Search the archive

AI EVALS ARE BECOMING THE NEW COMPUTE BOTTLENECK

INTRODUCING AUTOSP

AutoSP automates converting standard transformer training code into sequence-parallel code for long-context LLM training, integrated with DeepSpeed. It enables longer sequence training on multiple GPUs without significant runtime overhead, eliminating the need for complex manual code changes. AutoSP also offers an advanced activation-checkpointing strategy for better memory management, enhancing performance with minimal cost.

GRANITE 4.1 LLMS: HOW THEY'RE BUILT

Granite 4.1 LLMs utilize a dense, decoder-only architecture with models of 3B, 8B, and 30B parameters, trained on 15 trillion tokens and using a five-phase pre-training approach. The 8B model matches the performance of the previous 32B Mixture-of-Experts model through a multi-stage reinforcement learning pipeline focused on data quality. These models, designed for efficient, reliable enterprise use, demonstrate competitive instruction-following and tool performance while maintaining cost efficiency and stable usage.

→ Get your digital copy

LESSONS ON BUILDING MCP SERVERS

This post discusses how to make MCP toolchains work using a framework where the MCP servers do most of the work while models walk breadcrumbs. Models don't plan - they look at the conversation, scan the tool list, and grab whatever looks most probable. Making effective chains means making sure the server makes the next call blindingly obvious at every step.

LADIR: LATENT DIFFUSION ENHANCES LLMS FOR TEXT REASONING

LaDiR (Latent Diffusion Reasoner) is a novel reasoning framework that unifies the expressiveness of continuous latent representation with the iterative refinement capabilities of latent diffusion models for an existing LLM. The design allows efficient parallel generation of diverse reasoning trajectories, allowing models to plan and revise the reasoning process holistically. LaDiR consistently improves accuracy, diversity, and interpretability over existing autoregressive, diffusion-based, and latent reasoning methods. It is a new paradigm for text reasoning with latent diffusion.

MICROSOFT WORLD-R1 FOR 3D-CONSISTENT VIDEO GENERATION

World-R1 is a reinforcement learning framework that improves 3D consistency in video generation by leveraging feedback from 3D and vision-language models without modifying the base architecture.

RELIABLE DATA ANALYSIS AGENTS

DataPRM is an environment-aware process reward model that detects silent errors and better supervises data analysis agents, improving downstream performance and generalization across benchmarks.

ELON MUSK TESTIFIES HE WAS A ‘FOOL' TO FUND OPENAI

Elon Musk says he was a fool to back OpenAI when it was a nonprofit. Musk gave the startup $38 million of essentially free funding. OpenAI is now worth $800 billion. Musk has asked a court to unwind OpenAI's recent conversion to a for-profit entity and is seeking damages of more than $180 billion.

DARWINIAN SPECIALIZATION IN AI

The inference market is fragmenting because workloads are different. The model ecosystem has fragmented into latency tiers, multimodal models, and edge models. Each model type has different serving requirements, which fragments into infrastructure. The fragmentation creates room for several winners.

→ Get your digital copy

REVERSE ENGINEERING WITH AI UNEARTHS HIGH-SEVERITY GITHUB BUG

GitHub disclosed a high severity vulnerability, CVE-2026-3854, affecting GitHub Enterprise Server and other products, which allows remote code execution through manipulated git push options.

OPENAI CODEX SYSTEM PROMPT INCLUDES EXPLICIT DIRECTIVE TO “NEVER TALK ABOUT GOBLINS”

OpenAI appears to be fighting a new problem in its latest model where the model focuses on goblins in completely unrelated conversations.

BECOME A CURATOR FOR TLDR AI (3-5 HRS/WEEK)

TLDR is looking for an engineer/researcher at a major AI lab or startup to help write for 1M+ subscribers. Our curators have been invited to Google I/O and OpenAI DevDay, scouted for Tier 1 VCs, and get early access to unreleased TLDR products. Learn more.

AI AGENTS THAT BUILDS THEMSELVES

CrewAI built Iris, a Slack-native internal AI employee that writes code, files PRs, reviews teammates' work, and modifies its own codebase across CrewAI's engineering org.

DEEPMIND PROEVAL FOR GENAI EVALUATION

ProEval is a framework that reduces generative AI evaluation costs while identifying failure modes using surrogate models and transfer learning across benchmarks.

