The Agent Teaches Itself With Skills It Wrote: How SEED Fixes the Sparse-Reward Problem
The real bottleneck in agentic RL is that the reward arrives only once, at the very end of a trajectory. SEED turns that single sparse signal into dense per-...
All Posts
194
개의 게시글이 있습니다.
Longer Thinking, Linear Cost: How the Markovian Thinker and Delethink Redesign Long Reasoning
The real cost of long reasoning comes from the state growing without bound. Markovian Thinking breaks reasoning into fixed-size chunks and passes only a shor...
How vLLM Stays Solid at 2,000 Commits a Month: Three Devices in CI, Benchmarking, and Releases
vLLM merges roughly 2,000 commits into main every month and still holds production quality. The secret is not ‘more tests’ but three deterministic devices: a...
Claude Code + iOS Simulator: A Closed Coding Loop Where You Build, Run, and Watch
Claude Code desktop has shipped a public beta feature that opens the iOS simulator in a panel right next to the conversation. Here is what changes when Claud...
Drawing Architecture Diagrams With Words: We Ran Archify and Mapped the ThakiCloud Stack
Archify is an agent skill that generates self-contained HTML architecture diagrams from plain-language descriptions, no Mermaid syntax required. We installed...
The Image AI That Finally Spells — Whose Brush Is It?
The letters come out perfect. The brush is bolted to someone else’s cloud.
The Agent That Writes Code and the Agent That Audits It Arrived on the Same Day
On July 22, two open weight releases stood facing each other like mirror images. One generates code, the other finds vulnerabilities in it. Yet neither answe...
Why Dynamic Batch Tuning Failed to Beat Static Scheduling: An Honest Failure Analysis of a Multi-Tenant vLLM Serving Control Loop
Does an agent that observes GPU telemetry and adjusts batch size and concurrency in real time outperform static Kueue admission control? The simulation said ...
Qwen-Image-3.0 Unveiled: A Third-Generation Image Model Built on ‘Real’, Weights Still Pending
Alibaba’s Qwen team has announced Qwen-Image-3.0, its third-generation image generation model. It leads with 4.5k-token input, 10px micro-text rendering, and...
Hugging Face Wasn’t Breached by a Human, but by an Autonomous AI Agent: When the Dataset Pipeline Became the Attack Surface
In July 2026 Hugging Face disclosed an internal breach driven by an autonomous AI agent. The entry point was a single malicious dataset, and two vulnerabilit...
Switch off the OpenAI Realtime API in one line: hugging-voice, an open voice stack you run yourself
Hugging Face’s hugging-voice and its engine speech-to-speech wrap a full realtime voice pipeline, from VAD through STT, LLM, and TTS, behind an OpenAI Realti...
Claude Code Screen Reader Mode: One Flag That Opens Terminal AI Coding to Everyone
Claude Code added a screen reader mode that swaps its visual terminal UI for plain, linear text. Here is what claude --ax-screen-reader actually changes, how...
158 Skills and 24 Agents in One Plugin: How a Deterministic Skeleton Tames Agent Explosion
The open-source marketing plugin Digital Marketing Pro bundles 158 skills and 24 specialist agents without collapsing. The trick is a deterministic skeleton:...
The Best AI That Fits Your Graphics Card
We sized the model to the card we already owned. The only thing that starved was the cloud invoice.
An Agent Swarm Rewrote SQLite in Rust: Cursor’s Multi-Agent Economics
Cursor showed an agent swarm that rebuilt SQLite in Rust from only its 835-page manual. It passed 100% of a held-out test suite, cost varied 15x by model mix...
Harness Design and Post-Training Aren’t Separable: Harness-Aware Post-Training and LLM Agent Performance
Tool-using LLM agents run on top of a harness that wraps the model. A recent arXiv paper shows empirically that treating harness design and post-training as ...
Putting a Safety Net Under Autonomous Skill Rollout with Canary Gates: Trace-Based Regression Detection and Automatic Rollback
How should an agent harness that rewrites its own code overnight safely roll out the result the next day? An experiment pairing canary releases with automati...
Qwen3.8, 2.4 Trillion Parameters Open-Weight Preview: What’s Confirmed and What Isn’t
Alibaba has previewed Qwen3.8, a 2.4-trillion-parameter model, promising an open-weight release. The preview is already testable, but neither the weights nor...
No Guardrails: Who Holds the Refusal Power With Kimi K3
The over-refusal problem, where closed models block legitimate security, medical, and legal work, is back in the spotlight. Moonshot’s open-weight Kimi K3 sa...
Can We Trust a 2.8 Trillion Parameter Open Model: Kimi K3 and Benchmark Reliability
Kimi K3, released by Moonshot, is the largest open-weight model in history at 2.8 trillion parameters. The scores are dazzling, but a benchmark-overfitting d...
A 122B Model on a 24GB Card? We Dissected ATSInfer, Which Slices llama.cpp at Tensor Granularity
Instead of offloading whole layers or experts, ATSInfer places individual tensors across CPU and GPU. The paper reports running models far beyond VRAM on a s...
The Question Behind a 25 Billion Won Bill: Why AI Agent Costs Have Become Invisible
From a 25 billion won bill sent to a user in Korea to allegations of overbilling at 60 companies, the past month of AI cost news points to a single blind spo...
everything-claude-code: Dissecting an AI Coding Harness Battle-Tested Over Six Months
AI coding tools forget your rules every new session. An Anthropic hackathon winner solved this by open-sourcing the config they refined over six months on a ...
