🏠 Wayfair Scales Catalog Quality with OpenAI

Wayfair has moved OpenAI models from experiment to production infrastructure, embedding them into supplier support and catalog management workflows that touch 30 million products. The home goods retailer corrected 2.5 million product tags across over a million of its products, while automating 41,000 supplier support tickets per month.

Key Insights:

• Wayfair's catalog team manages 30 million products across nearly 1,000 product classes, with 47,000 different attribute tags requiring consistent accuracy

• A single tag-agnostic OpenAI model replaced expensive custom models - Wayfair expanded coverage to new attributes at 70 times the rate compared to a year prior

• The system has run in production on more than 1 million products, correcting 2.5 million tags across the most visible and purchased items in the catalog

• AI now automates 41,000 supplier support tickets monthly, hitting up to 70% automation in some workflows - freeing teams from manual triage

• Wayfair deployed over 1,200 ChatGPT Enterprise seats across its approximately 12,000-person workforce to support internal operations

The Bigger Picture: Wayfair's shift from bespoke models to a single general-purpose system signals a turning point for enterprise AI. That 70 times acceleration in attribute coverage proves the custom model era is ending. Retailers still building one-off solutions are betting against the trend - general-purpose LLMs now scale faster and cheaper than specialized alternatives. For AI practitioners, this is the playbook: embed models into operational workflows where complexity and volume are highest, measure impact in millions of corrections, and treat AI as infrastructure, not experimentation.

Read More →

🌍 AMI Labs Raises $1.03B for World Models

Yann LeCun's AMI Labs has closed a $1.03 billion round at a $3.5 billion pre-money valuation to build world models - AI systems that learn from reality, not just language. The French lab initially sought €500 million but ended up raising approximately €890 million, joining Fei-Fei Li's World Labs and SpAItial in the emerging world model category.

Key Insights:

• AMI Labs raised $1.03 billion at a $3.5 billion pre-money valuation - up from €890 million awarded last December

• The company is building world models based on JEPA (Joint Embedding Predictive Architecture), proposed by LeCun in 2022 as an alternative to LLMs

• CEO Alexandre LeBrun expects 'world models' to become the next AI funding buzzword within six months, with every company claiming the label

• AMI's first partner is Nabla, the digital health startup where LeBrun serves as chairman, targeting healthcare applications where LLM hallucinations pose life-threatening risks

• The team includes LeCun as chairman, Meta's VP for Europe, Laurent Solly as COO, and high-profile researchers Saining Xie, Pascale Fung, and Michael Rabbat

The Bigger Picture: AMI Labs represents a fundamental bet against the LLM paradigm - and investors are writing billion-dollar checks for it. LeBrun admits it could take years to ship commercial products, yet the round closed at a clear success. That premium reflects two realities: LLMs have hit a ceiling in domains like healthcare where hallucinations matter, and the talent behind AMI - LeCun, Solly, Xie - commands trust that buys time. For AI builders, this is the counter-narrative to scaling laws: sometimes the next breakthrough requires starting over, not just adding parameters.

Read More →

🧠 Microsoft Ships Phi-4-Reasoning-Vision-15B

Microsoft has released Phi-4-Reasoning-Vision-15B, a compact 15-billion-parameter open-weight multimodal model built on a mid-fusion architecture. Trained on 200 billion multimodal tokens, the model balances fast direct perception with deep chain-of-thought reasoning, targeting math, science, and computer-use agent applications.

Key Insights:

• Phi-4-Reasoning-Vision-15B is a 15B open-weight multimodal model trained on 200 billion multimodal tokens, now available on Hugging Face (HF)

• The mid-fusion architecture switches between fast direct responses for simple tasks and deeper reasoning when complexity demands it

• The model handles high-resolution screens well, making it particularly strong for computer-use agents that need to navigate GUIs

• Microsoft positions it as efficient for complex math and science tasks, rivaling larger models while staying compact enough for practical deployment

• This is the second launch from the Phi-4 Reasoning product line, following earlier releases in the small language model category

The Bigger Picture: Phi-4-Reasoning-Vision proves that multimodal reasoning doesn't require 100-billion-parameter models. At 15B parameters, it's small enough to run locally or deploy cost-effectively at scale - yet Microsoft claims it rivals larger models on math, science, and agent tasks. That efficiency gap matters; if a 15B model can handle GUI navigation and chain-of-thought reasoning, the economics of agentic AI just shifted. For developers building computer-use agents, this is the new baseline: compact, open-weight, and capable enough to make local deployment viable.

Read More →

📰 OpenAI Tackles Prompt Injection with Instruction Hierarchy Training

OpenAI released GPT-5 Mini-R, a model trained to prioritize instructions by trust level (system > developer > user > tool). The approach boosted robustness to impersonation attacks from 0.23 to 0.94 compared to baseline models, addressing a core safety challenge as AI systems juggle conflicting instructions from multiple sources.

Read More →

📰 Anthropic Launches AI Code Reviewer as Pull Requests Surge

Anthropic introduced Code Review in Claude Code to handle the bottleneck created by AI-generated code flooding enterprise workflows. The tool integrates with GitHub to automatically analyze pull requests and flag logical errors, targeting companies like Uber and Salesforce where the amount of pull requests from Claude Code has massively increased.

Read More →

📰 Google Embeds Gemini Chat Directly into Docs, Sheets, and Drive

Google Workspace now features an in-app Gemini chat window that generates formatted documents, builds entire spreadsheets from prompts, and matches writing styles across collaborators. The update aims to keep AI assistance where users already work instead of forcing context switches to separate tools.

Read More →

autoresearch 🧪

Andrej Karpathy's open-source project lets AI agents autonomously modify model-training code, run experiments, and iterate overnight on a single GPU - roughly 12 experiments per hour or around 100 while you sleep.

The system uses a three-file architecture where humans write research goals in Markdown while the agent handles the coding loop, shifting technical work from "write functions" to "design bounded environments and feedback loops." This could change how you approach ML experimentation: instead of manually tweaking hyperparameters, you define the research objective and let the agent grind through variations.

Try it →

Promptfoo 🛡️

OpenAI is acquiring Promptfoo, an AI security platform trusted by over 25 percent of Fortune 500 companies that automates security testing and red-teaming for LLM applications. The platform detects risks like prompt injections, jailbreaks, data leaks, and tool misuse before deployment, with integrated reporting for governance and compliance.

If you're deploying AI agents into production workflows, this matters: security testing moves from afterthought to native platform capability, and Promptfoo's open-source CLI will continue development alongside enterprise integration into OpenAI Frontier.

Try it →