Launch on Framer Today

CLAUDE CAN NOW PLUG DIRECTLY INTO PHOTOSHOP, BLENDER, AND ABLETON

Anthropic has introduced new connectors for its AI chatbot Claude that allow it to integrate with major creative tools like Adobe Creative Cloud, Blender, Ableton Live, and others. These connectors let Claude access data, assist with tasks, and interact directly within these apps, aiming to streamline creative workflows and reduce repetitive work. The move builds on Anthropic's push into the creative industry, positioning Claude as a tool for faster ideation and broader capabilities rather than a replacement for human creativity.

FREEPIK REBRANDS AS MAGNIFIC: A BOOTSTRAPPED, PROFITABLE $230M ARR AI CREATIVE PLATFORM

Freepik has rebranded as Magnific, unifying its stock assets, AI upscaling, and other creative tools under one platform. The bootstrapped company has achieved $230 million in annual recurring revenue with over one million paying subscribers and 250 enterprise customers including BBC, Puma, and Amazon Prime Video. CEO Joaquín Cuenca built this profitable AI creative platform without any venture capital investment since its founding in 2010.

PEDOMETER++ 8 BRINGS FRIENDLY DESIGN REFRESH AND EXPEDITION MODE TO APPLE WATCH

Pedometer++ 8.0 centers on a complete redesign of its Apple Watch app, introducing a bolder, more colorful interface that highlights key daily metrics like steps, distance, and flights climbed with improved clarity. A major part of the redesign is the new workout picker, built with larger touch targets and simpler navigation to make starting activities faster and more intuitive, especially for users frustrated with Apple's default Workout app. While the update also adds Expedition Mode for better battery life, the primary focus is on making the app more visually engaging, easier to use, and better suited for quick interactions on the go.

USERS OWN THE PRESENT. YOU OWN THE FUTURE

Smart users often provide convincing but wrong solutions because they're trained to always have answers, especially executives from consulting or finance backgrounds. User research should focus on understanding underlying needs rather than stated wants - when someone says they want ice cream, they actually need to cool down, which opens up many more solution possibilities. Analytics alone can't protect teams from bad user research, as the same problems that affect interviews also impact how metrics are interpreted.

THE AI ECONOMY: FIVE ADOBE SNEAKS WORTH WATCHING IN 2026

Adobe showcased seven experimental AI prototypes at its annual Sneaks event. Only 30-40% historically make it to production. Project Face Off, the audience winner, simulates A/B testing by creating synthetic user personas that predict which creative variants will perform best in seconds rather than weeks. Project Test Kitchen reimagines AI image generation as a collaborative workspace where multiple designers can work together without chaos.

ADVANCED ICON DESIGN: DOTS

Dots in icon design should typically be slightly larger than the stroke weight to appear visually balanced, since matching them exactly often makes them look too small. This optical adjustment—common in type design—can be applied across icons, with flexibility to vary dot size or even shape depending on context and emphasis. Different elements within an icon set may require different dot sizes to feel right, especially when dots are a primary feature, reinforcing that visual balance matters more than strict geometric consistency.

PROFESSIONAL, BRANDED FILE DELIVERY

Kinet is a branded file delivery platform that allows agencies, freelancers, and professionals to share files through custom-branded portals featuring their own logos, colors, and messaging.

THE COLOR API

The Color API is a fast, modern tool for color conversion, naming, and scheme generation that accepts any valid color format and converts it to other formats.

STARWIND UI

Starwind UI provides 45+ animated, customizable, and accessible components for Astro websites.

THE NEW DESIGNER VALUE PROPOSITION: HOW TO REPOSITION YOUR DESIGN SERVICES IN THE AGE OF AI ART

Designers can successfully compete with AI by repositioning themselves as strategic directors who know when to use AI tools and when human expertise is essential. While clients turn to AI for speed, cost, and convenience on routine tasks, they still need human designers for strategic thinking, understanding nuances, and creating compelling value propositions. The key is shifting focus from being just an executor to demonstrating how human expertise provides long-term investment value that AI cannot replicate.

HOW LA28'S "UNAPOLOGETICALLY TYPE-FORWARD" APPROACH NAILED OLYMPIC BRANDING

The LA 2028 Olympics branding uses a bold, typography-led system to reflect the diversity of Los Angeles, replacing a single fixed logo with a flexible identity built around multiple versions of the letter “A.” This approach draws from the city's street signage and visual culture, making typography the central expression of the brand. By balancing a clear structure with room for variation, the system captures multiple voices while staying cohesive, showing how design can both represent and embody a city's cultural identity.

GOOGLE IS REDESIGNING ITS APP ICONS TO FIX A BIG PROBLEM

Google is redesigning its Workspace app icons — including Meet, Calendar, Docs, and Sheets — to address a longstanding complaint that they all look too similar.

