Aleksa Gordić - The AI Epiphany
LLM architectures, training infrastructure, and open-source models with a focus on engineering trade-offs.
Nutrition Label
This channel provides high-signal technical interviews with primary authors of major AI models and tools, often diving into architectural nuances like kernels and training infrastructure. While most content is rigorously engineering-focused, occasional product demos serve more as friendly showcases than critical reviews.
Strengths
- +
- +
- +
Notes
- !Interviews with researchers offer deep technical value, while tool demos tend to be less critical.
- !Check video descriptions for clear disclosures on sponsorships or the creator's own startup ventures.
Rating Breakdown
Breakdown across the key dimensions we rate. Methodology →
Recent Videos
Archie: an engineering AGI for Dyson Spheres | P-1 AI | $23 million seed round
Build AI Agents using Integrail (Halloween special)
Cracked Engineers - Find Tech Jobs / Hire Tech Talent
Imbue - training a 70B model from scratch! (w/ Bowei - head of infra)
LLaMA 3 Deep Dive! (Thomas Scialom - Meta)
Best LLM? Qwen 2 LLM w/ author Junyang Lin
State Space Models w/ Albert Gu & Karan Goel (Cartesia AI)
Fine-tune LLMs 30x faster! With Daniel Han (Unsloth AI)
Hamel Husain - Building LLM Apps in Production
Building Julius AI to 500.000 users w/ Rahul (founder)
DeepMind's TacticAI: an AI assistant for football tactics | Petar Veličković
Ishan Misra (Meta) - Emu Video Generation
InstructPix2Pix (w/ OpenAI's Tim Brooks)
GPT-Fast - blazingly fast inference with PyTorch (w/ Horace He)
How does Groq LPU work? (w/ Head of Silicon Igor Arsovski!)
Why this rating
Evidence receipts showing why each dimension is rated the way it is.
“I can finally share what I've been up to since last summer! We just raised a $23 million seed round!! ... I co-founded P-1 AI”[Description] →
The creator explicitly discloses that he is the co-founder of the company being discussed, making the material connection the central theme of the video.
“Let's say I want to find a job in computer vision... I can just click here and it's going to filter out only computer vision jobs.”[01:15] →
Demonstrates the tool live in real-time, clicking through filters and showing immediate results rather than relying on slides or theoretical descriptions.
“For Qwen 2, we actually use GQA [Grouped Query Attention] for all the models... previously for Qwen 1.5 we only used GQA for 32B... we want to enjoy the benefits of faster inference speed.”[31:15] →
Guest details specific architectural decisions and trade-offs (inference speed vs parameter count) rather than generic specs.