Generative recommender systems reformulate personalization as sequence modeling over user behavior, and NVIDIA recsys-examples now provides an end-to-end HSTU inference workflow with Dynamo-Triton. PyTorch AOTI compiles HSTU ranking models into native C++ artifacts, while FlexKV-backed KV caching stores reusable attention state to avoid recomputing long user histories. On an @nvidia RTX PRO 6000 Blackwell Workstation Edition GPU, the eight-layer HSTU model achieved up to 5.93x lower latency at batch size 8 with a 100% GPU KV-cache hit rate compared to the same AOTI configuration without caching. Read the full blog post:

Quick lookback on technical AI themes of 2026: (dataset: daily ainews title stories) Jan: - Ralph Loops - RLMs - Filesystems are all you need - X LLM recsys open sourced - X Grok Imagine #1# Video model - Anthropic blocked OpenCode / OpenClaw - MCP Apps - Moltbook - Apple picks Gemini for new Siri - Notable models: Kimi K2.5 - Fast inference: OAI x Cerebras partnership - Big raises: xAI @ $230B + SpaceX, OpenEvidence @ $12B Feb: - Codex App launch - Claude Code 1 year anniversary - Context Graphs - Fast Inference: Taalas 17k tok/s/user Llama 3 8B - Anthropic vs Deepseek/Moonshot/Minimax distillation accusation - Notable models: Opus/Sonnet 4.6, 5.3 Codex and Codex Spark, Qwen-Image 2.0, Seedance 2.0, GLM-5, Gemini 3.1 Pro and Deep Think, Nano Banana 2 - Big raises: OpenAI @ $840B, Anthropic @ $380B, ElevenLabs @ $11B, Cerebras @ $23B Mar: - T3 Code - Autoresearch - Claude Code source leak - Claude Cowork Dispatch - OpenAI buys Astral - Cloudflare-Next.js "Slop Fork" - Meta buys Dreamer - NVIDIA GTC: OpenClaw, Vera CPU, 1T sales backlog, "The Inference Inflection" - ARC-AGI 3 introduced, SOTA models are 1% - Notable models: GPT 5.4 unifies chat and codex, Gemini 3.1 Flash-Lite/Live, Nemotron 3 Super, Composer 2, MiniMax 2.7, Arcee’s Trinity-Large-Thinking - Big Raises: Yann LeCun AMI Labs $1B seed Apr: - Claude Design - OpenAI Symphony & Harness Engineering discourse - @LilianWeng writeup later - OpenAI and Microsoft break up - Tokenmaxxing fears - Codex Superapp - AIE Europe, AIEi Miami - Hermes Agent gains traction w/ Manim skill - Cyber: Anthropic Project Glasswing - Notable models: Claude Mythos "Capybara", Opus 4.7, Muse Spark, Gemma 4, Kimi K2.6, GPT-Image-2, GPT 5.5 "Spud", DeepSeek v4 Pro/Flash Base/Instruct on Huawei Ascend, Poolside Laguna XS.2, Grok 4.3 - Big Raises: not much but Anthropic $19B->$30B ARR in one month... May: - Cyber: OpenAI Daybreak, Tanstack Mini-Shai Hulud - FDE: OpenAI DeployCo, Ode by Anthropic - OpenAI solve Erdos planar unit distance problem - OpenAI deprecates finetuning - "How to