LOGO DESIGN AFTER AI: HOW DESIGNERS CREATE POWERFUL BRAND LOGOS IN 2026

AI tools have transformed the initial stages of logo design by enabling faster concept generation and visual experimentation, but they cannot replace human strategic thinking and brand understanding.

PIA SALZER: A LOVE LETTER TO EVERYDAY AESTHETICS

Schedule a demo to learn more

THE REACT COMPILER AT EIGHTEEN MONTHS: THE ARC, THE DEBATES, AND WHAT'S NEXT

Eighteen months after React 19 shipped the stable compiler, its true impact is seen in eliminating bugs like forgotten dependencies and missing memos rather than delivering benchmark wins, though adoption in brownfield projects is hindered by legacy libraries that break the Rules of React. Open debates continue regarding whether those Rules should be a hard contract and the risk of 'use no memo' becoming technical debt, while future plans include finer per-component control, compiler-aware Server Components, a stable useEvent, React Native parity, and improved DevTools.

EVALUATING NETFLIX SHOW SYNOPSES WITH LLM-AS-A-JUDGE

Netflix developed an LLM-as-a-Judge system to score show synopses on four quality dimensions, anchoring it to a golden set of roughly 600 expert-labeled examples built through a rigorous calibration process. This approach, which uses judges and techniques like tiered reasoning and consensus scoring, achieves 83–92% accuracy and allows them to flag and fix weak synopses weeks before a show ships.

THE LAST SOFTWARE ENGINEER

As AI agents automate implementation, software engineering value is shifting from writing code to exercising judgment about what to build and why. Product engineering now requires humans to own implementation consequences and user impact, making strategic decision-making the most durable skill for engineers.

THE END OF "JUST ASK SARAH"

AI agents shift the role of docs from organizational memory for humans to execution context for machines. Unlike human engineers who can rely on unwritten institutional knowledge, agents only have their limited context window, forgetting information between sessions. This absence of durable intent leads agents to perpetuate patterns without understanding their underlying reasoning, resulting in "intent debt" and potentially incorrect decisions.

HOW TO BE SUCCESSFUL INTERVIEWING FOR BIG TECH

Traditional Big Tech interviews, often focused on LeetCode problems, inaccurately assess an engineer's true capability to ship working software. The interview process, like at Postman, should be designed to mirror the actual job, prioritizing real-world relevance, the use of modern tools including AI, and collaborative problem-solving.

ZED 1.0

Zed 1.0, a code editor, was reimagined with a "video game" architecture and a custom Rust-based GPUI framework to achieve high performance with multi-language support and deep AI-native capabilities.

OPENSPEC

OpenSpec is a spec-driven development framework that brings predictability and organization to AI coding assistants. It solves the issue of unpredictable AI output by creating a lightweight specification layer, making sure humans and AI align on requirements before code is written.

TWO FOUNDATION MODELS BUILT FOR AGENTIC CODING.

Poolside develops foundation models specifically for agentic coding, training models from scratch using its own data, infrastructure, and reinforcement learning. It has two main models: Laguna XS.2, an open-weight, lighter, and faster Mixture of Experts (MoE) model, and Laguna M.1, its most capable and largest MoE model.

COPY FAIL

"Copy Fail" (CVE-2026-31431) is a critical Linux kernel flaw that has existed since 2017. It allows unprivileged local users to gain root access via a simple Python script. This vulnerability poses a severe risk to multi-tenant environments like container clusters, requiring kernel updates or disabling the `algif_aead` module for mitigation.

I BENCHMARKED CAVEMAN AGAINST TWO WORDS

Benchmarks comparing the "Caveman" Claude code compression plugin against the simple prompt "be brief" found that the two-word prompt matched Caveman in both token reduction and output quality across most tested categories. While "be brief" is recommended for simple token reduction, Caveman is ideal for those needing predictable and structured output formats.

FASTCGI: 30 YEARS OLD AND STILL THE BETTER PROTOCOL FOR REVERSE PROXIES

FastCGI, a 30-year-old protocol, is still better than HTTP for reverse proxy communication due to HTTP's inherent vulnerabilities, like desync attacks and insecure handling of trusted headers.

DRAM CRUNCH: LESSONS FOR SYSTEM DESIGN

Due to rising DRAM costs and tightened supply, AI system design is shifting towards edge architectures that use lower-memory solutions or on-chip inference to achieve more cost-effective and power-efficient AI workloads.