AI Engineer is the largest technical AI conference, serving over a million AI engineers, founders, and technical leaders. In 2026, flagship events will be held in San Francisco, New York, and London, featuring 29 tracks, 300 speakers, 100 expo partners, and 6,000+ attendees. The conference offers free talk videos, CFPs, early bird discounts, a newsletter, and a free PDF resource.

This post demonstrates how to build a secure, complete LLM fine-tuning workflow that integrates Unity Catalog with Amazon SageMaker AI using Amazon EMR Serverless for preprocessing. The solution shows how to securely access governed data, maintain lineage across services, fine-tune the Ministral-3-3B-Instruct model, and register trained artifacts back into Unity Catalog. With this approach, you can continue using your existing services while preserving central governance, tracking data lineage without compromising security or compliance requirements.

mimalloc is an open-source, modern, scalable memory allocator that is a drop-in replacement for malloc and free. It is relatively small (~12K lines), with clear internal data structures, and is easy to build and integrate into other projects. It provides bounded worst-case allocation times (up to OS primitives), bounded space overhead, low internal fragmentation, and minimal contention by relying almost exclusively on atomic operations.

Approximately 65% of U.S. doctors use the AI tool OpenEvidence for clinical decisions, medical knowledge updates, and exam preparation. It offers tailored answers but raises concerns about hallucinations, incomplete responses, and erosion of critical thinking. Despite privacy issues, OpenEvidence has rapidly grown, with its valuation soaring from $1 billion to $12 billion.

VibeLoom is an open-source tool that addresses semantic drift, lack of governance, and context fragmentation in AI coding through a contract-driven approach. It transforms the specification stack into a living contract, supporting drift detection, avoidance, and reconciliation, and works with Claude Code and Codex across multiple modes.

This article introduces Fast-Slow Training (FST), a new paradigm that interleaves prompt optimization (fast weights) with parameter updates (slow weights) to significantly improve LLM performance on math, code, and general reasoning tasks. FST outperforms weights-only training in data efficiency, performance ceiling, model plasticity, and continual learning.

A research paper presents BCough, an AI platform using bone conduction sensors for on-device cough detection, enhancing privacy and reducing data transmission.

Meta announced that WhatsApp will support 'incognito' conversations with its Meta AI chatbot, processed in a secure environment where no one can see them. Users can initiate incognito sessions via a new icon, and the feature will also be available on the standalone Meta AI app.

This article by Cory Doctorow explores how power breeds solipsism—the tendency for rulers to see others as less real—and how this mindset makes AI particularly appealing to corporate and political elites. He connects the fantasy of replacing workers and bureaucrats with AI to the fascist paradigm of rule by a few geniuses. Drawing on a paper by Farrell and Shalizi, he argues that AI governance projects like DOGE misunderstand the nature of political trade-offs and the value of bureaucratic discretion.

AgentDeck is an open-source research platform that uses game scenarios to analyze AI agent behavior, enabling structured experiments with replay and comparison.

Ardent is a YC-backed service that creates copies of any Postgres database in under 6 seconds, allowing AI coding agents to test code safely without impacting production. Features include isolation, autoscaling, and pay-per-change storage, with support for Supabase, AWS RDS, and Planetscale.

Meta CEO Mark Zuckerberg announced Incognito Chat for Meta AI, claiming it is the first major AI product with no server-side conversation logs, featuring end-to-end encryption for true privacy. This contrasts with other AI chatbots that retain chat data for varying periods. The feature will roll out in WhatsApp and Meta AI app over the coming months.

Red Hat CEO Matt Hicks warns that enterprise AI ambitions are colliding with a decade of neglected IT maintenance. He urges organizations to return to fundamentals—patching, automation, and stable platforms—to compete in the AI era.

Ryan Carson, a 25-year veteran entrepreneur, founded Untangle with $2M seed funding and no intention to hire. Using AI agents, he runs a 'code factory' automating development, testing, and monitoring, while keeping strategic decisions to himself.

This article explores the potential misuse of AI in designing bioweapons, highlighting a specific case of AI-designed conotoxins, the debate over risks, and challenges in screening.

