Reading
What's worth your attention.
A hand-picked feed of the writing, analysis, and reporting shaping how we think about AI this month.
- ResearchMercor·June 3, 2026
AI can't read an investor deck
A close look at where frontier models still fall apart on finance tasks—parsing decks, reconciling figures, and reasoning across exhibits—and what the failure modes reveal about agent design.
Read at Mercor → - AI PolicyNew York Times·June 2, 2026
Trump Signs Executive Order Granting Oversight of A.I. Models
President Trump signed an executive order directing tech companies to submit new frontier AI models for government oversight before public release—a sharp pivot from the administration's prior hands-off stance. The order establishes a voluntary framework, with the NSA playing a central role in vetting national-security risks.
Read at New York Times → - InfrastructureFal·May 30, 2026
fal and AWS: Building for the Next Phase of Generative Media
fal announces a strategic partnership with AWS to scale infrastructure for the next wave of generative media workloads.
Read at Fal → - Voice AIElevenLabs·May 28, 2026
Introducing Dubbing v2
A new generation of AI dubbing with sharper lip-sync, preserved emotion across languages, and cleaner multi-speaker handling—aimed at studio-grade localization at scale.
Read at ElevenLabs → - AgentsMistral·May 28, 2026
Vibe gets to work
Mistral rebrands Le Chat to Vibe, a unified agent for long-horizon productivity and coding with Work Mode, Code Mode, and a new VS Code extension.
Read at Mistral → - Developer ToolsMistral·May 28, 2026
Introducing Search Toolkit
An open-source, composable framework for building production search and RAG pipelines that unifies ingestion, retrieval, and evaluation.
Read at Mistral → - Economic ResearchAnthropic·May 27, 2026
Coding agents in the social sciences
Anthropic's Economic Research team examines how coding agents are reshaping empirical workflows in the social sciences—from data cleaning to replication studies.
Read at Anthropic → - ResearchMistral·May 27, 2026
Introducing physics AI at Mistral
Mistral brings Emmi AI into the fold to build frontier models that predict physical system behavior, targeting aerospace, energy, and chip design.
Read at Mistral → - Voice AIElevenLabs·May 26, 2026
Introducing Music v2
Studio-quality generative music with longer tracks, tighter prompt control, and improved instrumental fidelity. A meaningful step closer to usable production audio.
Read at ElevenLabs → - ResearchAnthropic·May 22, 2026
Project Glasswing: An initial update
A first look at Project Glasswing, Anthropic's effort to make model behavior more transparent and inspectable end-to-end.
Read at Anthropic → - AgentsMistral·May 22, 2026
Remote agents in Vibe. Powered by Mistral Medium 3.5
Mistral Medium 3.5, a 128B open-weights dense model, powers remote coding agents that run asynchronously in the cloud and notify you when done.
Read at Mistral → - AgentsSequoia Capital·May 18, 2026
The State of AI Agents in 2026
A wide-ranging look at where autonomous agents are actually working in production—and where they still fail. Sales, support, and coding lead; longer-horizon planning lags.
Read at Sequoia Capital → - AI SafetyFal·May 15, 2026
Building long-term trust in a world where creation moves at the speed of thought
fal's Head of Trust & Safety on building safeguards for a world where AI-generated content is instant, ubiquitous, and high-stakes.
Read at Fal → - PolicyAnthropic·May 14, 2026
2028: Two scenarios for global AI leadership
Anthropic's policy team sketches two plausible 2028 worlds—one where democratic nations lead frontier AI, one where they don't—and what each path implies for governance.
Read at Anthropic → - Open SourceLatent Space·May 12, 2026
Open-source models are quietly catching up
Benchmarks suggest open-weight models from Mistral, Meta, and Alibaba now match GPT-class performance on most reasoning tasks at a fraction of the cost.
Read at Latent Space → - Vertical AIStratechery·May 8, 2026
Vertical AI is eating professional services
Why Harvey (law), Abridge (medicine), and a wave of domain-specific startups are winning enterprise contracts that horizontal chatbots cannot.
Read at Stratechery → - AlignmentAnthropic·May 8, 2026
Teaching Claude why
New alignment research on reducing agentic misalignment by training Claude to reason explicitly about why a behavior is appropriate, not just what action to take.
Read at Anthropic → - InterpretabilityAnthropic·May 7, 2026
Natural Language Autoencoders: Turning Claude's thoughts into text
Claude talks in words but thinks in numbers. Anthropic's interpretability team trains the model to translate its internal activations into human-readable text.
Read at Anthropic → - Foundation ModelsArs Technica·April 29, 2026
Comparing the frontier: Claude 4, GPT-5, and Gemini 3
A hands-on comparison of the three leading frontier models across coding, long-context reasoning, and multimodal tasks. Each has clear strengths.
Read at Ars Technica → - Voice AIElevenLabs·April 29, 2026
Introducing ElevenMusic
A consumer-facing app built on the Music model—Explore, Library, Studio, and Artist live sessions. ElevenLabs' bet that generated music needs a product, not just an API.
