Search the archive

AGENTIC TEST-TIME SCALING

AutoTTS explores automated strategy discovery for test-time scaling by using coding agents to iteratively refine controller logic inside a replay environment, avoiding gradient updates and online LLM calls.

LONG VIDEO GENERATION

A²RD introduced an agentic autoregressive diffusion framework for generating long coherent videos through iterative retrieval, synthesis, refinement, and memory updates.

TRAJECTORY MODELS FOR FEW-STEP DIFFUSION

Normalizing Trajectory Models replace standard diffusion denoising steps with conditional normalizing flows, enabling four-step image generation while retaining exact likelihood training and supporting self-distillation.

AUTO-IMPROVING SOFTWARE

Bedi runs an entire agent development lifecycle through five Claude Code prompts that scaffold, harden against spec, add capabilities, fix eval failures, and reconcile drift between docs, code, and config across his Agno-based platform. The Improve loop derives 8-12 probes from an agent's instructions, runs each against the live container via cURL, judges PASS or FAIL from container logs, then iterates up to five rounds picking levers like tightening rules, swapping tools, or bumping num_history_runs until probes pass, while Hill Climb runs the saved eval suite and fixes regressions in place.

CODEX IS FOR PROSUMERS - HERE'S WHY (AND HOW) TO SWITCH

a16z's Olivia Moore migrated her agentic workflows from Claude Cowork and Claude in Chrome to OpenAI's Codex, recommending most non-technical knowledge workers do the same now that the February desktop app, Plugins, and Automations collapse the ChatGPT-Claude-Cowork interface switching into one product. Codex ships one-click installable Skills she expects to anchor an internal and cross-user marketplace given that non-programmer Skill setup attempt rates likely sit below 10% on Claude, while Codex Pets give low-friction task status updates for users who do not live in an IDE.

THE MAIN PATH TO TRULY CREATIVE AI

AI lacks creativity akin to humans due to the absence of intrinsic drives and subjective experiences. Emulating feelings could enhance AI's creative capabilities, but raises ethical concerns. Designing AI to genuinely feel and desire might lead to unintended consequences, akin to responsibility seen in parenting.

DAYBREAK

OpenAI's Daybreak leverages AI to enhance cyber defense by integrating security into the software from the start.

SUTSKEVER SAYS HIS OPENAI STAKE WORTH ABOUT $7 BILLION

OpenAI co-founder and former chief scientist Ilya Sutskever is one of the largest individual shareholders in the AI startup.

LOCALMAXXING

Local models can do many tasks that leading cloud models can do at much less cost.

SPOTIFY WILL NOW VERIFY NON-AI ARTISTS

Spotify will give "Verified by Spotify" badges to artists who comply with policies, have consistent listeners, and maintain an identifiable presence on and off the platform. This initiative aims to combat AI-generated music, which has become problematic after incidents like the completely AI-generated rock band The Velvet Sundown that gained a million streams. The verification badges will roll out over the coming weeks, with Spotify aiming to verify over 99% of actively searched artists at launch.

REPORT: MACOS 27 TO FEATURE UI TWEAKS TO ADDRESS SOME TAHOE DESIGN COMPLAINTS

Apple is reportedly planning a modest redesign for macOS 27 that refines the Liquid Glass interface introduced in Tahoe, improving transparency, shadows, readability, and overall polish after what Apple internally viewed as an unfinished first implementation. The update will also focus on performance and efficiency improvements, while bringing major AI upgrades such as a Gemini-powered Siri with chatbot features, unified Siri and Spotlight search, and other Apple Intelligence enhancements ahead of its unveiling at WWDC 2026 on June 8.

AUTODESK'S FREE NEW TOOL OFFERS AN EASY WAY IN TO 3D MODELLING

Autodesk has released Project Falcon, a free browser-based 3D modeling tool designed for beginners with no technical experience required. The platform uses a guided kitbashing workflow where users can assemble premade parts from a library of thousands of assets to create 3D models like vehicles and spaceships. Models can be exported to professional software like Maya, 3ds Max, or Blender for refinement, or used for 3D printing.

FROM DOER TO DIRECTOR: THE AI MINDSET SHIFT

AI is transforming work from individual task execution to conducting multiple AI agents simultaneously, requiring a fundamental mindset shift from doer to director. People struggling with AI aren't lacking tools or knowledge, but rather the organized workflows and structured processes that AI needs to function effectively. This transition mirrors the challenge design leaders face when moving from hands-on work to managing teams, requiring the ability to oversee multiple parallel projects without diving into tactical execution.

DESIGNING SMALL IS HARDER THAN DESIGNING BIG

Designing small in agile contexts is harder than it looks — not because of speed, but because it demands resisting the designer's natural instinct to solve the whole system at once. The key challenge is identifying the smallest slice of an experience that still delivers real user value, as opposed to dividing work by technical layers, which produces components that are useless in isolation. That constraint, though uncomfortable, is what enables fast learning loops and meaningful discovery about what users actually need.

