Robert-Rami Youssef’s Post

 Most AI systems treat time as a linear timestamp. We built a system that models human consequence instead. I’m excited to share a public blueprint of our latest R&D breakthrough: The Felt-Time Dual-Mode Algorithm (v4). The Problem: Standard LLM and governance filters look at text superficially. They can catch an unsafe keyword, but they are blind to situational friction, historical baggage, and emotional urgency. Our early multiplicative models ($v1$/$v2$) suffered from "zero-collapse"—if a highly critical institutional prompt was dryly worded, the math flatlined to zero. The Architecture: To fix this, we engineered a dual-mode hybrid framework that dynamically branches based on observable features: 1️⃣ Reinhabited Mode ($T_{\text{reinhabited}}$): Models past/present intensity using a geometric-additive mix of feeling, memory depth, and present attention. 2️⃣ Anticipated Mode ($T_{\text{anticipated}}$): Tracks future-facing uncertainty, countdowns, and active time-monitoring without relying on historical text proxies. https://lnkd.in/grwz_wdw

Very excited about this work! So much recent progress in understanding scaling of AI models. Importantly, the progress in theory has direct implications for implementation of models. Defense Advanced Research Projects Agency (DARPA) AIQ: Mathematical foundations for AI and evaluation!

🚨 New paper: μP-style hyperparameter transfer meets energy-based AI We initiate a study of a μP-style hyperparameter transfer methods for Dense Associative Memories (DenseAM) — the family of models behind Hopfield networks, Energy Transformers, and energy-based GPT. Today, scaling AI systems requires painfully expensive hyperparameter tuning at every model size. Most existing transfer methods (including μP-style ideas) were developed around standard feed-forward architectures. Energy-based systems are different. They introduce a different dynamics: • shared weights across iterative computation • native residual structure • steep nonlinearities enabling massive memory capacity • energy-descent computation instead of one-pass inference At present it is unknown how to reliably transfer hyperparameters in these systems. Our work is a first step toward solving this problem. We study a minimal Dense Associative Memory setting that still preserves the key dynamical behavior of modern energy-based systems — and show that μP-style transfer ideas may extend beyond conventional feed-forward networks. Here is the paper: https://lnkd.in/eQpkKsjr. Curious to hear what others think about the future of energy-based architectures.

A new lab called Subquadratic shipped a 12M-token LLM at 1/5 the cost of frontier models. Most people won't care. They should. Here's why. Half the AI infrastructure stack exists to work around one thing: transformer attention scales O(n²). Double the context, quadruple the cost. So the industry built workarounds — RAG, embeddings, chunking, sub-agents passing notes, recursive prompting, memory folders. SubQ's pitch is that you don't need most of that. Their architecture scales linearly. Early signals: – 97% on RULER 128K at $8 (frontier models: ~$2,600) – 92% recall at 12M tokens, 100M targeted by year-end – A CLI agent that loads your whole repo in one pass — no chunking, no retrieval theater Frontier models still lead on hard reasoning. Mamba made similar noise and didn't displace transformers. One launch isn't a paradigm shift. What this means for product teams: – If you're building a RAG-heavy product, your moat just got narrower – Code agents stop being orchestration problems and start being context problems – Long-context use cases (legal review, large codebases, long-form research) drop a full order of magnitude in cost – Several AI tooling categories quietly start to commoditize The caveats matter. Frontier models still win on harder reasoning benchmarks (Opus 4.7 leads SWE-Bench at 87.6%, SubQ at 81.8%). Mamba promised something similar a few years back and didn't replace transformers. But here's the second-order question worth sitting with: How much of the AI stack you're building today is solving an actual customer problem, and how much is solving for transformer constraints that may not be permanent?

AI reasoning is quietly moving beyond chain-of-thought. A new wave of research is revisiting recursive architectures not by scaling models bigger, but by changing how they think. Traditional LLMs reason in token space. They generate one token at a time, feeding every intermediate thought back into the model. That works. But it becomes slow, memory-heavy, and expensive for complex reasoning. Recursive models take a different approach. Instead of generating “thinking tokens,” they reason inside latent space using internal memory loops. Architectures like HRM and TRM are showing why this matters. A 5M–7M parameter Tiny Recursive Model achieved: • 87.4% on Sudoku-Extreme • 45% on ARC-AGI-1 • 85% on maze navigation tasks Meanwhile, RecursiveMAS applies the same idea to multi-agent systems, allowing agents to exchange latent representations instead of text. Result: • up to 2.4× faster inference • 75.6% lower token usage • higher reasoning accuracy The industry spent years scaling parameters. Now it’s rediscovering recursion, latent reasoning, and compute-efficient intelligence.