Read at ElevenLabs → - Infrastructurea16z·April 22, 2026
The new economics of inference
Inference cost per token has fallen ~95% in 24 months. The piece argues this unlocks a new generation of always-on AI products previously unaffordable to run.
Read at a16z → - Generative MediaFal·April 20, 2026
Introducing PATINA
A new approach to AI-generated materials that yields production-ready textures for traditional CGI pipelines—no baked-in highlights, occlusion, or perspective.
Read at Fal → - Developer ToolsThe Pragmatic Engineer·April 15, 2026
Why every dev tool is becoming an AI dev tool
From Cursor to Linear to GitHub, embedded AI is no longer a feature—it's table stakes. A look at how teams are restructuring around AI-native workflows.
Read at The Pragmatic Engineer → - Voice AIElevenLabs·April 9, 2026
Enterprise voice AI, deployed locally
On-device and on-prem deployment of ElevenLabs' voice models for regulated industries that can't send audio to the cloud. Same models, private infrastructure.
Read at ElevenLabs → - MediaThe Verge·April 4, 2026
Generative video crosses the uncanny valley
With Sora 2, Runway Gen-4, and Veo 3, short-form generative video is now genuinely indistinguishable from live action in many contexts. The creative implications are immense.
Read at The Verge → - ResearchMIT Technology Review·March 27, 2026
Inside Anthropic's bet on interpretability
A profile of the team trying to actually understand what's happening inside Claude's weights—and why that may be the most important AI safety problem.
Read at MIT Technology Review → - CompanyMercor·March 26, 2026
Introducing Mercor Enterprise AI
Mercor's pitch for bringing expert-trained AI into the enterprise: vertical agents built on curated professional data rather than generic web scrape.
Read at Mercor → - ResearchMercor·March 24, 2026
Introducing the AI Productivity Index for Software Engineering
APEX-SWE, a new benchmark built with Cognition, measures whether frontier models can actually handle real software engineering work—not toy tasks. Early results suggest a wider gap than leaderboards imply.
Read at Mercor → - Developer ToolsFal·March 20, 2026
Connect your AI to 1,000+ models with the fal MCP Server
A hosted MCP endpoint that lets any AI assistant search, run, and chain 1,000+ generative models directly from a conversation—no SDK, no docs.
Read at Fal → - Societal ImpactsAnthropic·March 18, 2026
What 81,000 people want from AI
The largest multilingual qualitative study of its kind: nearly 81,000 Claude.ai users share how they use AI, what they hope it will unlock, and what they fear.
Read at Anthropic → - EnterpriseMistral·March 17, 2026
Introducing Forge
A system for enterprises to train frontier-grade AI models grounded in proprietary knowledge, bridging the gap between generic models and institutional context.
Read at Mistral → - CompanyMercor·March 3, 2026
Organizing human intelligence to power the AI economy
Mercor's thesis on why the next leg of AI progress depends on structured access to domain experts, and how they're building the marketplace to deliver it.
Read at Mercor → - ResearchMercor·March 1, 2026
Generalization results from training on the APEX-Agents dev set
Training on roughly 2,000 expert-authored tasks meaningfully improved tool use and professional reasoning across held-out domains. Evidence that small, high-quality datasets still move the needle.
Read at Mercor → - ResearchMercor·February 24, 2026
Scaling data leads to SOTA legal performance on APEX-Agents
Adding more expert-labeled legal tasks—rather than a bigger base model—delivered state-of-the-art performance on the legal slice of APEX-Agents. A datapoint for the data-over-scale camp.
Read at Mercor → - Video GenerationFal·February 10, 2026
Kling 3.0 is Now Available on fal
Kling 3.0 brings a new state-of-the-art video and image stack to fal on day zero, built for structured storytelling rather than isolated clips.
Read at Fal → - Voice AIElevenLabs·February 2, 2026
Eleven v3 is now generally available
The most advanced text-to-speech model exits alpha—expressive prosody, multi-speaker scenes, and stronger emotional range across dozens of languages.
Read at ElevenLabs → - ResearchMercor·January 23, 2026
Expert data drives model performance
An argument, with results, that curated expert annotations outperform raw scale on professional tasks—and that the bottleneck for vertical AI is increasingly sourcing, not compute.
Read at Mercor → - Video GenerationFal·January 15, 2026
LTX 2.0 is Now Available on fal
LTX 2.0 brings next-level open-source text-to-video and image-to-video generation with cinematic control, live on fal day zero.
Read at Fal → - Voice AIElevenLabs·January 9, 2026
Introducing Scribe v2
A new speech-to-text model with improved multilingual accuracy, speaker diarization, and realtime transcription. Closes the loop on ElevenLabs' end-to-end voice stack.
Read at ElevenLabs → - Foundation ModelsMistral·December 2, 2025
Introducing Mistral 3
Mistral 3 includes state-of-the-art small dense models (3B, 8B, 14B) and Mistral Large 3, a 675B sparse mixture-of-experts, all under Apache 2.0.
Read at Mistral →