Wayne (@waynoir) header
Wayne (@waynoir) profile picture

building systems from scratch, writing about llms and agents along the way

Joined February 2026
Photo by Wayne (@waynoir)Photo by Wayne (@waynoir)Photo by Wayne (@waynoir)Photo by Wayne (@waynoir)Photo by Wayne (@waynoir)Photo by Wayne (@waynoir)Photo by Wayne (@waynoir)
Pinned Tweet
Higgsfield just dropped a full tutorial on building a faceless YouTube channel with Claude Fable 5.2 and Higgsfield MCP Everything it walks through, from picking a niche to hitting 4,000 watch hours: 00:00 How faceless channels make $10K a month 00:42 Step 1: Which faceless niche pays the most 02:48 Shorts or long videos: which is better to monetize? 03:22 Step 2: Is YouTube demonetizing AI channels? 04:14 Step 3: How to find topics that get millions of views 05:01 How to write a script with Claude Fable 5.2 06:18 How to generate a faceless video with AI 08:05 Our AI documentary 08:43 How to fix a bad AI shot without re-rendering 09:54 How to make thumbnails with AI 10:47 How to upload AI videos and hit 4,000 hours 11:53 How to make faceless Shorts with AI 14:37 The truth about viral channels 15:22 Start your faceless channel today The video claims these channels make over $10,000 a month. Watch it, then read 8 more ways to make money with Fable 5.2
Call me crazy... fable 5.2 feels like AGI One person with a laptop can now ship work that used to need a small team I wrote down 8 ways to turn that into money, each with who pays, how to do it and your first step
1
1
7
363
Wayne retweeted
Higgsfield just dropped a full tutorial on building a faceless YouTube channel with Claude Fable 5.2 and Higgsfield MCP Everything it walks through, from picking a niche to hitting 4,000 watch hours: 00:00 How faceless channels make $10K a month 00:42 Step 1: Which faceless niche pays the most 02:48 Shorts or long videos: which is better to monetize? 03:22 Step 2: Is YouTube demonetizing AI channels? 04:14 Step 3: How to find topics that get millions of views 05:01 How to write a script with Claude Fable 5.2 06:18 How to generate a faceless video with AI 08:05 Our AI documentary 08:43 How to fix a bad AI shot without re-rendering 09:54 How to make thumbnails with AI 10:47 How to upload AI videos and hit 4,000 hours 11:53 How to make faceless Shorts with AI 14:37 The truth about viral channels 15:22 Start your faceless channel today The video claims these channels make over $10,000 a month. Watch it, then read 8 more ways to make money with Fable 5.2
Call me crazy... fable 5.2 feels like AGI One person with a laptop can now ship work that used to need a small team I wrote down 8 ways to turn that into money, each with who pays, how to do it and your first step
1
1
7
363
Wayne retweeted
Someone just built a website that could sell for $50,000 with Claude Fable 5.2 He did it in just 8 minutes. Here are more ways to make money with it
Call me crazy... fable 5.2 feels like AGI One person with a laptop can now ship work that used to need a small team I wrote down 8 ways to turn that into money, each with who pays, how to do it and your first step
1
5
354
Someone just built a website that could sell for $50,000 with Claude Fable 5.2 He did it in just 8 minutes. Here are more ways to make money with it
Call me crazy... fable 5.2 feels like AGI One person with a laptop can now ship work that used to need a small team I wrote down 8 ways to turn that into money, each with who pays, how to do it and your first step
1
5
354
Wayne retweeted
Call me crazy... fable 5.2 feels like AGI One person with a laptop can now ship work that used to need a small team I wrote down 8 ways to turn that into money, each with who pays, how to do it and your first step
3
3
1
12
1,818
Call me crazy... fable 5.2 feels like AGI One person with a laptop can now ship work that used to need a small team I wrote down 8 ways to turn that into money, each with who pays, how to do it and your first step
3
3
1
12
1,818
Wayne retweeted
Jensen Huang Says He's a Big Fan of "Chinese Models"
3
1
7
406
Jensen Huang Says He's a Big Fan of "Chinese Models"
3
1
7
406
Wayne retweeted
What if AI could learn continuously, just like the human brain? The Dragon Hatchling (BDH) is a brain-inspired, post-transformer architecture developed by Pathway that takes a different approach to how AI processes and remembers information Instead of relying on dense self-attention, BDH represents computation as a sparse network of neuron-like nodes connected through weighted synapses Here's how it works: > Sparse neural networks: Information flows between connected nodes rather than being processed through dense attention. > Hebbian learning: Connections between simultaneously active nodes become stronger, allowing the system to reinforce useful patterns. > Competitive selection: Only the strongest activations survive, keeping computation efficient. > Dynamic memory: Unlike traditional transformers, BDH can update its synaptic state while processing new information. The architecture aims to achieve linear-time inference, bounded memory usage and recurrent reasoning. The bigger idea is interesting, instead of building AI systems that simply process information using fixed weights, can we build systems that continuously learn, adapt, and develop memories as they interact with new information? BDH is an interesting step toward exploring that possibility
1
1
7
221
