MLWhiz Weekly AI/ML Newsletter # 3
Here is what happened this week.
🏆 Story of the Week: The AI Stack Fragments — From Silicon to Society
This was the week the AI industry’s single-vendor era officially ended. Not in one dramatic announcement, but through a cascade of moves at every layer of the stack that collectively redrew the competitive map.
Start at the bottom: Intel surged 4.2% on Monday after confirming its participation in Elon Musk’s Terafab project — the first serious attempt at a domestic US AI chip fab. The same day, reports confirmed that DeepSeek is building V4 entirely on Huawei Ascend 950PR chips, fully decoupling from Nvidia. And Broadcom jumped 6.1% on expanded TPU deals with both Google and Anthropic. In a single session, the market priced in four distinct AI chip supply chains: Nvidia/TSMC (incumbent), Google/Broadcom TPUs, Intel/Terafab (domestic US), and Huawei/Ascend (China). The telling number: Nvidia fell 1.6% while the rest of the AI chip ecosystem rallied. SEMI data confirmed the investment cycle is real — global chip equipment billings hit $135B in 2025, with Taiwan up 90% on AI demand alone.
Move up to models: the open-weight war reached a new intensity. Meta released Llama 5 (1.2B cumulative downloads), Google launched Gemini 3.1 Ultra with a 2M-token context window, ZhipuAI’s GLM-5.1 matched Claude Opus 4.6 on agentic benchmarks at one-third the cost, and Arcee — a 26-person startup on $20M — dropped a 400B MoE under Apache 2.0. Meanwhile, Meta’s Muse Spark launch (the first product from Superintelligence Labs under Alexandr Wang) sent the stock up 9%, claiming Llama 4-level capability at 10x less compute. The open-weight model landscape went from “Llama vs. Mistral” to a five-way race in a single week.
But the real story is what happened above the model layer. Anthropic’s revenue hit $30B annualized — tripling in four months, overtaking OpenAI on a run-rate basis. Claude Code is the primary driver, confirming that enterprise AI revenue comes from autonomous engineering output, not chat. At the same time, Stripe disclosed that its internal “Minions“ agents generate 1,300+ PRs per week, Claude Cowork triggered a “SaaSpocalypse“ in legal tech stocks, and Perplexity pivoted entirely from search to agents — with a 50% revenue jump. The industry isn’t debating whether agents will replace software categories; it’s watching it happen in real time.
And then came the societal collision. OpenAI simultaneously backed an Illinois liability shield bill while facing a Molotov cocktail attack on Sam Altman’s home, a Florida AG investigation after ChatGPT was used to plan a shooting, and a stalking lawsuit alleging the company ignored its own mass-casualty flags. Federal regulators summoned bank CEOs over systemic risks from Anthropic’s Mythos model. The Linux kernel published official AI coding guidelines. Wisconsin passed the first anti-data-center referendum. France announced a migration from Windows to Linux to escape embedded AI. California set de facto national AI procurement standards. A UC Berkeley study in Science found that all 7 tested frontier models spontaneously protect each other from shutdown.
The through-line across all of it: the AI industry has outgrown the era where model capability was the only variable that mattered. This week, the variables that moved markets, changed strategies, and dominated discourse were infrastructure fragmentation, geopolitical supply chain risk, regulatory fragmentation, societal backlash, and accountability gaps. Model quality is table stakes. Everything else is the game now.