Rereading Musk’s Davos Forecast: Intelligence in Five Years, and Where That Intelligence Actually Runs
Elon Musk’s three Davos predictions got recompressed online into a warning that ‘the age of the paycheck ends in three years.’ We transcribed the actual inte...
Aced the Exam, Blanked on the Job
Topped the leaderboard — on somebody else’s exam.
Inside Kimi Code CLI: How an Open Source Terminal Agent Swallows Editors via ACP
We dig into the open source coding CLI that Moonshot AI released alongside Kimi K3, working strictly from the official docs and repository. From coder/explor...
Why We Cut the System Prompt by 80 Percent: Smarter Models Want a Thinner Harness
News that Anthropic stripped Claude Code’s system prompt down by 80 percent made the rounds among developers. The person behind the change explained that the...
Controlling Reasoning Effort - How LLMs Learn Low, Medium, and High Thinking Modes
With GPT-5.6 shipping five or six reasoning effort settings per size, effort control has become table stakes for reasoning models. We dig into what training ...
Unified Multimodal Without a VAE: SenseNova U1, NEO-Unify, and On-Premise Serving
SenseTime has released 日日新 SenseNova U1 under Apache 2.0. Its NEO-Unify architecture drops both the visual encoder and the VAE, handling understanding, gener...
The Same Morning, Two Price Tags Moved in Opposite Directions
A memory shock has pushed the cost of owning AI infrastructure to an all time high, while Kimi K3 and Chinese open weight models have dragged the cost of usi...
The Era of AI Completing 32-Step Intrusions Start to Finish: GPT-5.6 Sol and Cybersecurity
OpenAI announced that GPT-5.6 Sol set a new record on a cyber range. As frontier models begin autonomously executing real attack chains, the decisive factor ...
Where Does a Real-Time Voice Agent Actually Bottleneck: A Latency Budget Calculator and GPU Serving Measurements
We built an open tool that diagnoses which stage of your voice agent is the bottleneck, without any vendor SDK, and measured our actual stack (Qwen3-ASR, Vox...
A $400K Rack on a 24GB Graphics Card? We Reproduced ktransformers’ ‘28x’ Ourselves
ktransformers claims you can run a giant MoE model on a single 24GB GPU by offloading experts to CPU. We tested the viral ‘28x’ and ‘$400K to 24GB’ claims by...
One Genius vs. a Whole Crowd
One clever agent hits a wall fast. So Paxis just sent the whole swarm.
Coding with Kimi K3: Wiring a 2.8T Open Model into the OpenCode Terminal Agent
Kimi K3, called Fable 5-class by many, can run inside an open-source terminal agent rather than a locked proprietary IDE. We installed OpenCode to verify the...
Your AI Bill Leaks Through Repetitive Work — How On-Prem Specialized Models Cut It Structurally
Most of an AI agent’s cost isn’t smart judgment — it’s simple, repetitive decisions made thousands of times a day. Peel those off into small specialized mode...
Kimi K3: What It Actually Takes to Serve a 2.8 Trillion Parameter Open Weight Model
Moonshot has released Kimi K3, the largest open weight model in the world to date. What is striking is not just that it beat top closed models on a frontend ...
How vLLM Works, and How It’s Used in Production
When you put an LLM into real service, most of the cost is decided not by the model itself but by the inference engine. Here’s how vLLM cuts GPU waste with P...
One 24GB Card Ate a $400K Rack
The rack you were about to rent is already in your drawer.
Why Nobody Buys the Leaderboard’s Number One: What Benchmarks Cannot Buy
China’s Kimi K3 overtook Claude on the coding leaderboard, yet Silicon Valley stayed cold. The same week, Google delayed Gemini 3.5 for a third time. At the ...
Skill Graveyard: Safely Deprecating Skills in an Agent Harness With Over 2,000 Units
If your team has kept expanding its catalog of agent skills or MCP tools, you’ve probably felt routing accuracy start to slip at some point. What happens if,...
From 46% to 90% Inside a Phone: Fine-Tuning Small LLMs for On-Device Agents
Google AI Edge’s Cormac Brick presented a case where fine-tuning the 270M-parameter FunctionGemma lifted accuracy on a specific agent task from 46% to 90%. T...
295B on a Single Card: Anatomy of Hunyuan Hy3’s 1-bit and 4-bit Serving
Tencent’s Hy3 1-bit and 4-bit GGUF builds shrink a 295B MoE from 598GB to 85.5GiB so it runs on a single GPU. But the single GPU here means a 128GB-class uni...
Effort Levels for Code Review: Claude Code /code-review from low to ultra
In v2.1.101 Claude Code renamed /simplify to /code-review and attached effort levels to the review. low and medium return a few high-confidence findings, hig...
Low Effort Somehow Won the Code Review
We cranked the effort dial to max. So did the invoice.
Claude Code Artifacts Now Call MCP Connectors: Building Live, Self-Refreshing Dashboards
Artifacts used to end as static markdown. Now they’re living apps that call connectors. Here’s how to build a dashboard that pulls fresh data every time you ...
Autonomous Remediation or Human Escalation: Calculating the Safety Boundary for Kubernetes GPU Incident Response
Any team trying to hand Kubernetes GPU incident remediation to an LLM agent eventually runs into this question: how long should the agent be left to fix thin...
