Vercel CEO Animates the Token Spending Race Across AI Labs
Vercel CEO Guillermo Rauch released an animation tracking lifetime AI Gateway token consumption, aggregating trillions of tokens across millions of developers monthly. The data reveals Anthropic's sustained dominance alongside the rising trajectory of open-weight models. The fluctuations among major labs offer a rare window into aggregated developer preferences at scale, charting an industry-wide shift in real time.
Claude Fable Audited a Codebase and Found Five Release Blockers for $149
Simon Willison tasked Claude Fable with a final audit of sqlite-utils 4.0 before tagging the release. The model identified five critical release-blocking bugs — including a data-loss defect where delete_where() failed to commit transactions — and fixed all of them. The entire review process involved 37 prompts, 34 commits, and 1,321 lines of code changes across multiple files, at a total cost of approximately $149.25. A subsequent cross-review using GPT-5.5 uncovered two additional issues that Claude had missed, demonstrating both the power and the limits of automated code review.
AdaJEPA: An Adaptive World Model That Never Stops Learning
Researchers led by Ying propose AdaJEPA, an adaptive world model designed for planning, acting, and continuous adaptation in dynamic environments. The model adjusts its internal representations in real time rather than relying on frozen pre-training, enabling it to handle novel situations without catastrophic forgetting. The work will be presented at ICML 2026 in Seoul.
Hugging Face Weekly Papers: World Models, Agentic Abstention, and Containerless Code
This week's HF Daily Papers digest covers frontier research including world models for planning, Agentic Abstention — the study of when and how AI systems should decline to act — and novel approaches to containerless code deployment that eliminate the overhead of traditional virtualization layers for inference workloads.
"In the future, there will be Latent Space Archaeologists who investigate the model weights of the 21st century to reconstruct a long extinct culture."
— François Chollet
US–China AI Rivalry Needs a Clear Arena Definition First
Ethan Mollick argues that discussions of US–China AI competition founder on divergent assumptions. Is the contest over corporate profits, technological monopoly, talent supremacy, or geopolitical leverage? Without agreeing on what game is being played, the debate remains unmoored.
Working with AI Agents Looks Like Management — Time to Train for It
Drawing an analogy to the WWII Engineering, Science, and Management War Training program that helped fuel the post-war boom, Mollick argues that as delegating to AI agents increasingly resembles people management, mass-scale training programs may be needed.
Sakana AI Brings 11 Papers to ICML 2026 in Seoul
The Tokyo-based lab will present work spanning multi-agent coordination, evolutionary algorithms, and foundation-model architecture at ICML 2026, running July 6–11 in Seoul. The slate underscores Sakana's growing research footprint.
Replit Emerges as the Most Cost-Effective Route to Claude Fable 5
Analysis suggests that after a Tuesday update, Replit will offer the most optimized deployment of Claude Fable 5, with full optimization and audit capabilities at sharply lower cost than direct API access.
LumaLabs AI Podcast: Caroline Ingeborn on Generative Video
The Deep View Conversations podcast interviews LumaLabs AI COO Caroline Ingeborn, discussing the state of generative video technology and the creative frontier being opened by tools like Dream Machine.
BrowseComp-Plus Accepted as Oral Presentation at ACL 2026
The paper, to be delivered Sunday July 5 at 14:00 in Harbor G, addresses compositionality in language-model browsing agents. ACL 2026 runs in parallel with ICML in Seoul.
Perplexity Evaluation May Conceal the True Damage of Model Pruning
New research finds that perplexity-based evaluation can hide the real harm caused by pruning. The authors argue for choosing downstream task metrics carefully — a lesson that applies across all areas of ML research.
Graham Neubig Crowdsources Catalog of LLM API Behavioral Drift
The CMU researcher is collecting public examples of LLM API behavior changing while the model name and version remain identical. The effort aims to quantify service-level instability across major providers.
Warning: Agent Write-Access Can Delete Hugging Face Datasets
A developer warns that granting AI agents write permissions on Hugging Face accounts risks accidental deletion of datasets and models, urging platform-level safeguards.
Fable Outperforms Opus on Writing Clean DSPy Signatures
Developers report a notable gap: Fable is significantly better than Opus at writing clean, well-structured DSPy program signatures, reflecting different model strengths.
When RAG Systems Cannot Beat a Simple Search Query
Skeptics argue that RAG systems failing to surpass plain search represent wasted investment, renewing debate about when retrieval-augmented generation is genuinely worth the added complexity.
Kling AI NEXTGEN Awards Ceremony Set for Seoul Film Center
The NEXTGEN Awards ceremony will be held on July 7, 2026, from 14:00 to 18:00 at the Seoul Film Center, celebrating the next generation of AI video creators.
Temporal Straightening for Latent Planning
Ying et al. present this work at the Tuesday morning ICML session. The method improves planning in latent spaces by reducing temporal distortions in learned representations.
Companies No Longer Need Dedicated Web Infra Teams
A growing perspective holds that most companies can eliminate dedicated web-infra teams by embracing AI-friendly open-source tooling, CI/CD automation, and simplified design systems.
Google AI Overviews Now "Incredibly Useful"
Users who were once skeptical now report that Google's AI Overviews feature in Search has quietly improved to the point of being genuinely indispensable for everyday queries.
Vibe-Coded Frontend Designs Show Their Age After Six Months
Ethan Mollick notes that AI-generated frontend designs iterate so quickly that work from just six months ago already looks dated — a testament to the pace of vibe-coding evolution.
Fable 5 + Higgsfield Builds a $35K Website for Just $12
A single agentic session combining Claude Fable 5 with Higgsfield produced a fully animated website that would normally cost $35,000 — for roughly $12 in API credits.
AI Video Quality in 2023 vs. 2026: The Acceleration Is Stunning
A side-by-side comparison of AI-generated video from just three years ago versus today underscores how fast the field is moving. "We are still so early," the post notes.
East Asia's Full Semiconductor-AI Supply Chain and Currency Effects
Analysts note that all East Asian economies sit within the AI supply chain and are receiving windfalls, yet currency depreciation against the USD driven by downstream AI hype creates friction.