From Fine‑tuned CNNs to ChatGPT’s Pixel‑perfect Descriptions
It do be like that these days... 😅 I remember when we had to fine-tune an ImageNet-pre-trained CNN and preprocess the image to then send it and get a top-5 classification that was somewhat accurate... Now ChatGPT describes every pixel of it accurately and spots things better than I do 🙃
Thinking Less Can Outperform Chain‑of‑Thought Prompts
Last week, I had the chance to attend the Conference on Language Modeling (COLM 2025) here in Montreal - a must-attend conference for anyone starting out in AI/LLM I was lucky to get an invite, it was awesome. Almost every session circled...
This Week in AI Engineering
This week’s AI engineering roundup highlights three developments reshaping developer workflows and compute economics: OpenAI launched Agent Builder, a no-code drag-and-drop tool that lets non-developers assemble multi-agent workflows alongside ChatKit, Evals and reinforcement fine-tuning; NVIDIA unveiled the DGX Spark, a...
Claude Haiku 4.5 Just Matched Sonnet 4... At 2x Speed!
Anthropic has launched Cloud Haiku 4.5, a lighter-weight variant of its Sonnet model that the presenter says matches Sonnet’s performance while running twice as fast and costing one-third as much. The update claims Haiku 4.5 is marginally superior to GPT-5...
Qwen3-VL Just Changed Multimodal AI (Again) 🔥
OpenAI competitor Qwen released two compact vision-language models, Qwen-VL 4B and 8B, that pack multimodal capabilities into highly efficient, small architectures. They support FP8 for lower-precision inference, offer both dense and Mixture-of-Experts (MoE) variants, and expand language coverage to 32...
Petaflop Desktop AI: Powerful Yet Hybrid Future
A petaflop on your desk. Let that sink in. NVIDIA’s new DGX Spark squeezes what used to be an entire data center into desktop size — Grace Blackwell chips, 128GB unified memory, and a full AI stack out of the box....
DGX Spark vs Cloud: Who Wins for AI Work?
NVIDIA’s DGX Spark is being touted as the world’s smallest portable AI supercomputer, packing up to 1 petaflop of compute, 128GB of memory and the capacity to train ~70B-parameter models or run inference on models up to 200B parameters (two...
More Thinking. Worse Results. Wait, What?
From face recognition to grammar learning, both humans and models perform better when they don’t reason step-by-step. Here’s the paradox.
Fine-Tuning Explained in 60 Seconds (No Math!)
Fine-tuning adjusts a pre-trained language model’s billions of parameters to make it specialize on a specific task or domain rather than teach it entirely new knowledge. Instead of full retraining—costly in compute—practitioners often tune small parameter subsets using methods like...
ChatKit + Agent Builder = Instant AI Apps (No Coding Needed)
OpenAI has launched two no-code tools—Agent Builder and ChatKit—allowing users to assemble agentic workflows and embed chatbots via drag-and-drop interfaces without programming. Agent Builder supports complex multi-step agents and integrations, primarily using OpenAI models but permitting external models for evaluation,...
Show End‑to‑End AI Projects, Not Just Buzzwords
In interviews, most candidates talk about AI. But few build something that truly stands out. After reviewing hundreds of projects and interviewing candidates, here’s what actually makes the difference. 👇 1️⃣ Complete end-to-end projects A strong portfolio project isn’t just a chatbot or a...