land a job at a frontier lab" - Anthropic rents Colossus I from SpaceX for $5B/yr - Thinking Machines Interaction Models (still underrated!!!) - Notable models: GPT-Realtime-2/Translate/Whisper, Gemini 3.5 Flash, Omni, Antigravity 2.0, Opus 4.8, Cosmos 3, Minimax M3 - Big Raises: Fireworks @ $15B, Baseten @ $11B, Cognition @ $26B, Cerebras $60B IPO, Anthropic @ $965B Jun: - Claude Tag - OpenAI Jalapeno - "The Untrainable" - @saranormous - Loopcraft: The Art of Stacking Loops - Metaharnesses discourse? Flue v Eve - Zai Prof @jietang predicts Open Fable by Dec - "SpaceX is a $28B/yr neocloud" - @jaminball - Anthropic-Fable White House issue (resolved) - Notable models: GPT 5.6 Sol/Terra/Luna, Fable 5, Sonnet 5, MAI-Thinking-1 and MAI Family, Nemotron 3 Ultra, GLM 5.2, Reve 2/ Ideogram 4 - Big Raises: SpaceXai buys Cursor for $60B, Supabase @ $10B Jul: - AIE World's Fair 3 - OpenCode 2.0 - Cyber: OpenAI-HuggingFace incident - Chinese Model ban discourse - GPT 5.6 Luna 80% price cut - Notable models: Grok 4.5, Opus 5, Muse Image, Thinky Inkling, Kimi K3, BFL FLUX 3, DeepSeek-v4-Flash-0731 - Big Raises: Aug: - EU AI text watermarking - Microduck - Megakernels - Memory prices up 500% in 12 months - How to steal a Reasoning Trace - Grok @bot! - OpenAI shuts off Cursor - Fast Inference: AMD buys Taalas - @andrewng stars AI Engineering series - Hot Chips: Jalapeno, Cerebrase CS-5, Groq 3, Apple M6 - Demis steps down from CEO, & Dean et al leave - Notable models: Qwen 3.8 Max, Muse Spark 1.2, Muse Glimmer, Grok 4.6, Gemini 3.7 Flash, GLM 5.3, @Fal H3 Max Live - Big Raises: Stripe buys OpenRouter $7B, Poolside $12B reverse-execuhire to NVIDIA, NVIDIA buys HuggingFace for $13B Sep: - AIEi Paris - OpenAI Navier-Stokes singularity - collusion . wiki - Yegge shuts down Gas Town - AEF-1 standard for Third Party Evaluators - Meta Muse App launch. Also new glasses and Charm - OpenAI blocks Cursor - OpenAI Dots - Opus 5.5 explainer videos - Pi 1.0 & Pi Durable - Notable Models: Fable 5.1, Mythos 5.1, Opus 5.5, GPT-6 Astra, GPT Image 2.5, DeepSeek v4.1-Flash, @typesafeai Jev, MiMo-v2.6-Pro new Open SOTA, GPT 6.1 Sol & Ultrafast & Decisions API, Gemini 4 Argon - Big Raises: Cognition @ $48B, Mistral @ $24B, AMD buys World Labs for $8.2B Oct to date: - AIE NYC - @dhh stops using Rails - ArtCraft clones Adobe suite - Interpretability vs Polysemanticity - OpenAI Quasi-Riemann-Hypothesis & 722 Math - Notable Models: Reflection Beam, Claude Haiku 5.5, Muse Wonder & Wonder Full reach Opus 5.5 tier - Big Raises: TypeSafe $7.5B 3 weeks after launch, Modal @ $16B (this was done for research purposes for the keynote in Nov, just sharing notes in case helpfuil)