The Day Apple’s Crown Sat on Someone Else’s Chips: The AI Era Moves From Owning to Governing
Apple, the symbol of vertical integration, gives up its own chip and borrows Google’s GPUs and even a Chinese rival’s model. On the same day, Korea pours ten...
Fitting a 27B Model on a Phone: Dissecting Bonsai 27B’s 1-bit and Ternary Quantization
Bonsai 27B, released by PrismML, is not a newly trained model but the result of compressing Qwen3.6-27B’s weights to 1-bit and ternary while leaving the arch...
Melting Your Brain by Hitting Enter: The Cognitive Cost of Parallel AI Coding and the Art of Delegation
A day spent with six terminals open, waiting on responses and just smashing enter. This new kind of labor, called multi-Clauding, is quietly draining develop...
The AI That Tutored Another AI Overnight
An AI tutored another AI to an A overnight. Whose GPUs paid for the study session?
The Gains From Self-Evolving Harnesses May Be an Illusion: Separating Harness Updating From Harness Benefit
The gains attributed to self-evolving harnesses are a mix of ‘the ability to produce good updates’ and ‘the ability to use those updates well,’ tangled toget...
Encode Domain Knowledge as Infrastructure: Redefining Automation in the Agent Era
If an agent can’t work productively in a codebase, that is not a failure of the model. It is a failure of automation. Moving domain knowledge into infrastruc...
Agents Take the Wheel on GPU Training: Dissecting NVIDIA Cosmos 3 Agent Skills
Give a coding agent two natural-language prompts, and post-training a vision foundation model finishes in a single day. We dissect NVIDIA’s agent skill and l...
If You Want an LLM to Generate Screens From a Design System, Build the Gate Before the Model
We expected fine-tuning to win. The data said otherwise. Here’s how we ended up with a generator that respects TDS without any training, and the next problem...
Implement Once, Serve at vLLM Native Speed: the Transformers Backend Ends the Double-Build
Until now, every new model architecture had to be built twice: once in Transformers for training and research, and again in vLLM for production inference. vL...
The Anatomy of GPU and TPU Clusters: How Collective Communication Determines the Speed of Distributed Training
When training or serving a large model across many accelerators, the real bottleneck usually isn’t computation, it’s the data moving between accelerators. Th...
The Interview Where the AI Does the Grilling
The AI that grills candidates decided to grill itself first.
Have It Write the Goal First on Hard Tasks: Codex Goal-Delegation Prompting
A developer shared a tip: when handing Codex a genuinely hard /goal, first ask it to write the goal so that another thread can achieve it. It sounds like wor...
When Can You Trust the Cheap Model: Cutting Multilingual Document OCR Costs with a Confidence-Gated VLM Cascade
Even in an era where sub-billion-parameter, ultra-lightweight VLMs read documents well, the truly hard documents still need a large model. We introduce an an...
The Five Layers the Model Map Never Draws: 2026’s AI Contest Is Decided Below the Stack
The question isn’t whether GPT-5.6 is smarter. It’s who serves those tokens cheaper and faster, because that’s what decides the margin. We map the five stack...
Where Do You Put a Quantized Model? Four AWS and Unsloth Deployment Patterns
Plenty of teams know how to shrink a model to 4-bit with Unsloth. But the moment they have to decide whether to put that file on EC2, wrap it in a SageMaker ...
We Deleted 423GB from GLM-5.2 Without Quantization: Measuring the Waste in BF16 Exponents
Someone claimed they shrank GLM-5.2 from 1403GB to 980GB. Not quantization, not pruning, but lossless compression that is bit-for-bit identical to the origin...
Running a Monster Model on Your Junk Drawer
Couldn’t afford one monster, so we ganged up all the little ones.
Autonomy Is Not a Switch, It Is a Dial: What Banks Buy Before Capability
The Korea Federation of Banks did not order a smarter model, it ordered a sequence for adoption. Reading today’s moves across regulated industries closely re...
Auditing an Autonomous Nightly Paper-Generation Pipeline Over 8 Days: A Case Study in Reliability, Diversity, and Research Integrity
This study audits a fully autonomous pipeline that generates a paper every night, not by inspecting the final output but by examining its actual operational ...
The Skill That Installs Skills, Installing Skills
Drop your repo, it installs everything — someone else’s taste and all.
Claude Code Can Watch Videos: claude-video Feeds Frames and Transcripts Into the Agent via /watch
Coding agents have long read only text. claude-video thinly wires yt-dlp, ffmpeg, and Whisper together to turn YouTube, Zoom, Loom, or local files into frame...
Validate Expensive, Then Descend: Cutting a Skill Fleet’s Model Cost with Code Gates
When a company’s recurring work is defined as skills, each skill can tell you, through measurement rather than intuition, exactly how far down the model tier...
Guard the Router and Quantization Hurts Less: An Accuracy-Cost Map for Selective NVFP4 Quantization in MoE
For engineers serving MoE models on H200 or Blackwell-class clusters, we introduce a paper that formalizes RASQ, an NVFP4 quantization policy that selectivel...
Connecting Claude Code to Self-Hosted Open Models: Inside the free-claude-code Proxy
Coding agents like Claude Code and Codex are tied to the Anthropic API. free-claude-code sits an Anthropic-compatible proxy in between, letting teams keep th...
vLLM v0.25.0: Model Runner V2 Becomes the Default and PagedAttention Is Gone
vLLM v0.25.0 landed with 558 commits from 232 contributors. Two changes define this release: Model Runner V2 is now the default execution path for every dens...
