Gemini 3.7 Flash Release Puts Google Back on TopPath Integral Unifies Flow, Diffusion, and Adversarial ModelsMARCH: Content-Routed State Anchors Enable Linear-Complexity Long-Context MemoryLoKiFormer: Locality-Aware Attention with Decoupled Memory Boosts LLM Pretraining EfficiencyLinearKV: Single Cached State Accelerates Position-Independent Caching in Hybrid LLMsWorld-Model-Driven AutoResearch Agents Close the Loop Without Human InterventionGoogle Makes Private AI Practical with Homomorphic EncryptionVercel CDN Adopts ECH to Fully Encrypt HostnamesGitHub Agent Apps Unify Software Delivery WorkflowSAPO: Segment-Level Prompt Optimization Improves One Behavior Without Hurting OthersSkip Classification, Let the LLM Hallucinate TagsAI by Hand: A Hands-On Guide to Manual AI ComputationAI Model Atlas: Visualizing ML Model Populations as Interactive 3D GraphsHashAgent: Share AI Agents as URLs, Running Locally via WebGPUWhat Is a Personal Agent? Ben ExplainsCommentary: Intellectual Arrogance of AI Labs Becoming a LiabilityClaude Code v2.1.233 Adds GitLab MR and Identity ForwardingOpenAI Codex 0.148.0-alpha.18 ReleasedGemini 3.7 Flash Release Puts Google Back on Top

Google released Gemini 3.7 Flash, marking a strong return to the forefront of AI model capabilities. Latent Space’s report notes significant breakthroughs in performance that could reshape the competitive AI landscape. Following a period of lower performance for the Gemini series, this update is seen as a powerful comeback. Specific technical details and benchmark scores have yet to be published, but the industry widely regards Gemini 3.7 Flash as a formidable competitor to GPT and Claude.

A new arXiv paper proposes unifying generative modeling as a path integral, in which flow-based, diffusion-based, variational, and adversarial models arise as different evaluation principles for a single master action. Using the Martin-Siggia-Rose-Janssen-de-Dominicis (MSRJD) formalism, the paper offers a unified theoretical perspective on the commonalities of different generative models. This finding could guide the design of a new generation of more efficient and versatile generative AI models.

MARCH introduces content-routed state anchors to replace the growing key-value cache in Transformers with fixed-size memory states. This reduces quadratic training complexity to linear while preserving long-context retrieval. The architecture significantly cuts training and inference costs for LLMs handling ultra-long documents.

LoKiFormer tackles two pretraining inefficiencies in LLMs: self-attention’s lack of locality bias and growing KV cache. It introduces locality-aware attention for local dependencies and decouples long-term knowledge into a separate memory module. Experiments show faster pretraining loss reduction and better long-sequence performance compared to standard Transformers, offering a new path for efficient LLM training.

Existing position-independent caching methods target full-attention models, relying on token-indexed KV caches for chunk matching and concatenation. LinearKV, designed for hybrid LLM architectures, requires caching only a single state to enable position-independent reuse, eliminating per-token KV indexing. It notably reduces cache memory and matching overhead, boosting serving throughput for hybrid models.

AutoResearch agents aim to let LLMs independently handle the full research loop from experiment design to implementation and analysis, but struggle with inefficient trial-and-error. A new approach trains a world model of the environment, letting agents simulate experiments virtually before executing them. On small benchmarks, it improves average success rate by 30% over existing agents, bringing AI-driven research automation closer to practical use.

Google published a blog post detailing advancements in homomorphic encryption, making ‘private AI’ practical. Homomorphic encryption allows computation directly on encrypted data without decryption, thereby protecting user data privacy during AI inference. Google’s post indicates the technology has moved from theory to practice, with potential key applications in privacy-sensitive fields like healthcare and finance. The post scored 261 points and 160 comments on Hacker News.

Vercel announced support for Encrypted Client Hello (ECH) on its CDN for domains managed by Vercel DNS. ECH encrypts the Server Name Indication (SNI) in the TLS handshake, the last part of an HTTPS connection that revealed the destination hostname. With ECH, network observers only see a connection to Vercel, not the specific site visited. The feature is now fully available across all domains managed by Vercel DNS.

GitHub Blog published a post detailing how four GitHub agent apps help developers scope, secure, roll out, and ship a feature across the SDLC—all without leaving GitHub. These apps cover four key phases: scoping, security, rollout, and delivery. Developers can invoke these agents directly within the GitHub interface to automate workflow management from requirements to launch, significantly reducing the need to switch between multiple tools.

Automatic prompt optimization often rewrites entire prompts monolithically, fixing one behavior while breaking others. SAPO decomposes prompts into role, context, task, and output format segments, optimizing only the problematic segment. It outperforms monolithic optimization on instruction-following benchmarks while preserving previously correct behaviors, offering a precise alternative for prompt tuning.

A popular Hacker News post recommends ‘AI by Hand’, a tutorial demonstrating how AI models work through manual computation. The project breaks down complex neural network calculations into concrete steps that can be done by hand, helping learners deeply understand the underlying logic of AI. The post scored 184 points and 14 comments on Hacker News, widely regarded as an educational and impactful way to demystify AI.

AI Model Atlas is a visualization tool that maps relationships between large populations of machine learning models as interactive 3D graphs. Users can explore the landscape on Cosmograph, zooming and clicking to inspect individual models and their similarity to neighbors. It gives researchers an intuitive view of model evolution, revealing clusters and gaps in the research landscape.

HashAgent is a lightweight tool that packages an AI agent into a shareable URL. Recipients click the link and run the agent locally in their browser via WebGPU, with no installation required and data staying on-device. It scored 45 points on Hacker News, useful for rapid prototyping and collaborative debugging without cloud deployment.

Ben’s Bites released a new session focusing on the definition of a ‘personal agent’. As AI assistants evolve from simple Q&A tools to proactive agents capable of executing complex tasks, the ‘personal agent’ is seen as the next key development. The session explores the core capabilities a personal agent should have, how it differs from existing AI assistants, and its potential applications in daily life and work.

A viral essay titled “When Genius Fails: The Intellectual Arrogance of the AI Labs” sparked heated discussion on Hacker News with 170 points and 184 comments. It critiques the “genius culture” inside AI labs for fostering insular decision-making and dismissing external criticism, with commenters debating impacts on AI safety, transparency, and public trust.

Claude Code released v2.1.233. This update adds GitLab merge request URL support to the –worktree flag and the claude agents view (MRs display as !N). It also introduces an opt-in forwarduseridentity apps gateway setting on Anthropic upstreams that sends the signed-in user’s identity as headers, allowing proxies behind the gateway to identify users.

OpenAI Codex released version 0.148.0-alpha.18. This alpha version follows a rapid sequence of test releases (alpha.14 through alpha.18). Specific changelog details have not yet been published, but the frequent iteration suggests rapid bug fixes and feature refinements. For developers, this update may include improvements to coding task stability and tool-calling accuracy.