Category: Machine Learning

7 Min Read

onOctober 15, 2026

Nobody Taught Models to Absorb Criticism Because Nobody Measures It: SycEval Says 58%, Defensiveness Says Nothing

We can score a model’s over-agreement to three decimal places and have no public number for a model refusing to listen. That asymmetry…

7 Min Read

Artur Markus

onOctober 14, 2026

Agent Lightning v1.0 in Practice: What 3,500 Lines of Agentic RL Buy You, and What They Do Not

Can you train your existing LangGraph or Claude Code agent with reinforcement learning without rewriting it as an RL environment? Microsoft…

7 Min Read

Artur Markus

onOctober 2, 2026

Decontaminating Benchmarks Is a Fantasy: Ten Detectors Drop to Coin Flip After One Round of GRPO

The sentence “we ran contamination checks and found nothing” now carries about as much information as a coin flip. Ten detection…

6 Min Read

Artur Markus

onSeptember 17, 2026

New Research: Agentic Reasoning Burns Up to 136.5× More GPU Energy Per Query Than a Single-Turn LLM Call

348.41 Wh. That is what one Reflexion agent on Llama-3.1-Instruct 70B spent answering a single HotpotQA question, 136.5 times the energy of…

7 Min Read

Artur Markus

onSeptember 9, 2026

AI Model Leaderboards: Myth vs Reality, 2 Votes Flipped #1 and 78% of Codeforces Was in Training

“Dropping just a handful of preferences can change top large language model rankings.” That is the title of an ICLR 2026 poster…

6 Min Read

Artur Markus

onSeptember 4, 2026

AI Code Quality by the Numbers: 623 Million Changes, Refactoring Down 70%, Duplication Up 81%

Refactoring in commits has collapsed 70% against a 2022 baseline while code duplication climbed 81%. That comes from 623 million code changes,…

8 Min Read

Artur Markus

onSeptember 1, 2026

The Harness, Explained: Why the Same Scaffold Scores 65% or 74% on SWE-bench

A 100-line Python harness scored 65% on SWE-bench Verified in July 2025. The same project now claims above 74% — and boots faster than Claude…

6 Min Read

Artur Markus

onAugust 28, 2026

172 Billion Tokens Show Every LLM Fabricates Above 10% at 200K Context

The best of 35 open-weight models still invented answers 1.19% of the time about entities that provably did not exist in the source document.…

10 Min Read

Artur Markus

onAugust 12, 2026

NVIDIA Nemotron 3.5 Lightning Launches at 1,200 Tokens/Second—30B MoE Outpaces Gemma 4 by 29× and Qwen 3.6 by 35% on Agent Tasks

NVIDIA just shipped a 30B parameter model that outputs 1,200 tokens per second—29 times faster than Gemma 4 26B on identical hardware. This…

11 Min Read

Artur Markus

onJuly 25, 2026

Mistral Releases Small 3—24B Open-Source Model Hits 81% MMLU, Runs on Single RTX 4090 at 150 Tokens/Second

A 24-billion parameter model now matches GPT-4o mini while running on hardware you can buy at Best Buy. Mistral just made the $400 billion…

What are You Looking For?

Machine Learning

Nobody Taught Models to Absorb Criticism Because Nobody Measures It: SycEval Says 58%, Defensiveness Says Nothing

Agent Lightning v1.0 in Practice: What 3,500 Lines of Agentic RL Buy You, and What They Do Not

Decontaminating Benchmarks Is a Fantasy: Ten Detectors Drop to Coin Flip After One Round of GRPO

New Research: Agentic Reasoning Burns Up to 136.5× More GPU Energy Per Query Than a Single-Turn LLM Call

AI Model Leaderboards: Myth vs Reality, 2 Votes Flipped #1 and 78% of Codeforces Was in Training

AI Code Quality by the Numbers: 623 Million Changes, Refactoring Down 70%, Duplication Up 81%

The Harness, Explained: Why the Same Scaffold Scores 65% or 74% on SWE-bench

172 Billion Tokens Show Every LLM Fabricates Above 10% at 200K Context

NVIDIA Nemotron 3.5 Lightning Launches at 1,200 Tokens/Second—30B MoE Outpaces Gemma 4 by 29× and Qwen 3.6 by 35% on Agent Tasks

Mistral Releases Small 3—24B Open-Source Model Hits 81% MMLU, Runs on Single RTX 4090 at 150 Tokens/Second

Install AI Unfiltered