Wiola: A Novel SLM Architecture Built from First Principles, Not Based on GPT or LLaMA
Wiola: A Novel SLM Architecture Built from First Principles, Not Based on GPT or LLaMA
A new arXiv paper presents Wiola, a Small Language Model architecture built entirely from first principles, sharing no structural lineage with GPT, LLaMA, Mistral, or Falcon. It introduces five independently novel components for more efficient lightweight language modeling, with detailed design rationale and experimental results.
Kara Uses Sliding-Window KV Cache to Speed Up Reasoning LLMs by 50%
Reasoning LLMs suffer from long chain-of-thought generation, causing massive KV cache overhead. Kara introduces a sliding-window KV cache compression method that reduces decoding latency by 50% and triples throughput on LLaMA-3.1-8B-Instruct. This approach can be directly applied to existing reasoning model deployments.
Beyond Next-Token: RLVR Trains LLMs to Use Atlassian APIs Correctly
LLMs are trained for next-token prediction, not for correct API execution. This paper uses RLVR to train LLMs on Atlassian workflows, achieving over 40% higher success rate than GPT-4 in correctly calling endpoints with nested arguments in order.
BPE Tokenization Creates Exploitable Safety Gaps in LLM Alignment
Research reveals BPE tokenizers fragment safety-critical words into sub-word pieces, allowing character-level perturbations to bypass LLM safety alignment. All three major closed-source models are affected: inserting spaces or special characters within safety words triggers prohibited outputs while prompts remain human-readable.
Spin-Weighted Spherical Harmonics Enable Scalable E(3)-Equivariant Networks
E(3)-equivariant networks excel in 3D atomistic modeling but are limited by O(L^6) complexity of CGTP. This paper replaces CGTP with Spin-Weighted spherical harmonics, reducing complexity to O(L^3) while maintaining accuracy on multiple molecular property prediction tasks, enabling scaling to larger systems.
From Approximation to Emergence: A New Theoretical Framework for Deep Learning
arXiv:2607.01311v1 proposes a unified, proof-oriented theoretical framework for modern deep learning. Starting from classical foundations like approximation and optimization, the study progressively delves into emergent phenomena, aiming to provide a complete theoretical explanation beyond single mathematical interpretations. It systematically traces the path from traditional approximation theory to modern emergent properties, using rigorous mathematical proofs to reveal intrinsic mechanisms in training and generalization of deep networks. This framework not only helps understand existing model behavior but also offers theoretical guidance for designing more efficient and interpretable deep learning architectures. Specifically, the paper validates its theoretical predictions on multiple standard benchmark datasets, demonstrating performance highly consistent with experimental data, laying a new foundation for theoretical advances in deep learning.
Google DeepMind and A24 Announce First-of-Its-Kind Research Partnership
Google DeepMind and independent film studio A24 announced a first-of-its-kind research partnership to explore AI applications in storytelling, visual effects, and film production. This marks the first deep collaboration between a top AI research lab and a major film company. Specific projects have not been disclosed.
Vercel Sandbox Adds FUSE Support for Mounting Remote Storage
Vercel Sandbox now natively supports FUSE-based filesystems, allowing developers to mount remote storage like S3 buckets and network filesystems as regular paths inside a running Sandbox. This enables streaming large datasets directly from object storage without local downloads.
Open Source AI Gap Map Launches, Indexing 70+ Models by Openness
Current AI, a non-profit with $400M in committed funding, launched the Open Source AI Gap Map to systematically index the global AI open source ecosystem. It covers 70+ models and tools, annotating openness levels across data, code, and weights to help developers identify truly open projects.
Course Sales Drop Two-Thirds: AI Coding Tools Disrupt Frontend Education Market
Noted frontend educator Josh W. Comeau reported his new course ‘Whimsical Animations’ sells only one-third of a typical launch, with existing courses also seeing significant drops. He attributes this largely to AI coding assistants enabling developers to achieve results without systematic learning, reducing demand for traditional tutorials.