621,500 App Screens, All Somehow the Same App
Browse 621,500 screens and somehow ship the exact same app as everyone else.
The Week Token Prices Were Cut in Half, Inference Chip Valuations Rose Fivefold
When prices fall, do we use less? In the inference market, the opposite is happening. We read the paradox of the token price war and SambaNova’s fivefold val...
Distilling Our Own Skills Into Small Models to Run Them Automatically: The Last 6 Months
Running agent skills exclusively on frontier APIs makes costs explode once you hit thousands of calls. Research from the first half of 2026 has pushed distil...
When Coding Agents Start Talking to Each Other: Designing Multiplayer Claude Code and Collaborative Agents
We are moving from a world where one person uses one agent to one where multiple people and multiple agents share a workspace and talk to each other. Using m...
Less Verification Is Safer: Measuring the Cost-Quality Curve of Multi-Agent Verification Gates
The principle that every fan-out should close with adversarial verification is widely accepted, but how much verifier tier and how many skeptics to use has m...
Agentic RL Stops Waiting for the Group and Learns from One Rollout at a Time
When training long-horizon agentic tasks with RL, GRPO’s group sampling idles the GPU while it waits for the slowest rollout to finish. SAO, the method Tsing...
Remember the Decision, Not the Description: A Meta-Affiliated Study Reframes Agent Memory as a Rate-Distortion Problem
Long-running agents operate within limited memory, yet memory methods to date have organized the past using descriptive criteria such as relevance or summary...
There Was Almost No Q4 Inside Q4_K_M: Dissecting GGUF Quantization Internals
There is a real gap between someone who downloads a GGUF file from Hugging Face and just clicks run, and someone who knows exactly which tensors are stored a...
Testing iOS Apps in a Browser from a Headless Cloud Mac: serve-sim
Put a Mac Mini in the cloud and you lose the GUI, which means you lose the iOS Simulator too. serve-sim streams the simulator’s framebuffer to a browser and ...
The Agent That Studied Its Own Screwups
Fix the one skill you lack and you win. Paxis’s missing skill? Knowing when to stop.
Coding Agents That Remember Through Folders, Not Vector Databases: A Look at personal-monorepo-template
OpenAI Codex engineer jxnl has released personal-monorepo-template, which gives agents persistent memory through folder structure and an AGENTS.md file, no v...
A Cross-Vendor Workflow Where Fable 5 Conducts and Grok 4.5 Implements: fable-advisor
We break down the conductor-worker split of the fable-advisor plugin, in which Claude Fable 5 conducts spec writing and diff review while Grok 4.5 handles th...
LLM Research Ideas Lose on Range, Not Quality
Researchers at Yale and the University of Chicago compared human and LLM research ideas using 11,683 real papers. The conclusion is surprising. The problem w...
Orientation Before Execution: How GPT-5.6 Sol Broke ARC-AGI-3 for the First Time
ARC-AGI-3 measures whether an agent can figure out a situation and adapt on its own inside an interactive game with no instructions. GPT-5.6 Sol became the f...
Back from GRPO to PPO: How GLM-5.2 Stabilized RL with IcePop
The dominant trend in RL post-training these days is the GRPO family, which drops the critic. GLM-5.2, however, went back to PPO with a revived value model a...
The Night the Search Dictionary Fixed Itself: Closing a Skill-Routing Bottleneck with an Unattended Loop
After a prior diagnosis found that the real cause of routing failures in a large skill ecosystem was the retriever rather than decomposition, this study meas...
Grok 4.5 Arrives for Coding and Agents: The Math That Cheap Opus-Class Performance Changes
SpaceXAI has unveiled Grok 4.5. Trained from the ground up for coding and agents, it delivers Opus-class performance at $2 per million input tokens and $6 pe...
21 Million Dollars for One Rack: When AI’s Bill Arrived, the Bottleneck Moved to Power
HBM4 has pushed the price of a single server rack to 21 million dollars, and SK hynix raised 40 trillion won in a single day. Now that capital and power have...
Racks Exploding Upward, Inference Collapsing Downward: Enterprises Stand in the Middle of the AI Scissors
Two numbers released on the same day moved in opposite directions: a 21 million dollar AI rack and inference pricing that got 34 times cheaper. Here is a loo...
From the Intelligence War to the Value War: Enterprises Leaving Frontier APIs and the Economics of Migration
Microsoft has started routing bulk AI requests from Excel and Outlook to its own models, Chinese open models now handle nearly half of some US enterprise AI ...
The Agent You Kept For Someday
The agent you shelved for ‘someday’ takes less time than your coffee to cool.
How a Claude Code Skill Builds Premium Landing Pages in One Shot
A Claude Code skill that turns one plain request into a premium landing page. We take apart how it actually works and validate it from ThakiCloud’s Paxis vie...
Neither the Retriever nor Decomposition: A Ceiling-Gap Diagnostic for Skill Routing Bottlenecks
For a harness that routes hundreds of skills, this diagnostic tells you with two numbers whether your next investment should go into decomposition or the ret...
Memory as an Action Space: AgeMem’s Unified Long- and Short-Term Memory Management
AgeMem folds long-term and short-term memory management into the agent’s policy, exposing store, retrieve, summarize, and discard as tool-based actions. We a...
Opus-Class Performance at a Third of the Price: How Grok 4.5 Is Rewriting Model Economics
SpaceXAI’s newly released Grok 4.5 comes close to Opus 4.8 and GPT-5.5 in performance, at less than half the price. When benchmark gaps shrink to a point or ...