What happens when you let AI agents autonomously run research on recommender models that take days to train? This paper finds out. Auto-RecSys: Harnessing Autonomous Research Agents for Industry-Scale Recommender Systems ❓ Why aren't existing autonomous research agents enough? 💡 Prior systems assume feedback loops of minutes to hours. Industry-scale recommender models take days per training run and involve thousands of lines of configuration, so serial iteration breaks down and you need a fundamentally different design that runs many ideas in parallel across distributed servers and recovers from failure. ❓ How does it juggle parallel experiments and recover from crashes? 💡 Each idea keeps its own independent state file moving through ideating, implementing, validating, training, and analyzing. When a session restarts, it rereads a global registry and trajectory logs on shared storage, letting work resume seamlessly across servers. ❓ Does the system actually get better over time? 💡 Experiment trajectories are distilled into model-specific playbooks that record dead ends as "do not" rules. Across 31 iterations, major fixes needed per iteration dropped from 4.0 to 0.5, with 83% of iterations completing with zero fixes. ❓ How autonomous is it in practice? 💡 One session executed 970 consecutive log entries with zero human intervention, and the system even switched its own training workflow after repeated publish failures. #AIAgents# #MachineLearning#

TL;DR: HiRank is a recommender ranking model that fuses sequential and non-sequential features two ways in a single backbone, and uses effective-rank-aware attention to keep representations diverse. In an online test on the Taobao homepage, orders per user rose 2.27%. Title: Unified Heterogeneous Feature Interaction via Effective-Rank Scaling for Recommender Systems URL: Key points ・🧩 Drops the encode-then-interact pipeline; event-level sequence states stay accessible across layers ・🚪 Models the main behavior sequence with HSTU-style gated attention ・🔎 Retrieves from long-term history memories with both interest queries and candidate queries for diverse retrieval ・🔀 Fuses sequence, retrieval outputs, features, and the candidate via token mixing ・🎯 Applies High-Rank Attention over scenario-aware compressed composite tokens ・📊 Offline: homepage GAUC 0.7801 (+0.0186 over the strongest baseline) ・🛒 Online: orders/user +2.27%, GMV/user +2.86% 💡 I like the takeaway that sequence length and representation diversity matter, not just model size. #RecSys# #MachineLearning#

TL;DR: Adding features to a recommender system is easy, but safely removing them is brutally hard. Pinterest built a production framework that automates that whole cleanup process. Title: FORTE: A Production Framework for Feature Evaluation and Deprecation in Industrial Recommender Systems URL: Key points 🧩 Adding a feature just needs a metric win, but removing one must not regress any metric, a built-in asymmetry that stalls cleanup ⏱️ Daily perturbation experiments automatically track feature importance and rank stability over time 🧪 Grouped perturbation evaluates candidate feature sets together, catching interactions before costly retraining 🚀 Pooled execution resolves all experiments for a model in a single distributed inference pass, cutting runtime 54x for 300 experiments 🔗 Automated lineage tracking finds every downstream consumer model, making deprecation safe 📉 On the Pin Ranker model, removing 8 features cut parameters by 29.9%, boosted throughput 19.8-25.5%, and saved hundreds of thousands of dollars a year Feature removal infrastructure deserves as much engineering investment as feature addition, and this looks like a solid blueprint for it. #RecSys# #MLOps#

In generative recommendation, everyone has assumed they know what makes a "good" semantic ID design. So what happens when you actually put that assumption to a large, controlled reproducibility test? 🔍 Generative recommenders turn items into discrete token sequences called semantic IDs, then generate the next item autoregressively. Designs like RQ-VAE, OPQ, and RQ-Kmeans have proliferated, but prior comparisons used different datasets and backbones, so nobody could say which design actually wins. So the authors reran 12 methods side by side under the exact same data splits and evaluation protocol, and the assumptions started falling apart. No single SID design dominates across datasets. Codebook utilization balance barely correlates with recommendation quality. Longer codes don't always help, and bigger backbones can even hurt performance. And when you measure how well an SID preserves an item's local semantic neighborhood, the winning method flips depending on whether you use Jaccard or RBO. The lesson from What Makes a Good Semantic ID for Generative Recommendation? A Reproducibility Study is that good SID design isn't about chasing one metric, it's about balancing several at once. URL: #RecSys# #Reproducibility#