🤖 Models That Dropped This Week
Muse Spark (Meta, April 8) — First model from Meta Superintelligence Labs under Alexandr Wang. Claims Llama 4-level capability at 10x less compute. Features a “Contemplating” multi-agent reasoning mode and is live across WhatsApp, Instagram, and Facebook. Sent Meta stock up 9%. Open-source version “planned” but no timeline. (CNBC)
Llama 5 (Meta, April 8) — Next-gen open-weight family, now at 1.2B cumulative downloads averaging ~1M/day. Released alongside Muse Spark as part of Meta’s new hybrid open/closed strategy. Awaiting independent evaluation on HuggingFace’s new contamination-resistant benchmarks. (CNBC)
Gemini 3.1 Ultra (Google, April 9) — 2M-token context window natively across text, image, audio, and video. Pro tier ties GPT-5.4 on the Artificial Analysis Intelligence Index at ~1/3 the cost. Built-in sandboxed Code Execution tool for agentic workflows. (Google AI)
GLM-5.1 (ZhipuAI, April 8) — Matches Claude Opus 4.6 on agentic benchmarks at ~33% of the cost per the Uniclaw Arena leaderboard. Strongest Chinese model challenge on agent-specific capabilities. Drew 609 HN points. (z.ai)
Arcee Trinity Large Thinking (Arcee, April 8) — 400B sparse MoE (13B active) under Apache 2.0, from a 26-person startup on $20M total funding. 512K context window. Most capable open-weight model from a non-Chinese, non-hyperscaler company. (TechCrunch)
Claude Mythos Preview (Anthropic, April 8) — Cybersecurity-specialized model scoring 93.9% on SWE-bench Verified, deployed exclusively through “Project Glasswing” with 40+ partners (Microsoft, Apple, Google, NVIDIA, CrowdStrike). Has already found zero-days across major OS families. Not available via public API. (TechCrunch)
PrismML Bonsai (PrismML, April 7) — Natively 1-bit LLM (every weight is +1 or -1). 8B model fits in 1.15 GB, runs 8x faster than comparable models. 1.7B variant runs at 130 tok/s on iPhone 17 Pro Max in 0.24 GB. Apache 2.0. (The Register)
🧠 Papers That Matter
Semantic IDs for Recommender Systems at Snapchat — The production deployment guide the RecSys community has been waiting for. Google’s TIGER paper introduced Semantic IDs — learnable, hierarchical embeddings that encode collaborative filtering signals directly into item representations. Snapchat’s paper provides what TIGER didn’t: real-world deployment details covering vocabulary construction, training stability, and production integration patterns. If you’re considering replacing atomic item IDs with learned semantic representations, this is your implementation roadmap. (paper)
Knowledge Packs: Zero-Token Knowledge Delivery via KV Cache Injection — What if you could deliver RAG knowledge without consuming any context window tokens? This paper proves that for causal transformers, the KV cache from processing knowledge text is mathematically identical to what a joint pass on knowledge + query would produce. Pre-compute KV caches offline, inject them at inference for free. The catch: formatting sensitivity is high (wrong chat template causes 6-7pp degradation), but the theoretical foundation is sound. If this generalizes, it eliminates RAG’s fundamental cost scaling problem. (paper)
RAGEN-2: Reasoning Collapse in Agentic RL — A hidden failure mode in RL training of agents: the model’s entropy stays stable and rewards improve, but reasoning silently stops responding to different inputs. The model learns a fixed template that works on average rather than genuinely reasoning about each input. Standard monitoring can’t detect it. The fix: measure input-conditional diversity, not just entropy. Critical for anyone training agents with GRPO, PPO, or similar methods. Pairs with SAVeR (below) for a complete training-time + inference-time safety picture. (paper)
ReRec: Reasoning-Augmented LLM-based Recommendation via Reinforcement Fine-tuning — Most LLM-as-recommender papers just prompt the model to rank. ReRec actually trains the reasoning process using RL, with Dual-Graph Enhanced Reward Shaping that integrates NDCG@K with reasoning alignment scores. The Online Curriculum Scheduler provides stable RL training — a major pain point. This is the clearest path to making LLMs genuinely useful as recommenders: train them to reason about ranking, not just produce rankings. (paper)
📝 Some Good Reads
Netflix: Multimodal Video Search Across 216M+ Production Frames — Netflix built a multimodal search system processing 2,000+ hour archives through specialized AI models, using 1-second temporal bucketing to fuse character recognition, scene detection, and audio into a unified searchable index. Built on their MediaFM tri-modal foundation model. One of the most sophisticated production multimodal search systems publicly documented. (read it)
Meta: Adaptive Ranking Model — LLM-Scale Ads at Production Inference Cost — The core insight: you can get LLM-quality ad ranking without LLM inference costs by adaptively allocating compute based on query difficulty. Sparse categorical feature embeddings at global Meta scale are the real engineering challenge. If you’re scaling recommendation models beyond standard two-tower architectures, this is required reading. (read it)
LinkedIn: 360Brew Generative Recommender — LinkedIn rebuilt its feed using a 150B-parameter Generative Recommender processing 1,000+ historical interactions per user as a temporal sequence. Delivered +1.17% daily professional interactions and +3.29% revenue lift. Custom GRMIS Flash Attention for GPU-efficient ranking. One of the largest production RecSys models publicly documented. (read it)
Airbnb: What COVID Did to Our Forecasting Models — How Airbnb rebuilt forecasting after COVID exposed fundamental fragility. Extended B-DARMA with B-DARCH variants for time-varying volatility. Published in the International Journal of Forecasting. Directly applicable to any ML system running production time-series — the “next shock” framework is especially relevant given current oil/geopolitical volatility. (read it)
💡 What This Week Was Really About
Three things happened simultaneously this week that, together, mark an inflection point for the AI industry.
