☕️ Meta doubles ad AI training efficiency

Hi there, this is your daily ☕️ AIpresso.

In today's AIpresso:

📊 Meta doubles ad AI training efficiency

💳 Cloudflare launches wallet for AI agents

💾 SK hynix unveils new AI memory standard

🤖 Cloudflare open-sources Astro bug-fixing tool

📌 Alibaba's AI coded 16 days straight

Plus: 💡 6 strategies & tactics, 🎁 7 other news you might like, 🧰 6 tools, and 📚 5 papers.

📊 Meta doubles ad AI training efficiency LINK

Meta doubled end-to-end training efficiency for GEM, the foundation model behind ads recommendations across Instagram and Facebook, reaching 20-25% MFU while scaling total training FLOPs 4x over 12 months on several thousand latest-generation GPUs.

Gains came from a custom recommendation kernel library-Jagged Flash Attention, GDPA, and BlockAttention-plus MXFP8 attention and MLP, with JFA eliminating up to 50% padding waste and JFA v4 delivering 40-140% TFLOPS improvement over v2.

Scaling used topology-aware 5D parallelism with SM-free collectives, combining 2D FSDP plus Expert Parallelism for dense parameters and Fully Sharded 2D Model Parallelism for the trillions of sparse embedding parameters, though CTR/CVR objectives remain highly sensitive to naïve low-precision training and risk quality regression.

💳 Cloudflare launches wallet for AI agents LINK

Cloudflare launched a wallet-and-identity service that lets AI agents shop on your behalf, pairing a bot-readable identity called cloudflare.id with a virtual wallet holding stablecoin funds that agents can spend at participating merchants.

Users fund Cloudflare Wallets via bank transfer converted to dollar-pegged stablecoins, then grant agent access with optional guardrails like spending limits and merchant whitelists, while merchants get tools to communicate with incoming customer agents.

The service builds on the open-source X402 protocol from a consortium of Coinbase, Cloudflare, and Stripe, though Cohen expects only tech-savvy developers and AI firms buying data to adopt it before mainstream consumers.

💾 SK hynix unveils new AI memory standard LINK

SK hynix and SanDisk have introduced High Bandwidth Flash (HBF), a new memory standard that stacks NAND dies to bridge the performance gap between HBM and enterprise SSDs for AI accelerators.

HBF scales up to 512GB per configuration by stacking NAND in eight or 16 layers, delivering 0.4TB/s to 3TB/s bandwidth and letting firms park entire 500B-parameter models too costly to hold in HBM.

The spec supports UCIe, allowing HBF to attach to various CPUs and GPUs as a flexible layer, though it is designed to complement HBM's capacity limits rather than replace it for immediate compute.

🤖 Cloudflare open-sources Astro bug-fixing tool LINK

Cloudflare has open-sourced triagebot-action, the agent-based GitHub tool that Astro used to shrink its issue backlog from over 200 at the start of the year to around 20, targeting zero within a month.

Running as a GitHub Action, the tool triggers a four-stage pipeline-reproduce, diagnose, verify, and fix-where each stage is a separate AI agent driven by a state machine encoded in GitHub labels for visible tracking.

The underlying Flue agent framework runs on any infrastructure with no Cloudflare dependency, though the "fix" stage remains the weakest, with Schott saying he can't predict when its auto-fixes will be trustworthy close to 100% of the time.

📌 Alibaba's AI coded 16 days straight LINK

Qwen3.8-Max uses a sparse mixture-of-experts design with hybrid attention, holding 2.4 trillion parameters total but activating about 95 billion per token, drawing only on the parts of the model needed for each specific task.

A harness handles everything around the model during long jobs, letting the agent decide next steps while the system manages the rest, and Alibaba tested it with QwenWork, Claude Code, Codex, OpenClaw, and Hermes.

In Alibaba's tests, the agent spent 16 days building "oh-my-cli" with 265 commits, 127 pull requests, and 151 GitHub issues, reproduced a research paper in five days, and cut a cryptographic circuit from 8,298 gates to 678.

