OpenAI launches GPT-5.5 Instant chat model

#1 📝 OpenAI News

OpenAI and Broadcom unveil LLM-optimized inference chip - OpenAI and Broadcom announced a collaboration to release a Jalapeño inference chip optimized for large language model workloads, aiming to boost inference performance and energy efficiency. The chip is intended to enable faster, more cost-effective deployment of LLMs.

#2 📝 OpenAI News

How agents are transforming work - By June 2026 Codex had become OpenAI’s primary AI tool—accounting for 99.8% of weekly output tokens internally and more than 85% of output tokens for the average employee—while 80.6% of sampled individual users made at least one Codex request estimated to exceed 30 minutes, 70.2% exceeded one hour, 42.4% exceeded four hours, and 25.6% exceeded eight hours. Non-developer users grew fastest since August 2025 (137x for individual users, 189x for organizational users, and 12x within OpenAI), and the 99th-percentile users regularly generated more than 60 hours of Codex agent turns per day.

#3 𝕏

OpenAI launched GPT-5.5 Instant—a more engaging chat model with improved intent understanding, reliable complex-constraint handling, and smarter shopping and local recommendations—rolling out to paid users today and free users tomorrow.

#4 𝕏

Philipp Schmid showcases Google’s new Gemini 3.5 Flash “computer-use” model you can test live on Browserbase. He also links to the official API docs and blog post for easy integration.

#5 𝕏

Qwen open-sourced Qwen-AgentWorld-35B-A3B (MoE, 35B params/3B active, 256K context) and AgentWorldBench, unveiling a dual-path roadmap—scalable simulators and internal world modeling—to expand general agent capabilities.

#6 𝕏

Jason Zhou showcases Crabbox, which spins up isolated cloud boxes for each AI agent to sync uncommitted changes instantly, run the full stack, and capture screenshots/videos as proof—helping his team ship 10× more PRs.

#7 𝕏

NVIDIA AI launched Metropolis Blueprint VSS 3, an open-source video search and summarization framework with 16 natural-language agent skills—search, summarize, alerts, reports and clip review—plus production-ready, #1 SOTA 3D multi-camera tracking.

#8 𝕏

NVIDIA AI introduces NeMo AutoModel on Hugging Face Transformers v5, adding Expert Parallelism, DeepEP and TransformerEngine kernels to Mixture-of-Experts models and boosting training throughput by 3.4–3.7×.

#9 📝 Anthropic Engineering

How we contain Claude across products - Layered defenses—environmental sandboxes/VMs/egress controls, model-layer system prompts/classifiers/training, and limiting external content—are used across claude.ai, Claude Code, and Cowork; telemetry shows users approved roughly 93% of permission prompts, Claude Code auto mode blocks about 83% of overeager behaviors before execution, Claude Opus 4.7 holds prompt-injection success to ~0.1% on single attempts (~5–6% after 100 adaptive attempts), and Claude Mythos Preview was judged too high-risk to ship in April 2026.

#10 𝕏

OpenAI built Jalapeño, its first AI chip co-developed with Broadcom and optimized for LLM workloads in ChatGPT, Codex, the API and future agentic products to boost scalability and broaden AI access.

#11 𝕏

Philipp Schmid launched computer use in Gemini 3.5 Flash, letting agents drive browser, mobile & desktop screens with built-in safeguards like user confirmation, auto-stop & prompt-injection defense.

#12 𝕏

Harrison Chase demos LangSmith Engine’s “sleep time compute” (aka dreaming) to trace agent trajectories, run background analyses, and auto-update memory in Context Hub.

#13 ▶️

Hermes Full Course: Build Your 24/7 AI Chief of Staff in 45 Minutes

Peter Yang

#14 𝕏

Cognition: Devin built an automated QA workflow where you review and approve a test plan before PR review, then receive a screen recording with a visual, step-by-step QA checklist.

#15 𝕏

Google Research demonstrates that reasoning steps unlock latent parametric knowledge in LLMs via a computational buffer effect and factual priming. They outline strategies to harness these mechanisms for more reliable models.

#16 𝕏

bolt.new released four AI-for-real-estate app templates—including three turnkey tools and a parcel viewer—ready to clone at https://bolt.new/solutions/ai-for-real-estate

#17 𝕏

clem 🤗 open-sourced Kog’s 2B-parameter Laneformer model on Hugging Face, demonstrating latency-first inference at over 3,000 tokens/sec.

#18 𝕏

Santiago shared a fully open-sourced RAG assistant for navigating airline policies, complete with code and a @lenadroid walkthrough.

#19 𝕏

Santiago previews Tripo AI’s Project Eden, a world-first model that constructs a persistent, editable world map before rendering and supports multiple users and AI agents. The startup has raised nearly $200 M across two rounds and will demo it at SIGGRAPH 2026.

#20 𝕏

Cursor now lets you delegate tasks directly from Notion using the Cursor SDK. Agents run on the same models and runtime behind Cursor, so you can assign specs or even open PRs for your team to review.

#21 𝕏

xAI launched the official @MongoDB plugin for Grok Build, enabling users to query data, optimize indexes, and manage databases directly within the tool.

#22 𝕏

Google DeepMind explores how AI agents can self-organize into “agentic economies,” negotiating, transacting, and delegating tasks at scale. It argues that diversifying agent decision-making is crucial to avoid security traps, cognitive monocultures, and AI groupthink.

#23 📝 Claude Code Blog

Building effective human-agent teams - Discusses strategies and best practices for designing effective collaborations between humans and Claude agents, focusing on organizational, product, and workflow considerations. Aimed at teams building with Claude to improve outcomes and human-agent coordination.

#24 📝 Mario Zechner

Slow down to speed up: AI and software engineering - An AI-generated and AI-reviewed Meta feature allowed attackers to take over Instagram accounts by faking their location and asking Meta AI to send verification codes, prompting a SEV investigation and the resignation of Meta’s CISO. The breach unfolded amid a “token-maxing” culture where engineers were implicitly measured on AI token usage, massive layoffs announced around May 20 (about 8,000 people, ~10%), and a reorganization that reassigned roughly 40% of Instagram’s trust-and-safety team to manual AI data labeling (around 5,000 developers company-wide), leaving many teams less than half their former size and some services without on-call coverage.

#25 𝕏

Peter Yang questions the role of design when AI agents interact solely via APIs or CLIs, suggesting that traditional UX becomes irrelevant in agent-driven product access.