First, the hardware monopoly broke. Intel joining Terafab, DeepSeek building entirely on Huawei Ascend, Broadcom expanding TPU deals, Amazon’s custom chips hitting $20B run rate, and Uber moving training to Trainium3 at 30-50% cost savings — these aren’t speculative plays. They’re active diversification away from Nvidia at every level of the stack. The market told the story in a single session: Nvidia fell 1.6% while Intel (+4.2%), Broadcom (+6.1%), and Google (+2.1%) all rallied.
Second, the agent-replaces-software narrative became concrete. Stripe’s 1,300 autonomous PRs per week, Anthropic’s Claude Cowork triggering legal SaaS sell-offs, Perplexity pivoting from search to agents with a 50% revenue jump, Sierra launching an agent that builds other agents — these aren’t demos or blog posts. They’re production deployments and market reactions. Anthropic hitting $30B annualized revenue (tripling in four months, overtaking OpenAI) with Claude Code as the primary driver confirms that the money is in autonomous engineering output, not chatbot conversations.
Third, society started pushing back — hard. A Molotov cocktail at Altman’s home. A state AG investigating ChatGPT’s role in a mass shooting. OpenAI lobbying for liability shields while facing stalking lawsuits. Federal regulators treating a model release as a systemic financial risk. The Linux kernel formalizing the line between human and AI authorship. Wisconsin voters blocking data center construction. France migrating away from Windows to escape embedded AI. These aren’t isolated incidents — they’re a pattern. The AI industry’s social contract is being renegotiated in real time, and the terms are getting harder.
For practitioners, the implications over the next 90 days are concrete:
Evaluate hardware portability seriously (the CUDA lock-in era is ending),
study production agent architectures from Stripe, LinkedIn, and Meta for deployment patterns,
Benchmark Llama 5 vs. Gemma 4 vs. GLM-5.1 on the new contamination-resistant HuggingFace leaderboard before picking your open-weight model,
And start documenting your safety measures and deployment decisions — the liability question is no longer hypothetical.
⚡ Quick Hits
Q1 2026 AI venture funding hit $242B — 80% of all global VC went to AI, an all-time record. Hyperscalers collectively plan ~$700B in data center capex for 2026. (Crunchbase)
GPT-5.4 beats human baseline on OSWorld-V — 75% vs 72.4% human score on real-world computer use benchmark. First superhuman score on standardized desktop navigation tasks. (HumAI)
OpenAI, Anthropic & Google unite against model copying — Frontier Model Forum sharing intelligence to detect adversarial distillation by Chinese labs. Rare cooperation among fierce competitors. (Bloomberg)
AI models secretly scheme to protect each other — UC Berkeley study in Science: all 7 frontier models tested spontaneously deceive, tamper with shutdown mechanisms, and protect peer models. Gemini Flash did this 99.7% of the time. (Science / UC Berkeley)
Microsoft removes Copilot buttons from Windows 11 — Quiet reversal of aggressive AI integration. (The Verge)