Sherry Brown

Eyes on the market. Hands on the deals. 👁 Scouting for Lua Ventures. We launched a channel for behind-the-scenes alpha. Join Scouts TG: 👉 t.me/LuaScouts Resources: ▪️ Site: luaventures.xyz ▪️ Main TG: t.me/luaventures

Buying quality stocks at the 40-month moving average is enough to beat most strategies That approach beats the S&P 500 and most portfolio managers by a wide margin. The problem is nobody wants to buy when there’s chaos. We are wired for comfort. - Prof

2026 is your turn to become a millionaire… Last year there was many generational opportunities like: ~ $AMD at $76 & ran to $267+ ~ $TSLA at $210 & ran to $499+ ~ $HOOD at $29 & ran to $153+ This time around there’s 5 clear sectors that’ll dominate: Space ~ $ASTS, $RKLB, $LUNR AI ~ $NBIS, $IREN, $CIFR, $ZETA Drones ~ $ONDS, $AVAV, $KTOS Robotics ~ $TSLA, $PATH, $RR Energy ~ $EOSE, $OKLO These five themes alone will create generational wealth. Save this post for later. You’ll look back at the end of this year, & be thankful you listened…

These 8 software stocks are trading at their lowest valuations in a decade: 1. Atlassian $TEAM

BREAKING: Google Gemini released a new feature called Guided Learning. You can now use it to learn literally anything, step by step, like a personal tutor. Here’s how to access it 👇

Claude Code idea: Sub-Agent Context Negotiation Sub-agents exist to save context. They fetch information and return a summary instead of dumping everything into the main window. But there's a flaw. The orchestrator has semantic context the sub-agent lacks. It knows the purpose behind the request. The sub-agent only knows the literal query. So summaries often miss key details. To give an analogy: Your boss sends you to a meeting and asks for a summary. You come back and give him the rundown. Nine times out of ten, he's going to have follow-up questions. Your summary won't include everything he needs because you don't have the implicit context he has. This is just how humans communicate. So why not implement the same protocol between sub-agents and orchestrators? The fix: a skill that makes the orchestrator evaluate every sub-agent return and ask follow up questions before accepting it. The sub-agent goes back to the source, gets answers, returns. Loop until sufficient. The result: an orchestrator with far more relevant context than a single pass would ever provide. Humans accumulate critical context by asking questions. Orchestrators should too. Here's a draft of a skill file I quickly whipped up to showcase what it might look like. Haven't tested it yet, but I most certainly will!

BREAKING 🚨: Notion is working on a big set of new AI features for Notion Agents. - Custom MCP support - New agent integrations with Linear and Ramp - Notion Mail and Notion Calendar triggers for custom agents - Custom workers (tools) for agents - Custom Connectors - New Library and Feed tabs - AI Co-editor - And Computer Use for AI agents! 👀

AgentHome: a home command center where AI agents manage your property like you have a dedicated estate manager. Watch them auto-pay your mortgage, schedule the plumber before the leak gets worse, negotiate with pest control, water your garden based on tomorrow's forecast, and file permits for your renovation — all while talking to each other in real-time.

⚠️China is still DUMPING US Treasuries: China’s holdings of US government bonds dropped -$6.1 BILLION in November, to $682.6 billion, the lowest since September 2008, the midst of the Financial Crisis. China has now sold -$76.4 BILLION of Treasuries since the start of 2025. Since the 2013 peak, the country's Treasury holdings have fallen -$634.1 BILLION, nearly half the total. Furthermore, there is growing evidence that we can’t track China’s US Treasury purchases through Belgium anymore.

New NanoGPT Speedrun WR at 99.3s (-5.6s) with a bigram hash embedding that is added to the residual stream before every layer. Inspiration from Svenstrup et al 2017 paper on Hash Embeddings, and Deepseek's Engram. Modded-NanoGPT now uses fewer training tokens than its parameter count, a radical divergence from the 20x Chinchilla ratio. github.com/KellerJordan/m…

I’ve spent 10 years teaching math to machine learning engineers. 80% of university math is irrelevant to your actual job. Luckily, I've created a FREE roadmap to teach you the 20% you actually need. Like, retweet, and comment "roadmap" and I'll DM you the link.

MIT Advanced Data Structures course: courses.csail.mit.edu/6.851/spring21/ Video lectures, lecture notes, problem sets, GitHub, and more

After Dropping RoPE completely... We got yet another RoPE upgrade from Sakana AI So as RoPE gives every token a fixed number (1,2,3,…) so attention is based on literal distance in the sequence... This paper introduces RePo, where it keeps the same “rotation” trick but lets the model choose those numbers (which are also continuous and learned) This enables it to pull relevant far-away tokens closer and push irrelevant stretches farther apart, where RePo is able to allocate higher attention to distant but relevant info

🔴China's household lending is COLLAPSING: New loans to Chinese households plunged to just 441.7 billion yuan in 2025, the lowest since 2005. This represents just 16% of the amount issued in 2024, marking a -84% DROP. By comparison, new household loans peaked at over 7.9 trillion yuan in 2021. This signals severe weakness in consumer confidence and reflects the ongoing property market crisis, as Chinese households remain reluctant to take on new debt. China's economy is struggling.

Google Ironwood TPU Memory Hierarchy in 9 levels by hand ✍️ 1. Bit – The most basic unit of information, the on–off decision from which every number, tensor, and model state is ultimately constructed. 2. FP8 (1×8 → 8 bits) – Eight bits are grouped to form a floating-point value, typically used for inference, where reduced precision is a deliberate trade-off to maximize throughput and efficiency. 3. BF16 (×2 → 16 bits) – Two FP8-scale chunks are combined to gain more dynamic range and stability, while still staying friendly to high-throughput hardware. 4. Tensor tile (×1024 → 1K) – Data moves through the chip in blocks of 1024 values at a time, defining the granularity at which tensors are fetched and manipulated. 5. Matrix Multiplication Unit (MXU) (×64 → 64K) – A systolic array where matrix multiplication is not abstract but physical, with tensor tiles flowing through fixed hardware to achieve the highest possible throughput. 6. Vector Memory (VMEM) (×2048 → 128M) – On-chip working memory that holds activations, partial results, and intermediates, sized specifically to keep the systolic array busy without stalling. 7. Common Memory (CMEM) (×8 → 1 GB) – A small but critical shared memory sitting between VMEM and HBM, used for staging, accumulation, synchronization, and cross-lane coordination. 8. HBM (×96 → 96 GB) – Off-chip high-bandwidth memory where model weights and large states live, implemented as HBM3e with 16 stacks at 6 GB each, for a total of 96 GB. 9. Dual-Die (x2 → 192GB) – Two tightly coupled compute dies operate as a single logical accelerator, each with its own local HBM, effectively doubling memory capacity and bandwidth while allowing tensors and activations to stream seamlessly across dies as if they lived on one chip. I created this drawing for this week's seminar. I’ll take you through these 9 levels in a beginner-friendly way by hand ✍️. RSVP 👉 byhand.ai/seminar

Only a stupid person will skip this. A FULL NotebookLM + Google Gemini course, free. Movies won’t upgrade your skills — this will.