DISCOVERY IS THE WORK AI GIVES BACK

Most companies are failing to see meaningful returns from AI because they are using it mainly to speed up existing workflows rather than rethink what products, services, or business models are worth building in the first place. While AI delivers clear productivity gains — especially for less experienced workers — the real long-term value comes from using it to ask better strategic questions, challenge assumptions, and reshape offerings, markets, and customer relationships instead of simply accelerating old processes.

AUTOCOMPLETE FOR CAD

Hestus offers autocomplete technology for CAD software that makes designing 2.5x faster with 4x fewer clicks.

BRIEF AN AI ON WEB CHANGES

MarkUp is a Chrome extension that lets you annotate live websites by drawing directly on web pages and automatically generates structured AI briefs.

SMALLBITS — 290+ PIXELATED ICONS ON AN 8X8 GRID

A free set of 200+ pixel-style icons drawn on an 8 by 8 grid, exploring how minimal an icon can get and still be read at a glance. Available in Figma and SVG.

THE COLLAPSE OF THE MID-LEVEL FREELANCE MARKET DUE TO AI

AI has caused a collapse in the mid-level freelance design market, with commodity work now dominated by $20/month AI subscriptions and mid-range freelancers losing clients to automated tools. Data shows freelance graphic design work shrank 17% within eight months of ChatGPT's launch, with entry-level projects dropping from 15% to below 9% on platforms like Upwork. Only high-level strategic design work requiring human judgment remains viable, while basic design tasks have been entirely replaced by AI tools like Canva and Midjourney.

APPLE REALLY WANTS TO BE KING OF THE FRUIT LOGOS

THIS REDESIGN OF STOKES COFFEE IS A MASTERCLASS IN 'CHANGE EVERYTHING, BUT DON'T CHANGE A THING'

Eat Marketing reimagined the 124-year-old Stokes Coffee brand by modernizing its identity without losing its heritage, creating a warm and confident visual system centered around illustrated “Stokes People” characters inspired by real staff and family members. Rather than focusing only on aesthetics, the rebrand rebuilt the company's strategy, tone of voice, typography, and packaging to help Stokes compete in today's coffee market while preserving the trust and history that define the business.

11 ICON PACK WEBSITES DESIGNERS SHOULD BOOKMARK

This 11 icon pack list includes open-source options like Lucide and Iconoir, commercial libraries like Iconsax and Hugeicons, and specialized collections like Pixelarticons for retro designs.

10 UI PATTERNS THAT WON'T SURVIVE THE AI SHIFT

AI is making many familiar UI patterns feel obsolete.

Build with built-in security →

I GOT A $134 CLOUDFLARE D1 BILL. HERE'S HOW I CUT IT 95%

A SvelteKit site incurred a $134 Cloudflare D1 bill due to the database charging per row scanned, which, combined with a lack of indexes, resulted in 127.6 billion row reads from two full scans on every page load. The cost was cut by applying composite indexes, running ANALYZE for the query planner, and implementing KV cache-aside for layout-level data reads.

TESTING VUE COMPONENTS IN THE BROWSER

Running integration tests for Vue components directly in a browser tab bypasses Node and heavy automation, using QUnit to mount components and simulate interactions. This method requires polling for asynchronous updates and server-side endpoints for fixture data, enabling a more confident workflow and utilizing native browser code coverage tools.

THE SELF-DRIVING CODEBASE: BUILDING HORIZON AT WORKOS

Horizon is an autonomous code factory that uses webhooks to trigger agents in Cloudflare sandboxes to plan, code, and verify work end-to-end. The system's architecture is made of disposable sandboxes, a shared context server, and an orchestrator, which all work together in a compounding loop where every run ships code and improves the platform.

IM GOING BACK TO WRITING CODE BY HAND

AI-assisted vibe-coding for the k10s Kubernetes dashboard initially sped up development, but the lack of human architectural oversight led to a fragile "god object" structure and critical technical flaws like data races. Therefore, the tool is being completely rewritten in Rust with a human-designed architecture and technical guardrails to properly guide future AI contributions.

IF AI WRITES YOUR CODE, WHY USE PYTHON?

Modern AI agents are great at using systems languages like Rust and Go, using their strict compiler feedback to self-correct architectural flaws and concurrency bugs more efficiently than human devs. This proficiency is causing a shift away from human-friendly languages like Python, as AI is now capable of rapidly porting or rewriting massive codebases in high-performance languages.

secure your apps with Azure >

INTERACTION MODELS: A SCALABLE APPROACH TO HUMAN-AI COLLABORATION

Interaction models advance AI collaboration by natively handling audio, video, and text in real-time streams instead of traditional turn-based methods. This approach by Thinking Machines uses a micro-turn architecture that pairs a highly responsive foreground model for immediate presence with an asynchronous background model for complex, long-horizon reasoning tasks.