💡 Strategies & Tactics

> The Inference Engineering Masterclass, Philip Kiely & Ali Taha, Baseten: Explains how to turn newly released open AI models into fast, cheap production services by reusing cached inputs, quantizing models, and predicting tokens with smaller helper models.

> How to Run Isolated Tenant Kubernetes Clusters on Shared GPU Infrastructure: Give each team its own private Kubernetes cluster on shared GPU hardware by pairing KAI Scheduler for fair GPU splitting with vCluster for isolation.

> NVIDIA Vera Storage Benchmarks: Faster Encryption, Compression, Integrity Checking, and Recovery for AI-Native Storage: NVIDIA's Vera storage processor runs encryption, compression, and data-recovery tasks up to 3.67x faster than x86 chips, so AI storage keeps pace without more power.

> How to build a cloud software factory - computer use verification: Give your coding agents the ability to click through your app so they can reproduce bugs and record videos proving fixes work, cutting review time.

> How Stripe Built Kai on Deep Agents in 1 Week: Stripe built its company-wide AI assistant Kai in a week by using an open-source agent framework, letting the team skip infrastructure and focus on Stripe-specific workflows.

> How to use Vercel eve: The Next.js framework for AI agents: Explains how to build custom AI agents faster with Vercel's open-source eve framework, which organizes agents into files and folders like a Next.js app and handles infrastructure like sandboxing and approvals automatically.

Other news you might like

• Anthropic signs a $10bn compute deal with a week-old cloud startupLINK

• Kimi K3, The Manos, The Mythos, The LegendosLINK

• Google ADK flaws reveal what happens when AI agents trust the wrong messageLINK

• Copilot prompt injection goes viral in your documentsLINK

• Deliberate Alignment Faking as a Defense Against Model PoisoningLINK

• Attackers are crafting malicious AI instruction files to turn your agentic workflows into quiet criminal helpersLINK

🧰 Trending tools

AgenticX: a production-ready multi-agent platform offering an SDK, CLI, and desktop app for building, orchestrating, and managing intelligent AI agents.LINK

free-one-api: a tool that converts reverse-engineered AI chat services like ChatGPT, Bard, and Claude into standard OpenAI-compatible API endpoints.LINK

Wondering: runs AI-moderated user interviews and usability tests at scale, letting product teams collect qualitative feedback without manual recruiting or live sessions.LINK

Stynar: an autonomous AI agent that researches prospects, personalizes outbound emails, manages replies, and books meetings automatically, replacing manual SDR work.LINK

Crodo AI: a voice-first macOS assistant that lets you dictate, ask about your screen, capture meeting notes, and manage Gmail, Calendar, and Slack without app-switching.LINK

Atlaso: a shared memory layer that syncs context, decisions, and preferences across Claude Code, Cursor, Codex, and ChatGPT, eliminating repetitive re-explaining.LINK

📚 Trending papers & reports

Competitive debate archives compiled into a 3.5 million document dataset now let AI tools be trained to summarize complex arguments, opening the door to sharper research and writing assistants for any argument-heavy business or legal work.LINK

Curriculum-guided model training grows a language model's depth in step with progressively harder training material, outperforming standard pretraining on question-answering tests like PIQA and ARC while using less compute.LINK

AI writing detectors systematically flag non-native English writers as machine-generated more often, while separate tracking shows AI-assisted text now appears widely in peer reviews, scientific papers, corporate press releases, and job postings.LINK

AI self-correction gaps show chatbots catch errors when a user points them out but miss the exact same mistake in their own answers ~65% of the time, though adding one word, "Wait," cuts that blind spot by ~89%.LINK

Multilingual farm chatbots get fine-tuned on locally-generated question-and-answer examples in English, Hindi, and Punjabi, making their farming advice notably more accurate and relevant than off-the-shelf chatbots.LINK

See you tomorrow for a new dose of ☕️ AIpresso!