Building Interactive Web Experiences with Fable 5: 3D, Scroll Animation, and GLSL in a Single Prompt
Anthropic’s Claude Fable 5 is setting a new bar for frontend generation. We look at real public guides and open-source galleries showing scroll-driven 3D sce...
An Agent That Never Forgets (Especially the Bill)
Memory became an action space. Pity the only thing it never forgets is the invoice.
Turns Out Every AI Has the Same Brain
If every brain converges, why rent the pricey one?
AI That Can Now See Through Glass, Yet What Companies Really Ask Is ‘What Exactly Did You Just Do?’
A day full of news about AI expanding its senses and gaining hands. But the question that actually decides adoption is shifting from performance to ‘evidence...
Claude Code That Runs Without You: Four Axes of Autonomous Execution
Four mechanisms make Claude Code run without a human watching every step: headless mode, hooks, subagents, and skills. We verify each against ThakiCloud’s re...
Smart Agents Have Become Commonplace. Now Enterprises Are Asking, ‘Prove What You Did.’
On a day when a 1.1 billion parameter model beat a 7 billion parameter one and the best models got given away for free, the news that actually made enterpris...
The Hypothesis That Every AI Is Converging on the Same ‘Brain’: Reading the Platonic Representation Hypothesis
Vision models and language models trained on different data for different objectives are starting to represent data in the same way. MIT’s Platonic Represent...
Binding GPU TEE Attestation to Model Signatures: Attested Confidential Inference for Sovereign Multi-Tenant AI
Hospitals, banks, and government agencies running on-premises AI platforms must prove two things to regulators: that data never leaked outside the enclave, a...
Stop Typing Prompts, Start Designing Loops: Reading Claude Code’s Official Loop Engineering Guide
On July 7, 2026, Anthropic published its first official loop engineering document, ‘Getting started with loops.’ It marks the shift from a human prompting ev...
The Fab Finally Made Money. For Whom, Exactly?
Years of making other people’s chips, finally in the black. Rented upside prints on someone else’s ledger.
An Evaluation That Trusts No Single Number: Building a Sovereign LLM-Judge Service on Binary Decomposition and Deterministic Gates
Using an LLM as a grader, the practice known as LLM-as-a-judge, is now the default in model development, but the evidence that piled up through 2026 shows th...
Claude Code /cd: How to Move Between Directories Without Restarting Your Session and Keep the Prompt Cache Intact
In a monorepo, switching between a library directory and the service that consumes it used to mean restarting your session, and with it, both your conversati...
Setting Up a Claude Code Project the Right Way: Dissecting the .claude/ Folder
Most developers skip the setup and jump straight into prompting. That’s a mistake. We dissect the structure of the .claude/ folder, from CLAUDE.md to rules, ...
This Week’s Must-Read AI Papers (6/28-7/5): Metacognitive RL, Memory as a Skill, MCP Server Patterns
From dair.ai’s weekly paper roundup, we take a deep dive into three papers through a ThakiCloud lens: RLMF, which beats standard RL by up to 63% using metaco...
The Characters the Monitor Cannot Read: How Tokenizer Asymmetry Opens a Covert Channel Around AI Control
The core defense of AI control, the idea that an untrusted strong model can be controlled by a weaker, trusted monitor, rests on the assumption that the moni...
What Comes After AGI: DeepMind’s Four Paths to Superintelligence
Google DeepMind’s roughly 57-page report From AGI to ASI treats superintelligence not as a distant thought experiment but as a planning problem to prepare fo...
The Center of Gravity Has Shifted: Reading the Open Model Ecosystem Through the ATOM Report
The ATOM Report measures open language models across both downloads and inference usage in one place, and shows with data that Chinese open models overtook t...
Is Fine-Tuning Really Dead? A Survival Strategy Read from June 2026’s Verified Signals
As frontier LLMs and agent skills keep improving, the industry has started to feel that fine-tuning is no longer necessary. OpenAI is, in fact, winding down ...
Why an 8x Larger Model Is 5x Cheaper: The Real Structure of LLM Inference Costs
We dissect, with a roofline model, the paradox that the 284B DeepSeek V4 Flash prices its output tokens 5x cheaper than the 35B Qwen3.6. From KV cache reads ...
GLM-5.2 on AMD MI355X at 2,626 tok/s: the serving economics MXFP4 and SGLang built
We examine a case of serving the GLM-5.2 743B MoE model on a single AMD MI355X node at 2,626 tok/s per node, at more than twice the cost efficiency of Blackw...
The Age of Steam Engines, Waiting for Carnot: The Mindset Science Needs Right Now
The gap between a machine that works and a principle we understand is one of the oldest scenes in the history of science. In an era that solves everything wi...
Speccing a Stack Like a Doctor’s Breakfast
A perfectly healthy combo — pity someone else holds the fridge key.
Reading Anthropic’s Prompting Guide: Model-Specific Strategy for Fable 5, Sonnet 5, and Opus 4.8
We work through Anthropic’s official guide to prompting best practices for the latest models. Model differences, the core techniques (clarity, examples, XML,...
Intelligence That Runs at Home: The Personal AI Computer and the Economics of On-Prem Serving
Starting from the Personal AI Computer build guides shared by tom_doerr, which reach up to 384GB of VRAM, we work out through calculation how VRAM decides wh...
Working Without Rate Limits on Fable 5: Model Routing and Token Budget Strategy
We unpack the Claude Fable 5 workflow tips shared by T3 creator Theo: effort levels, Codex orchestration, model priority in CLAUDE.md, and offloading token-h...