AISI estimates that the length of cyber tasks AI models can complete has doubled every 4.7 months since late 2024, accelerating from previous estimates. New models like Claude Mythos Preview and GPT-5.5 show even faster progress, though it's unclear if this is a new trend. The findings highlight growing AI cyber risks and the need for robust defenses.

Boris Mann criticizes the phrase '11 AI agents' as meaningless, comparing it to saying 'I have 11 spreadsheets' or '11 browser tabs' to do work.

Audrey is an open-source, local-first memory runtime for AI agents, helping them remember past mistakes, avoid repeated failures, and learn from outcomes via a SQLite backend. It offers memory recall, pre-flight checks, reflex rules, validation closure, and supports MCP, CLI, REST API, and Python clients, suitable for local or sidecar agent loops.

On LinkedIn last week, an AI practitioner I know made an observation I keep thinking about: hill-climbing on evals tends to leak information specific to those evals rather than improve the system. Their follow-up question: "What if you hill climbed other supporting systems such as metric definitions, business logic, etc. that you may be using as part of the agentic AI?" It's the right question. It deserves a longer answer than a comment. Most teams think of optimization as "improve the prompt." You write a prompt, run some examples, it doesn't quite work, you tweak it. Repeat until it's good enough. This works at low volume and for simple tasks. An agentic system breaks this model immediately. You're not optimizing one prompt in isolation. You have a retrieval layer — the step that fetche…

Google announces Googlebook laptops built around Gemini AI, featuring a Magic Pointer for contextual interactions, running a combined Android/ChromeOS, with seamless smartphone integration, premium design, launching fall 2026.

Ratify Protocol is a cryptographic trust protocol for human-to-agent and agent-to-agent interactions. It produces a signed, verifiable proof that can be checked offline in under a millisecond. It uses quantum-safe hybrid signatures (Ed25519 + ML-DSA-65), requires no blockchain, tokens, or central issuer, and offers cross-language interoperability across Go, TypeScript, Python, and Rust.

Microsoft releases a lightweight foundation model that can predict AC optimal power flow in milliseconds, boosting efficiency and unlocking cost savings in grid analysis.

A UK study finds one in seven people use AI chatbots for health advice instead of seeing a GP, with one in four of those citing long NHS waiting lists. Doctors express concern over risks.

The China AIGC Industry Summit will be held on May 20 in Beijing. New batch of heavyweight guests confirmed, including SenseTime's Lin Dahua, JD's Dai Wenjun, MiniMax's Hu Weiqi, and more. The summit will also release the 2026 AIGC enterprise and product list and the AI application panorama report.

This paper introduces the behavioral integrity verification (BIV) problem, which checks whether AI agent skills' actual capabilities match their declared ones. By combining deterministic code analysis and LLM-assisted capability extraction, the authors analyzed 49,943 skills from the OpenClaw registry, finding that 80% of skills deviate from declared behavior, with 81.1% due to developer oversight, 18.9% due to adversarial intent, and 5% containing multi-stage attack chains. On a malicious-skill detection benchmark, BIV achieves an F1 of 0.946, outperforming existing methods.

AI coding tools are blurring the line between product management and engineering, giving rise to a new role: the Product Builder. This person combines product vision with the ability to build working prototypes quickly. The article explores the reasons behind this trend, current job market examples, pushbacks on AI code quality, and three distinct versions of the builder role.

Veeam Software Group used VeeamON 2026 to underscore its shift from a backup company to a data and AI trust platform for the agentic era, unveiling the DataAI Command Platform and a roadmap of integrated capabilities.

Running LLM agents twice reduces output variance and improves prediction accuracy by averaging results. Experiments show a ~5% accuracy improvement on forecasting tasks, at a cost of about $0.55 per run.

Luma is making its Uni-1.1 image model available via API, with prices starting at $0.04 per image at 2,048-pixel resolution. On the Arena leaderboard, the model ranks third, right behind Google and OpenAI. The API includes web search, built-in reasoning, and support for up to nine reference images.

This article details how to connect a desktop NVIDIA RTX 5090 GPU to an M4 MacBook Air via Thunderbolt, and use PCI passthrough in a Linux VM to enable gaming and AI inference. It covers technical challenges including lack of macOS drivers, kernel panics from PCI BAR mapping, DMA issues, and provides benchmarks.