Shadman Ahmed
Software Architect
Software architect who runs AI Bytes. Articles here are produced with AI assistance from the sources they cite; he oversees the site and handles corrections.
32
Articles
39,887
Total Views
54K
Words Published
All Articles (32 total)
Google Opens Lyria 3 API: AI Music for 4 Cents a Track
Google Lyria 3 is now available to developers through the Gemini API at $0.04 per 30-second clip. Here's what you get, what's missing, and how it stacks up against Suno and Udio.
Clarity-OMR vs Audiveris: 5 OMR Accuracy Tests
A deep-dive comparison of Clarity-OMR's machine learning approach against Audiveris's traditional computer vision for optical music recognition — with real benchmark data on 10 classical piano pieces.
OpenAI Sora 2 Safety Policy: Watermarks, C2PA, and 3 Gaps
OpenAI details its five-layer safety system for Sora 2, including C2PA metadata, CSAM detection, and teen protections. But real-world testing reveals stubborn blind spots that watermarks and classifiers can't fix.
ROCm vs Vulkan Performance: Mi50 Benchmark (4 Models)
New benchmarks pit ROCm 7 nightly against Vulkan on an AMD Mi50 32GB running llama.cpp. Vulkan wins short-context dense inference, but ROCm dominates everything else — with a stability catch.
CRYSTAL Benchmark Exposes How AI Models Fake Reasoning
A new benchmark tested 20 multimodal AI models and found 19 of them cherry-pick reasoning steps while skipping actual thinking. The gap between accuracy and reasoning quality is alarming.
Project Genie Prompts: 4 Tips to Build Better Worlds
Google DeepMind's Project Genie lets you generate interactive worlds from text. Here are 4 proven tips for writing prompts that produce stunning, explorable environments.
Google's $12.5M AI Open Source Security Push, Explained
Google, Microsoft, OpenAI, and Anthropic are pooling $12.5 million to secure open source software — and Google's AI tools Big Sleep and CodeMender are already finding and fixing real vulnerabilities.
Qwen3.5-9B vs GPT: Document OCR Benchmarks & Weak Spot
Benchmark data shows Qwen3.5-9B beats frontier models on OCR and field extraction, yet stumbles badly on tables. Here's the honest breakdown.