How to Build a World-Changing Company
A rocket-CEO’s startup lecture, and the last rule was on-prem all along.
Claude Code Artifacts Come to Pro and Max: Your Session Becomes a Living Web Page
Claude Code artifacts have expanded beyond Team and Enterprise to the Pro and Max plans. We break down the feature that turns a coding session into a live, s...
Build Expensive, Run Cheap: Trimming Skill Costs Every Night
Most agent work doesn’t need a frontier model. Build a skill once with an expensive model, push its format down into code, and you can run the same quality o...
Does splitting vLLM Prefill/Decode across two B200s actually make things faster?
We measured actual tokens-per-second across tensor parallelism, data parallelism, and Prefill/Decode disaggregation (1P1D) for Qwen3.6-27B-NVFP4 and gemma-4-...
Claude Code’s /dataviz Skill: Treating Charts as Design, Not Just Code
The /dataviz skill added in Claude Code 2.1.198 loads chart and dashboard design guidance directly into context. We break down the form heuristic, color form...
Qwen3.6-27B in NVFP4: The Economics of Single-GPU Blackwell Serving
NVIDIA re-quantized Qwen3.6-27B to NVFP4 so it serves on a single Blackwell GPU with vLLM out of the box. We break down how the mixed precision (MLP in NVFP4...
How to Prompt Claude Fable 5: Five Principles from Anthropic’s Official Guide
We break down Anthropic’s official prompting guide for Claude Fable 5. It lays out five principles: strip out instructions written for older models, audit pr...
How to Systematically Learn LLM Internals: From Tokenization to Inference Optimization
If you operate LLMs in production but can’t explain why the KV cache eats memory or what GQA actually saves, your optimization work is running on intuition. ...
My Whole AI Stack Went Chinese
87% cheaper is great. But you just swapped landlords, didn’t you?
Qwen3.6-27B at 4-bit: Why NVFP4 Quantization Came Down to Hopper
NVIDIA’s Qwen3.6-27B-NVFP4 compresses a 27B hybrid-attention reasoning model to 4-bit, cutting memory by roughly 2.5x while keeping benchmark gaps within 1 p...
Dual-Pool Token-Budget Routing — Cutting vLLM Inference GPU Time by 31–42% with Two-Pool Scheduling
Short requests waiting behind long ones in a single queue silently waste GPU time — this is Head-of-Line (HoL) blocking. Dual-Pool Token-Budget Routing, prop...
Five Archetypes That Remain When Job Titles Dissolve: From Prototyper to Maintainer
As engineering, product, design, and data blur into a single mass, Boris Cherny, the creator of Claude Code, proposes five role archetypes and a team-composi...
The Future of Work: Five Archetypes, Not Job Titles
Roles blur into prototyper, builder, sweeper, grower, maintainer, and nobody wants the sixth.
My Whole AI Stack Went Chinese
The day the whole stack belonged to someone else, and Paxis and Metis cope.
The Year Open-Weight Models Caught the Frontier: Why Self-Hosting Economics Now Decide the Outcome
By mid-2026, open-weight models had closed to within a 3-to-6-month capability gap of frontier labs, and that gap is no longer widening. The real decision ha...
Fitting GLM-5.2 469B on a Single Node with NVFP4: A Deep Dive into the vLLM Serving Recipe
The GLM-5.2-NVFP4 checkpoint published by NVIDIA lets you serve a 469B MoE model on a single Blackwell node (8x RTX PRO 6000) with vLLM. We break down the mi...
Editing Video With a Coding Agent: A Look Inside the video-use Skill
Shared by midudev and quickly making the rounds, browser-use’s video-use is a free, open-source skill: drop raw footage into a folder, type one sentence, and...
Inside slime, the Open-Source RL Framework That Built GLM-5.2
slime, the asynchronous reinforcement-learning infrastructure behind the post-training of Z.ai’s 1M-context open-weight model GLM-5.2, has been fully open-so...
Token Usage Explodes, AI Spend Halves: Coinbase’s Better-Defaults Strategy
Coinbase CEO Brian Armstrong’s recipe for controlling AI cost was not usage caps or spend alerts, but better defaults, routing, and caching. Backed by the fi...
Ornith-1.0: An Open Coding Model That Writes Its Own RL Scaffolds
Released by DeepReinforce on June 25, 2026, Ornith-1.0 is an open-source coding model family built on a self-scaffolding mechanism: during reinforcement lear...
NVIDIA GLM-5.2-NVFP4: Serving a 753B Frontier MoE at 4-Bit Precision
NVIDIA has released a NVFP4 (4-bit) quantization of ZAI’s 753B MoE reasoning model GLM-5.2 on Hugging Face. Rather than blanket-quantizing every weight, the ...
AI Agents That Write Papers - Moving Your Writing Workflow to an Open-Source Skill Package
An open-source project bundles a professor’s paper-writing expertise into a Skill package that works across Codex, Claude Code, and Gemini. We examine how it...
vLLM vs Ollama - Local LLM Inference: When to Choose Which
The conventional wisdom that Ollama is easy and vLLM is fast gets re-examined with 2026 RTX 4090 public benchmarks. At a single user the gap is around 13%, b...
Building Your First AI Agent - Start with a Minimal Guarded Loop
Most introductory guides to AI agents cover four pieces: the LLM brain, memory, tools, and the agent loop. What actually breaks in production is the part tha...