What if you could leave your recommendation retriever completely untouched and just bolt on one lightweight "verifier" to significantly boost recall? A team at Meta just showed exactly that. Title: Recommendation Retrievers Need Verifiers: Universal Generative Reranking for Sequential Recommendations URL: The idea is a "drafter-verifier" pipeline: freeze the existing retrieval model (the drafter) entirely, and add a lightweight verifier that generatively scores candidate items via their identifier tokens. 🌐 Highlight 1: It scales with catalog size On the YaMBDa music recommendation dataset, gains grew larger as the catalog got bigger — at the largest scale (4.65B interactions), Recall@10 improved by up to 29.8%. 🧩 Highlight 2: Works across wildly different drafters The same training recipe worked on 4 architecturally distinct drafters: attention-based SASRec, recurrent GRU4Rec, convolutional NextItNet, and even the LLM-based MiniOneRec. 🔬 Highlight 3: The gain really comes from generative verification Injecting content embeddings directly into the drafter actually hurt standalone performance, but adding the verifier on top recovered it — proving the improvement comes from the output-side generative verification process itself, not simple feature injection. The fact that it bolts on without touching the retrieval index feels like a big deal for adopting this in large-scale production. #RecSys# #MachineLearning#

The Underpriced Truth: Agentic AI Is a Paradigm Shift Centered on Memory 1/ The market will slowly realize: Agentic AI is memory-centric, not compute-centric. The new hardware stack is: ① Memory — HBM / DRAM / NAND ② Parallel compute — GPU / ASIC ③ Coordinator — CPU CPUs stopped doing the heavy lifting a long time ago. This isn't a cycle. It's a paradigm. 🧵👇 2/ First principles Humanity's ultimate pursuit of intelligence has always been two things: Infinite memory + infinite compute. When we say someone is smart, we mean two things: "good memory" + "fast thinking." Machine intelligence is walking the exact same path. 3/ The story the market already understands: HBM LLM inference's decode stage is a textbook memory-bound workload. Every token generated → drag the entire KV cache across memory. Bandwidth too low → expensive GPUs sit idle. That's why every new GPU generation ships with more HBM bandwidth and capacity. 4/ The story the market is missing The "1M context" you keep hearing about? It is not assembled inside the GPU inference cluster. So where is it actually built? 5/ It's built on the traditional servers running the agentic system Those CPU + huge-DRAM servers are quietly doing the heaviest lifting: • loading user long-term & short-term memory • loading the agent's system spec / prompt • loading skill / tool / subagent definitions • compressing the context once it overflows 1M tokens All of this lives in DRAM, not HBM. 6/ Compare this to the previous era In the web / mobile era, we barely stored any user context at all. Only search / recsys / ads kept a small user profile — maybe 1/20, even 1/100 of the data volume an agentic system needs today. That asymmetry is the real overlooked inflection point. 7/ The supply chain is already telling this story Server CPU : DRAM ratio is climbing fast: • Web / Mobile era: 1 core : 4 GB • Agentic AI today: 1 core : 16 GB • Deep agentic future: 1 core : 64 GB and beyond 8/ And it's NOT just "4x more memory" Under agentic workloads, a single CPU serves a fraction of the users it used to. When the entire IT stack migrates to agentic: • CPU count grows several-fold to ~10x • DRAM total grows tens-fold to ~100x That's the part nobody is pricing in. 9/ The conclusion Agentic AI is a paradigm shift centered on storage + parallel compute. The software paradigm changed. The hardware paradigm changed with it. Only those who deeply understand the technology will see it: This isn't a memory cycle. It's a memory paradigm. 10/ Time horizon Given how early we still are on: • user adoption rate • depth of usage per user We are at least 5 years away from the cyclical top of this memory wave. (Zoom out far enough and everything is a cycle — but this one is nowhere near peak.) $MU $DRAM $SNDK

Replacing CUDA with FBTriton for Table Batched Embedding (TBE) kernels delivered massive gains for our rec sys models. @Meta contributors achieved up to 1.28x faster forward passes, 2x faster backward passes, and a huge boost in developer velocity. Check out the engineering deep-dive: ✍️ Daohang Shi, Oleksandr Stashuk, Rupert Wu, Liangbei Xu, Rich Zhu