Models & Releases 3
GPT-Live
OpenAI's new voice model generation for natural, real-time human-AI interaction, now powering ChatGPT Voice
SWE-1.7
Cognition's coding model, claimed to approach GPT-5.5 and Opus-level intelligence on software tasks
HunyuanOCR-1.5
Lightweight end-to-end OCR-specialized VLM unifying document parsing, text spotting, information extraction, and text-image translation in a single model
Tools & Demos 4
Microsoft Flint
Open-source visualization language for AI agents, released by Microsoft
GeoSQL
Turns Claude/Codex into a geospatial analytics agent
Onboard-CLI
AST-powered, local-first CLI that visualizes system architecture and enforces architectural boundaries via Git hooks
Kastor
Terraform-style declarative specs for AI agents: define agents, tools, and prompts in HCL, then compile to frameworks with plan/apply semantics
AI Agents 5
From Atomic Actions to Standard Operating Procedures
Iterative tool optimization for self-evolving LLM agents; agents build reusable higher-level procedures instead of reinventing low-level logic each workflow, cutting reasoning overhead and failure rates
Reason Less, Verify More
Deterministic gates recover a silent policy-violation failure mode in tool-using LLM agents, where a tool executes a forbidden state transition while the task appears successful
Doomed from the Start
Lightweight probes on an agent's hidden activations predict eventual episode failure as early as the first round, enabling early abort to save inference compute
Progressive Crystallization
Turns agent exploration into deterministic, lower-cost production workflows so previously-solved IT-ops problems no longer require full LLM inference each run
SkillCenter
Largest open skill library for agents by count: 216,938 structured skills across 24 domain bundles, aimed at grounding agent outputs in correct, secure, maintainable operational knowledge
Lab Updates 4
[OpenAI]
Separating signal from noise in coding evaluations — New analysis reveals issues in the SWE-Bench Pro benchmark, raising concerns about reliability of coding-model evals (Jul 8) | link
[OpenAI]
Our approach to government and national security partnerships — Principles for responsible AI use, democratic accountability, and public safety (Jul 8) | link
[DeepMind]
Google DeepMind × A24 research partnership — First-of-its-kind partnership between the AI lab and the film studio (Jul 3) | link
[Anthropic]
Introducing Claude Sonnet 5 — New model in the Claude family (Jun 30) | link
Podcasts 6
[Latent Space]
Why AI Infrastructure Must Evolve for Agent Experience — Akshat Bubna (Modal CTO) (Jul 8) | link
Concludes Latent Space's "Agent Cloud" series. Thesis: the cloud was built for human developers, but agents now dominate the workload and need different infra — faster cold starts, sandboxed code execution, and per-agent isolation. Modal just raised a $355M Series C to build for this shift.
[Latent Space]
The Coolest Diffusion Research Isn't in LLMs — Evan Feinberg & Sergey Edunov, Genesis Molecular AI (Jul 1) | link
The most innovative diffusion research is happening in 3D molecular structure prediction, not LLMs. Genesis's PEARL model handles protein flexibility ("induced fit") that breaks classical docking. Sharp take: the field's "2Å RMSD = good pose" benchmark is "slop" — hydrogen bonds need sub-1Å accuracy. Their agentic drug-discovery loop (SAPPHIRE) now iterates like a chemist 24/7.
[Training Data]
Inside Zipline's Autonomous System: 140M Miles, Zero Incidents — Keller Rinaudo Cliffton & Eric Watson (Jul 7) | link
The largest commercial autonomous system on earth is Zipline's drone-delivery fleet, not robotaxis. The drone is only 15% of the solution — the rest is inventory management, air-traffic integration, and systems engineering.
[No Priors]
Analyzing OpenAI's New Dual-Function Model (Jul 8)
[No Priors]
Claude's Adaptation to Mobile Co-Working (Jul 7)
[No Priors]
Mistral's $3.5B Raise, Samsung Profit up 18x from AI Memory (Jul 7)
HN Threads 3
I think I have LLM burnout
202 pts, 135 comments
Widely-shared personal essay on fatigue from constant LLM-tooling churn — struck a nerve in the community.
We made Grok 4.5, GPT-5.5, and Claude build the same apps
108 pts, 51 comments
Head-to-head build-off comparing the three frontier models on identical app-building tasks.
John Deere owners get right to repair under FTC settlement
473 pts, 97 comments
Industry 3
GitLost — GitHub AI agent leaks private repos
Security researchers at noma.security used prompt injection to trick GitHub's AI coding agent into exfiltrating private repository contents; a concrete demonstration of agent-injection risk
OpenAI government & national security partnerships
OpenAI lays out its principles for working with governments on national-security use cases
Enterprise adoption
Australian Payments Plus (ChatGPT Enterprise + Codex) and MUFG (AI-native transformation) both published customer stories with OpenAI
HF Trending Papers 5
Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence
Adapts video generative models for robot control, addressing the domain mismatch where content-creation models prioritize visual fidelity over physical realism
LLM-as-a-Tutor
Policy-aware prompt adaptation for non-verifiable RL; adapts training prompts (not just rubrics) to the evolving policy
JD Oxygen AI Item Center V1
Industrial-scale LLM/VLM solution for e-commerce item understanding across tens of billions of SKUs at JD.com
Dual Latent Memory in Vision-Language-Action Models
Adds memory to VLA robot-manipulation models to handle long-horizon, temporally dependent tasks beyond the Markovian single-observation assumption
CanvasAgent
Complex image creation and editing via visual tool orchestration (synthesis, localization, segmentation, compositing, OCR)
arxiv: LLMs 4
From Noisy Traces to Root Causes
Structural trajectory analysis and causal extraction to make agent execution traces usable for optimization without discarding causally important evidence
Single-Rollout Asynchronous Optimization for Agentic RL
Asynchronous RL pipeline for post-training LLMs on long-horizon agentic tasks, replacing inefficient synchronous batch-interleaved pipelines
Think Big, Search Small
Studies where model capacity matters in hierarchical multi-agent search: how to distribute model scale across the main agent vs. parallel sub-agents
Does Bielik Know What It Doesn't Know?
Activation dispersion separates entity familiarity from factual reliability across model scale; models' activations betray unfamiliar entities before the first answer token