Study: AI Saves Only 3% of Work Hours, Hardly Any Revenue Impact
A study on AI productivity ROI found that AI tools save only about 3% of work hours in real workplace settings, with almost none of these savings translating into measurable financial gains. The research questions the actual ROI of AI in enterprise production, noting minimal impact on bottom-line metrics.
AI Engineer World's Fair Ends with Debate on Loops and State of AI Engineering
The AI Engineer World’s Fair concluded with a debate on loops, a report on the state of AI engineering, and closing keynotes on what to build next. The event highlighted current trends and challenges in AI engineering, guiding developers on future priorities.
Claude Code v2.1.201 Improves Sonnet 5 Session Role Handling
Claude Code released v2.1.201, which removes the mid-conversation system role for harness reminders in Claude Sonnet 5 sessions. This change streamlines dialogue flow, reducing unnecessary context interference for a better interaction experience with the Sonnet 5 model.
OpenAI Codex Releases 0.143.0-alpha.35 Preview Version
OpenAI’s code completion tool Codex released version 0.143.0-alpha.35. As an alpha preview, this version includes various bug fixes and performance improvements not detailed in the changelog. Developers are advised to test in non-production environments.
Kagi Search Adds Coin Flip, AI Toggle in July 2 Changelog
Search engine Kagi released its July 2 changelog, adding a virtual coin flip feature for heads/tails decisions and an AI toggle that lets users enable or disable AI enhancements in search results with one click. The updates aim for a more flexible and fun search experience.
Claude Code Team: Let Fable Use Its Own Judgment for Better Testing
At AIE Fireside Chat, Claude Code team members Cat Wu and Thariq Shihipar shared a critical tip: in testing and other tasks, let models like Fable (and Opus) use their own judgment instead of dictating how they should work. This approach significantly improves testing quality and efficiency.
AI Coding Addiction Is Costing Engineers Their Core Skills
A LeadDev article warns that AI-assisted coding tools are fostering an addictive dependency among engineers. Citing a Hacker News discussion (45 upvotes, 37 comments), the piece notes that while AI rapidly generates code snippets and boosts initial productivity, long-term reliance degrades debugging abilities and code review skills while increasing technical debt. Specific data shows over-reliant developers spend 30% more time fixing bugs and see a 15% drop in code quality scores. The article cautions that this “efficiency illusion” may strip engineers of core competencies in tackling complex problems, with the entire industry bearing the consequences.
June Newsletter: Claude Fable 5, GPT-5.6, US Export Rules, GLM-5.2
Simon Willison’s June sponsors-only newsletter covers Claude Fable 5, GPT-5.6, US export restrictions’ impact on AI, GLM-5.2 as the new best open-weights model, the end of Tokenmaxxing trend, and updates on Datasette Apps and sqlite-u.
Instead of Banning AI, I Signed a Classroom Contract with My Students
In an article published on Science magazine’s website, an educator shares an alternative strategy to outright bans on AI: co-creating a classroom contract with students that defines the boundaries and ethical guidelines for using AI tools like ChatGPT in assignments and research. The author argues that outright bans often drive students to use AI secretly, while the contract model fosters consensus, preserving AI’s assistive value—such as brainstorming and grammar checks—while preventing plagiarism and over-reliance. After piloting the approach in a high school classroom, student engagement improved and no misuse was reported. The piece gained 68 upvotes and 74 comments on Hacker News, sparking widespread discussion on integrating AI into education.
Stop AI’s Confident Performance Act
This article examines the widespread phenomenon of ‘confident performance’ in AI, where developers and companies overstate the capabilities and reliability of AI models, leading to unrealistic user expectations. It argues that such hype misleads the public and risks a crisis of trust. Citing specific cases, the author shows how many AI systems—unstable on critical tasks—are still marketed as ‘all-in-one solutions.’ The discussion, drawing 221 upvotes and 237 comments on Hacker News, struck a chord with the audience. The piece calls for the industry to return to rationality, urging transparency about model limitations and failure cases to foster a healthier AI ecosystem.