What if AI could learn continuously, just like the human brain? The Dragon Hatchling (BDH) is a brain-inspired, post-transformer architecture developed by Pathway that takes a different approach to how AI processes and remembers information Instead of relying on dense self-attention, BDH represents computation as a sparse network of neuron-like nodes connected through weighted synapses Here's how it works: > Sparse neural networks: Information flows between connected nodes rather than being processed through dense attention. > Hebbian learning: Connections between simultaneously active nodes become stronger, allowing the system to reinforce useful patterns. > Competitive selection: Only the strongest activations survive, keeping computation efficient. > Dynamic memory: Unlike traditional transformers, BDH can update its synaptic state while processing new information. The architecture aims to achieve linear-time inference, bounded memory usage and recurrent reasoning. The bigger idea is interesting, instead of building AI systems that simply process information using fixed weights, can we build systems that continuously learn, adapt, and develop memories as they interact with new information? BDH is an interesting step toward exploring that possibility
1
1
7
221
Wayne retweeted
How can we make LLMs think deeper without adding more parameters? If you take a transformer block and stack it 20 times, you get the popular Transformer architecture When generating tokens, it repeatedly processes previous tokens. To avoid recomputing everything, it stores previous attention keys and values in a KV cache, making autoregressive generation faster But the KV cache grows with every token. So, you replace much of the attention mechanism with linear attention Instead of storing the entire history, this compresses it into a fixed size summary, giving us a Transformer hybrid Now context is cheaper, but each token still gets only as many processing steps as there are layers You could stack 1,000 layers, but every layer adds more weights, making memory requirements explode So what if you reused the same layer multiple times? You take a layer's output and feed it back into that same layer, over and over Each pass requires more compute, but no additional weights That's a looped Transformer, an architectural approach used in models like GPT-6 Astra An idea dating back to 2018 is now one of the interesting directions in modern LLM architecture
1
2
9
180
How can we make LLMs think deeper without adding more parameters? If you take a transformer block and stack it 20 times, you get the popular Transformer architecture When generating tokens, it repeatedly processes previous tokens. To avoid recomputing everything, it stores previous attention keys and values in a KV cache, making autoregressive generation faster But the KV cache grows with every token. So, you replace much of the attention mechanism with linear attention Instead of storing the entire history, this compresses it into a fixed size summary, giving us a Transformer hybrid Now context is cheaper, but each token still gets only as many processing steps as there are layers You could stack 1,000 layers, but every layer adds more weights, making memory requirements explode So what if you reused the same layer multiple times? You take a layer's output and feed it back into that same layer, over and over Each pass requires more compute, but no additional weights That's a looped Transformer, an architectural approach used in models like GPT-6 Astra An idea dating back to 2018 is now one of the interesting directions in modern LLM architecture
1
2
9
180
Wayne retweeted
Attention is just 96 layers repeating the same idea over and over, build a query, check it against every key and grab the values that match best. That's basically what a frontier model is made of underneath The video breaks down the full pipeline covering queries, keys, values, attention scores and how the encoder blocks fit together If you want a more in depth structure for becoming an AI engineer, this article is worth a read
4
1
13
427
Attention is just 96 layers repeating the same idea over and over, build a query, check it against every key and grab the values that match best. That's basically what a frontier model is made of underneath The video breaks down the full pipeline covering queries, keys, values, attention scores and how the encoder blocks fit together If you want a more in depth structure for becoming an AI engineer, this article is worth a read
4
1
13
427
Wayne retweeted
Sam Altman on starting up today: If you're graduating now, you're lucky. This is the best time in history to build something A single founder with the right tools can build a billion-dollar company. What once took hundreds of people now takes one person who knows how to use AI
3
1
8
396
Sam Altman on starting up today: If you're graduating now, you're lucky. This is the best time in history to build something A single founder with the right tools can build a billion-dollar company. What once took hundreds of people now takes one person who knows how to use AI
3
1
8
396
Wayne retweeted
Don't waste 2 years learning to build LLMs like Claude & ChatGPT Stanford just dropped a 2 hour 34 min course on how to build LLMs from scratch 00:00 - LLM tokenization 25:42 - how LLMs decode user prompts 35:38 - training pipeline of LLMs 1:16:45 - LLM architecture from scratch Anthropic pays $750,000/year to engineers who understand this exact knowledge of LLMs Watch this, and read the following Roadmap article to become AI Engineer
5
11
86
7,417
Don't waste 2 years learning to build LLMs like Claude & ChatGPT Stanford just dropped a 2 hour 34 min course on how to build LLMs from scratch 00:00 - LLM tokenization 25:42 - how LLMs decode user prompts 35:38 - training pipeline of LLMs 1:16:45 - LLM architecture from scratch Anthropic pays $750,000/year to engineers who understand this exact knowledge of LLMs Watch this, and read the following Roadmap article to become AI Engineer
5
11
86
7,417
Wayne retweeted
"You just have to pick the things you care about and put in 10,000 hours. It doesn't matter where you start. There's a determinism to becoming an expert at something" Andrej Karpathy's advice for beginners getting into AI/ML
4
22
147
15,440