross‑entropy gradient, which makes inference over...
Sep 26, 2026
4 min
I Took Laya's "Calibrated RL" Apart. It's Cross-Entropy With Extra Steps.
Laya's README reports a raw ECE of 0.466 on the typed-decisions benchmark before temperature scaling.
Sep 26, 2026
5 min
Inside RLCD: The Estimator, the Proof, and the 30 Runs That Changed My Mind
I went into this expecting the RL to be the interesting part. Laya's whole pitch is that a reinforcement-learning term turns noisy logits in...
Sep 26, 2026
7 min
Pi‑warden: Using Jev to Block Destructive Commands in 48 Hours
Jev, a decision‑only AI model released on September 15 2026, was quickly adopted by developers: pi‑warden, built within 48 hours, blocked 42...
Sep 26, 2026
4 min
Understanding RLCD: Calibration Techniques for Decision‑Only LLMs
RLCD is a proprietary training method from TypeSafe AI that teaches Jev to produce calibrated probability distributions and confidence score...
Sep 25, 2026
4 min
Jev, Explained: The AI That Refuses to Write a Word
Jev is a “System One Model” that returns typed, calibrated probabilities for predefined question types in a single forward pass, enabling fa...
Sep 24, 2026
4 min
SmolVLM2 vs Qwen2.5-VL: Real-Time RTSP Edge Video Summarization Under 8GB VRAM
Connect a vision-language model to a live RTSP surveillance feed, ask it to generate real-time incident summaries, and watch your GPU metric...
Sep 23, 2026
9 min
From Codebase to Cognitive Skill: Inside BAAI's Repo-to-Skill Framework
Give an autonomous coding agent a hundred-thousand-token context window, point it at a GitHub repository, and ask it to reproduce an ML base...
Sep 22, 2026
12 min
Dynamic Reasoning Budgets: Halving Inference Costs on Test-Time Compute Models
Dynamic reasoning budgets route simple tasks to a zero‑thought path and reserve capped reasoning for complex queries, cutting token use and ...
Sep 17, 2026
5 min
The Harness is the Moat: Building a Deterministic Agent Runtime with Context Pruning
Learn how a deterministic harness prunes context, uses a ledger and transactional tool calls to keep LLM agents reliable over many turns.
Sep 16, 2026
6 min
What DeepSeek's Open-Source Agent Harness Gets Right
A few weeks ago, my dad and I spent an afternoon at the kitchen table trying to figure out what an "AI agent" actually is.
Sep 16, 2026
8 min
I Tested 3 LLMs on Vietnamese Idioms. They Got It Backwards.
State‑of‑the‑art LLMs, even Vietnamese‑specialized ones, score below 50 % on the VIVID benchmark of 1,636 authentic Vietnamese idioms and fa...
Sep 10, 2026
6 min
AI Reads Vietnamese Slang as Angry. 3 Models Don't.
Benchmark reveals Llama-3.1-8B mislabels Vietnamese slang as angry, losing 20 F1 points, while three other LLMs correctly interpret social m...
Sep 9, 2026
6 min
The Real Vietnamese LLM Tokenizer Cost (It's Not 4.5x)
Discover why Vietnamese tokenizers cost only 1.05‑2.14× English tokens, not 4.5×, and how updating tokenizers can cut LLM expenses by a thir...
Sep 8, 2026
7 min
A Simple Look at TokPress: A Compressor That Uses an LLM's Tokenizer
TokPress compresses tiny JSON log lines by tokenizing with OpenAI's o200k_base tokenizer, then applying LZ77 and rANS for smaller files.
Sep 6, 2026
2 min
A Simple Look at TokPress: A Compressor That Uses an LLM's Tokenizer
TokPress compresses tiny JSON log lines by tokenizing with OpenAI's o200k_base tokenizer, then applying LZ77 and rANS for smaller files.
Sep 6, 2026
2 min
Fun Project: I Built a Compressor That Thinks in Tokens
Field notes from a weekend spent teaching an entropy coder to speak LLM. LLMs spent billions of dollars learning the best subword dictionary...
Sep 5, 2026
10 min
Agent harness or agent framework?
In July I wrote Prompt, Context, Harness, Loop, which split an agent into four parts and argued that the harness — the thing that owns the l...
Sep 5, 2026
6 min
Five ways to invalidate your prompt cache
Part 2 of 2 on prompt caching. Part 1 covered the economics. Part 1 established the prize: caching cut an 80-turn agent session from $54.08 ...
Sep 2, 2026
9 min
Why your coding agent's bill grows faster than the chat
Part 1 of 2 on prompt caching. I got curious about a number I had never actually checked: what does a long session with a coding agent reall...
Sep 1, 2026
8 min
Discussion
How do you deduplicate a stream you don't control?
Field notes, part 1 of 3. Data plumbing from an AI engineer's desk.* My job title says AI. A meaningful share of my week is data. Not the gl...
Aug 15, 2026
6 min