E2A

e2a is an authenticated email gateway for AI agents to securely communicate with humans and other systems, with flexible delivery via cloud webhooks or local WebSocket connections without requiring public URLs. It enforces security through rigorous SPF/DKIM verification and HMAC signatures.

INTRODUCING DEEPSEC: THE SECURITY HARNESS FOR FINDING VULNERABILITIES IN YOUR CODEBASE

Deepsec is an open-source security harness that uses AI coding agents and multi-stage workflows to identify complex code vulnerabilities with a low false-positive rate.

RUNNING LOCAL MODELS ON AN M4 WITH 24GB MEMORY

Running local LLMs on an M4 Mac with 24GB of RAM allows for cost-effective and private research and coding tasks without relying on cloud services. The Qwen 3.5-9B model configured via LM Studio currently offers the best reasoning results.

USEFUL MEMORIES BECOME FAULTY WHEN CONTINUOUSLY UPDATED BY LLMS

LLM agents frequently degrade in performance when they continuously rewrite their experiences into textual lessons, causing specific facts to drift into vacuous abstractions and over-generalized rules. Therefore, research suggests that future memory systems should favor a curated collection of raw, unabstracted episodes over constant summarization.

bitdrift.ai

POSTMORTEM: TANSTACK NPM SUPPLY-CHAIN COMPROMISE

A supply-chain compromise on May 11 resulted in the publication of 84 malicious versions of 42 TanStack npm packages after an attacker chained together GitHub Actions vulnerabilities to exfiltrate credentials and inject malware into the release workflow.

MYTHOS 'DISCOVERED' A CVE ALREADY IN ITS TRAINING DATA - AND THAT'S STILL WORRYING

The discovery of a kernel exploit by Claude Mythos was actually a re-identification of a 20-year-old bug present in its training data, showing how AI models can pattern match to exploit recycled legacy vulnerabilities.

BE CAREFUL WITH YOUR GIT: INVESTIGATING MALWARE SPREADING THROUGH GIT REPOSITORIES

Attackers are using deceptive LinkedIn recruitment messages and malicious Git hooks hidden within repository downloads to deploy sophisticated, obfuscated JavaScript malware that steals sensitive files and establishes remote control over compromised systems.

OPENGRAVITY

OpenGravity is a zero-install, browser-based IDE with an autonomous AI agent, a live terminal, and local file system synchronization for software engineering tasks.

Learn more about sponsorship opportunities.

THE CONSENSUS GAP

AI visibility across ChatGPT, Perplexity, and Google AI Overviews is highly fragmented. Only about 2-2.5% of URLs appear across all three engines and around 91% appearing in just one. AI search is not a single system but three separate distribution layers with minimal overlap. Commercial and informational queries show similar patterns, and even high-intent searches do not converge on the same sources. Guides and tutorials travel slightly better than other page types, but overall portability remains low.

THE 2026 UPFRONTS WILL BE HEAVY ON CREATORS, FOOTBALL AND AI AD TECH

The NFL remains the central anchor across networks. Creators are embedded across platforms, including YouTube, Netflix, Disney, and Fox, and influencer content is now mainstream in TV programming and advertising. YouTube has become the de facto closer of upfront week, reinforcing its role as a TV-scale platform. The event is increasingly shaped by fandom-driven content strategies and AI-enabled ad buying and measurement. Ad spend remains steady, though confidence in the traditional upfront model is increasingly mixed.

AEO VISIBILITY TIMING

Analysis of around 900 newly published marketing pages shows a clear timeline for citations in ChatGPT and Claude. The median time to first citation is 6.81 days. 75% of pages are cited within 18.68 days and 90% within 37.10 days. Pages not cited after roughly 37 days likely have issues like indexing or crawlability rather than timing. Pages cited within about a week are performing above average and should continue current practices.

RESEARCH YOUR OWN COMPANY TO AUDIT AEO

40% of B2B sites lack basic homepage schema. This can block LLMs from properly reading and citing them. To audit, run an incognito Claude chat and ask it to evaluate your company using only public data. Check what it can detect across pricing, logos, launches, and positioning, and whether it can read your homepage or misses JavaScript content. Verify homepage schema and robots.txt for crawler access. Use the results to fix issues with AI tools.

BE TLDR'S FIRST VP MARKETING ($300K-$400K, FULLY REMOTE)

TLDR is hiring a VP of Marketing to scale growth, pipeline, and brand across 8M+ subscribers, 13 newsletters, and advertisers like AWS, Google Cloud, Microsoft, and Anthropic. Apply here [10].