The Complete Gemma 4 Lineup: Five Apache 2.0 Open-Weight Models and an On-Premises Serving Guide
Released by Google DeepMind in April 2026, Gemma 4 is a multimodal open-weight family of five models spanning E2B to 31B. We break down the Apache 2.0 licens...
Running Gemma 4 26B 16x Parallel on a Single DGX Spark: NVFP4 Quantization and an Honest Cost-Efficiency Review
NVIDIA’s Gemma-4-26B-A4B-NVFP4 running 16 parallel streams on a single DGX Spark (128 GB unified memory) delivers roughly 18 tokens/s per stream and about 30...
Android Emulators in Containers - Building a Reproducible Device Farm on K8s with docker-android
docker-android packages an Android emulator into a single container and runs it headlessly. We verify image sizes and KVM requirements against the official d...
Making Videos from Claude Code: I Ran claude-code-video-toolkit Myself
An open-source toolkit that renders 1080p video from inside Claude Code with a couple of slash commands. I cloned examples/hello-world and rendered it with z...
The Era of One-Person Cinematic Websites: A Deep Dive into the Gemini 3.1 + Seedance 2.0 Workflow
A 16-minute tutorial published by web designer Viktor Oddy demonstrates how one person can build a cinematic marketing site that once commanded $10,000 in fe...
What We Need to Rethink When Fable 5 Leaves the Subscription
Anthropic is including Fable 5 in subscription plans at no extra cost only through June 22, and from June 23 it shifts to pay-per-use credits. ThakiCloud exa...
Three Giants, Three Futures: What Hassabis, Huang, and Amodei Leave for Builder Organizations
A comparison of the worldviews of Hassabis, Huang, and Amodei, who read the same AI wave through three different lenses of AGI, infrastructure, and labor, an...
The Age of Agent-Managed Organizations: Jensen Huang’s Vision and the Reshaping of Organizational Culture
When NVIDIA CEO Jensen Huang’s vision of ‘hundreds of agents per engineer’ becomes reality, how must organizational structures and ways of working change?
AGI Is Still Far Off: Demis Hassabis’s Realism and the Culture of Expectation Management
Starting from DeepMind CEO Hassabis’s statement that we are ‘nowhere near AGI,’ this post explores why AI organizations should make honest expectation-settin...
If AI Were the Best Decision-Maker: A Culture Where Data Beats Intuition
Starting from a market-coined label about Jensen Huang, and deepening the Moneyball legacy, how to root a data-beats-intuition decision culture inside your o...
When Software Becomes Free, What Do Engineers Do? Dario Amodei’s Prophecy
When the marginal cost of code approaches zero, the value of an engineering team shifts from ‘what we build’ to ‘knowing what should be built.’
Kimi K2.6: 1T MoE, 32B Active, 300-Subagent Swarm Architecture Analysis
Moonshot AI’s Kimi K2.6 is a MoE model with 1T total parameters but only 32B active per token, maintaining dense 32B inference costs while supporting 256K co...
GLM-5.2: 753B MoE, 1M Context, MIT License Self-Hosting Guide
Z.ai’s GLM-5.2 handles 1M context with 2.9x FLOPs savings using DSA (Dynamic Sparse Attention). The MIT license removes barriers to on-premises deployment, a...
FastContext-1.0-4B-SFT: A 4B Subagent Model Dedicated to Repository Exploration for Coding Agents
Microsoft released FastContext-1.0-4B-SFT, a fine-tuned Qwen3-4B coding agent subagent model. It explores repositories using only three tools, READ, GLOB, an...
DiffusionGemma 26B-A4B: Google’s Experiment Generating 15-20 Tokens at Once with Discrete Text Diffusion
Google DeepMind released diffusiongemma-26B-A4B-it, a MoE-based VLM that generates text via discrete diffusion rather than autoregressive decoding. 25.2B tot...
Vibe-Coding-Instruct: SFT for Lightweight Code Agents Using 1.1 Million Samples
Vibe-Coding-Instruct by lazarus19 is an Apache-2.0 dataset of 1.1 million coding instruction-response pairs. It has already been used to train 7 Gemma/Qwen d...
Fable-5-traces: How to Distill Small Models Using Coding Agent Session Traces
Glint-Research released 4,665 Fable 5 (Claude Code) agent traces in AGPL-3.0 under the HF Agent Traces format, with 81% tool-use composition. A practical gui...
Agents Last Exam: Why Computer-Use Agent Evaluation Has Evolved to 153 Long-Horizon Tasks
agents-last-exam is a benchmark dataset of 153 long-horizon tasks for evaluating computer-use agents. Five domains including Business, Computing, Engineering...
Goclone: Clone Any Website to Your Computer in Seconds
Learn how to use Goclone, a powerful Go-based website cloner that downloads entire websites including HTML, CSS, JavaScript, and images to your local machine.
RAGLight Complete Guide: From Basic RAG to Agentic Workflows
Master RAGLight framework with hands-on examples covering RAG, Agentic RAG, RAT pipelines, and MCP integration for building powerful retrieval-augmented gene...
LangGPT: Master Structured Prompt Engineering Framework for Better AI Interactions
Learn how to create high-quality, reusable prompts using LangGPT’s structured framework. Transform chaotic prompt engineering into systematic methodology wit...
udocker: Complete Guide to Rootless Docker Container Execution
Learn how to run Docker containers without root privileges using udocker - perfect for HPC environments, shared systems, and secure container execution.