The Scaling War in Time Series AI is over. Data won. This week’s release of Datadog’s Toto-2.0 and its "Family and Friends" ensemble marks a possible "GPT-3 moment" for our field. For two years, we debated if we should treat time series as language (Chronos) or use specialized "any-variate" attention (Moirai). The answer? Neither. Just get bigger. I’ve spent the last few weeks connecting the dots between: 1️⃣ The ICLR 2025 paper on Scaling Laws. 2️⃣ The IBM Research deconstruction of "Generic" Transformers. 3️⃣ The 2.5 Billion parameter conquest of Toto-2. The result is a new era where "Generality is the ultimate bias." We no longer need models that understand seasonality; we have models that have seen 500 billion examples of it. Full breakdown of the 2026 Time Series landscape here:

𝗔𝗻 𝘂𝗻𝗱𝗲𝗿𝗿𝗮𝘁𝗲𝗱 𝗿𝗲𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 𝗳𝗿𝗼𝗺 𝘀𝗼𝗺𝗲𝗼𝗻𝗲 𝗯𝘂𝗶𝗹𝗱𝗶𝗻𝗴 𝗶𝗻 𝗔𝗜: 𝗔𝗱𝘃𝗮𝗻𝗰𝗲𝗱 𝗥𝗔𝗚 𝗶𝘀 𝗻𝗼𝘁 𝗷𝘂𝘀𝘁 𝗮𝗯𝗼𝘂𝘁 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗶𝗻𝗴 𝗱𝗼𝗰𝘂𝗺𝗲𝗻𝘁𝘀, 𝗶𝘁'𝘀 𝗮𝗯𝗼𝘂𝘁 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗶𝗻𝗴 𝘁𝗵𝗲 𝗿𝗶𝗴𝗵𝘁 𝗶𝗻𝗳𝗼𝗿𝗺𝗮𝘁𝗶𝗼𝗻 𝘄𝗶𝘁𝗵 𝗯𝗲𝘁𝘁𝗲𝗿 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴, 𝗹𝗼𝘄𝗲𝗿 𝗹𝗮𝘁𝗲𝗻𝗰𝘆, 𝗮𝗻𝗱 𝗵𝗶𝗴𝗵𝗲𝗿 𝗮𝗰𝗰𝘂𝗿𝗮𝗰𝘆. Over the past few days, I explored how modern RAG systems actually work beyond the basic "𝘳𝘦𝘵𝘳𝘪𝘦𝘷𝘦 + 𝘨𝘦𝘯𝘦𝘳𝘢𝘵𝘦" pipeline. I went deep into: - Vector quantization - LLM token generation flow - Traditional RAG architecture - Query optimization and hybrid search - Re-ranking and context compression - Multi-hop retrieval and embedding optimization - Evaluation metrics like Recall@K, MRR, Precision, F1, BLEU, ROUGE What really clicked for me was mapping the internal flow: User Query → Query Optimization → Retrieval Strategy → Re-ranking → Context Compression → Prompt Engineering → LLM Generation This gave me a much clearer picture of how production-grade AI systems are actually built. 𝗛𝗲𝗿𝗲'𝘀 𝘁𝗵𝗲 𝘁𝗵𝗶𝗻𝗴: most people talk about RAG like it's simple. But the real magic happens in the optimization layers, the reasoning between steps, and how you compress context without losing signal. Still exploring and learning more every day, but this was one of the most interesting deep dives into GenAI systems architecture so far. Stop thinking RAG is just retrieval; start building it like the complex system it actually is. #AI #GenAI #RAG #AdvancedRAG #LLM #AIEngineering

Spent time deeply understanding RAG (Retrieval-Augmented Generation) today — and one realization stood out: Modern AI systems are no longer just “models.” They’re becoming: • retrieval systems • memory systems • search systems • orchestration layers • context engineering systems RAG isn’t just “chat with PDFs.” It’s essentially an external memory architecture for AI. A few things I explored: • evolution from search systems → LLMs → RAG → agentic retrieval • chunking strategies • embeddings & vector representations • vector databases & ANN indexing • similarity search methods • ingestion vs retrieval complexity in production systems One particularly interesting insight: A lot of AI quality improvements now come from better retrieval and context systems — not necessarily bigger models. Sharing an illustration summarizing the concepts visually 👇 #AI #RAG #LLM #GenerativeAI #MachineLearning #AIEngineering #VectorDatabases #Embeddings #LangChain #LlamaIndex #AIAgents #ArtificialIntelligence

AI agent memory has been treating as a retrieval problem when it’s actually a governance problem. Even in single-agent systems, this leads to drift, stale beliefs, and inconsistent behavior over time. In multi-agent systems, the problem becomes largely unsolved, shared memory creates conflicts, while isolated memory leads to drift and broken ground truth across agents. Why current agent memory architectures are fundamentally misdesigned and what a better approach looks like. Read the full article:

RAG is evolving far beyond simple “retrieve + generate” systems. While learning and building with LangGraph, I explored how modern Retrieval-Augmented Generation architectures are becoming more adaptive, reflective, and agentic. This infographic summarizes the evolution of RAG systems: 🔹 Basic RAG 🔹 Hybrid RAG 🔹 Self-RAG 🔹 CRAG (Corrective RAG) 🔹 Graph RAG 🔹 Agentic RAG 🔹 Adaptive RAG 🔹 Multimodal RAG Key insight: Modern AI systems are no longer just single LLM calls — they are becoming intelligent workflows with retrieval evaluation, query rewriting, routing, memory, tool usage, and self-correction. Understanding these architectures is becoming essential for building reliable AI applications in production. Which RAG architecture do you think will dominate enterprise AI systems over the next few years? #AI #RAG #GenerativeAI #LLM #LangChain #LangGraph #MachineLearning #ArtificialIntelligence #DataScience #AgenticAI #GraphRAG #CRAG #SelfRAG