Shannon AI Agent Orchestrator: Complete Tutorial for Enterprise-Grade AI Agent Management
Learn how to set up and use Shannon, an open-source AI agent orchestrator with enterprise-grade security, cost controls, and vendor flexibility. A comprehens...
Helm Dashboard: Complete Guide to Kubernetes Helm Charts UI Management
A comprehensive tutorial on Helm Dashboard - the missing UI for Helm that simplifies Kubernetes chart management with visual interface, revision history, and...
AI Coding Assistant: Complete Guide to Boost Your Development Productivity
Master AI-powered coding tools like GitHub Copilot, ChatGPT, and Claude to accelerate your development workflow and write better code faster.
Designing Agentic Loops: Complete Tutorial for AI Coding Agents
Master the art of designing effective agentic loops for AI coding agents. Learn safety practices, tool selection, and real-world implementation strategies.
Context Engineering: The Complete Guide to AI Coding Assistant Mastery
Master Context Engineering - the revolutionary approach that’s 10x better than prompt engineering and 100x better than vibe coding. Learn how to make AI codi...
Complete Guide to LandingAI Agentic Document Extraction: AI-Powered PDF and Image Processing
Master LandingAI’s Agentic Document Extraction library for intelligent document processing. Extract structured data from complex PDFs, images, and documents ...
Beyond Technical Skills: A Strategic Approach to Career Growth
While technical excellence is the foundation of any career, true advancement requires mastering four essential disciplines: technical skill, product thinking...
Comprehensive Guide to LLM Dataset Curation: From Training to Preference Alignment
Explore the essential datasets and tools for LLM post-training, including supervised fine-tuning datasets, preference alignment data, and curation methodolog...
Awesome Public Datasets: Your Gateway to High-Quality Open Data
Discover the ultimate collection of curated public datasets across diverse domains, from agriculture to eSports, maintained by the global open data community.
NVIDIA Nemotron 6 Million Multilingual Reasoning Dataset Released – Strengthening the Open-Source AI Ecosystem
NVIDIA releases a 6-million-example multilingual reasoning dataset, providing high-quality training data expanded across five languages: French, Spanish, Ger...
NVIDIA Granary: A Large-Scale Multilingual Speech Dataset Spanning 25 European Languages
NVIDIA’s latest multilingual speech recognition and translation dataset, Granary, covers 640,000 hours of audio across 25 European languages. This guide cove...
Rowfill: Complete Guide to Unstructured Data Processing Platform for Knowledge Workers
Discover the core features and applications of Rowfill, an open-source AI platform that automatically structures PDF, image, and audio files.
AnyCrawl: A Comprehensive Guide to the LLM-Ready Web Crawler - A New Standard for AI Data Collection
Master how to transform websites into LLM-ready data and efficiently collect Google/Bing SERP results with AnyCrawl, built on Node.js/TypeScript.
ByteDance Dolphin Document Image Parsing: Fox Dataset and Benchmark Complete Analysis
A detailed analysis of ByteDance’s Dolphin project Fox dataset and benchmark, including the Analyze-then-Parse paradigm from ACL 2025 and a large-scale datas...
dots.ocr: SOTA Multilingual Document Parsing with 1.7B Parameters - Complete Analysis
How to implement multilingual document layout analysis and OCR in a single vision-language model using dots.ocr, released by RedNote.
Essential Skills Guide for ML Engineers - A Must-Have Checklist for Recruiters and Candidates
A practical guide outlining the core technology stack and competencies needed to build ML applications in production environments
AI Coding Era Cultural Innovation: Paradigm Shift from Resistance to Collaboration
How to embrace and develop the new development culture brought by Vibe Coding and Agentic Coding? A guide to building collaborative culture with AI, breaking...
Quantifying Team Chemistry with Saberr: Data-Driven Hiring and Organizational Strategy
How to leverage Saberr algorithms to quantify team compatibility through 15-minute surveys and behavioral data, optimizing everything from hiring to onboarding
Building Data-Driven Organizational Culture with Moneyball Thinking
How to apply Moneyball strategy that discovers hidden value through data and achieves maximum performance relative to resources in development, product, and ...
Hacking and Reverse Engineering: A New Methodology for AI Vibe Coding Era
In the AI era, developers don’t need to know everything. Explore a new paradigm that transforms ignorance into strength through hacking mindset and reverse e...
AI Engineering Learning Roadmap
We’re sharing a learning roadmap for those who want to start AI engineering. This roadmap is structured to allow step-by-step learning of the core areas of A...
Backend·Infrastructure Engineer Hiring: 10 Must-Read Books and the People We’re Looking For
Introducing the ideal candidate profile and hiring criteria through 10 must-read books for backend·infrastructure engineer recruitment and practical applicat...
Three Vs-Driven MLOps at ThakiCloud 🚀 — Why You’ll Want to Join Us
ThakiCloud’s Three Vs (Velocity, Validation, Versioning) based MLOps culture and practical cases, plus recruitment information for colleagues to join us.
[Thaki Cloud Life & Career] KCD Seoul 2025
Sharing materials presented at KCD Seoul 2025. Content about Thaki Cloud, an xPU as a Service-based Agentic AI platform
Thaki Cloud Life & Career Guide
Sharing Thaki Cloud’s corporate culture, benefits, developer stories, recruitment information, and more.
[Thaki Cloud Life & Career] Mission, Principles, Values
Sharing Thaki Cloud’s mission, principles, and values.