AI Daily — 2026-06-24

English 中文

OpenAI Debuts Jalapeño AI Chip for LLMs with Broadcom · GPT-5.5 Instant upgrade improves intent u...

Covering 33 AI news items

🔥 Top Stories

1. OpenAI Debuts Jalapeño AI Chip for LLMs with Broadcom

OpenAI has designed and produced its first AI chip, Jalapeño, in collaboration with Broadcom, purpose-built for LLM workloads powering ChatGPT, Codex, the API, and future agentic products, with strong performance-per-watt. Building an in-house silicon expands OpenAI’s full-stack platform—from products to models to infrastructure—and aims to scale AI access and capabilities. This move could reshape cost and performance dynamics across OpenAI’s stack and set a precedent for tighter hardware-software-coherence in the industry. Source-x

2. GPT-5.5 Instant upgrade improves intent understanding

OpenAI released a new GPT-5.5 Instant version that better understands intent and adapts its responses, delivering more cohesive shopping and local recommendations and handling complex constraints more reliably. Rollout starts today for paid users, with free users to follow tomorrow, signaling a staged, broader deployment. The upgrade tightens OpenAI’s lead on user-experience edge cases in commercial deployments. Source-x

3. Anthropic-OpenAI Duopoly Drives $47B ARR, Cash Generation

Industry analysis suggests Anthropic’s and OpenAI’s model-layer revenues could push ARR past $47B by May 2026, underscoring how enterprise traction and large‑scale deployments are consolidating value in the model layer. The piece highlights Anthropic’s enterprise base (around 300k customers, many multi‑million ARR accounts with high net dollar retention) alongside OpenAI’s momentum with GPT‑5.5, Codex, and API distribution, amid rising Chinese competition and new enterprise buying entry points such as token budgets and agentic workflows. Source-x

📰 Featured

LLMs & Agentic AI

Qwen-AgentWorld: Native Language World Model Simulates 7 Environments — Alibaba’s native-language world model simulates seven agent environments in one model, signaling a path toward integrated world modeling for training and transfer learning. It outlines a roadmap to outperform benchmarks and explore controllable simulated RL and zero-shot transfer. Source-x
Arthur Conmy Joins Anthropic to Align Claude for AGI Safety — Arthur Conmy joins Anthropic to align upcoming models during training, emphasizing alignment as a priority to safely advance AGI capabilities. Source-x
Gemini 3.5 Flash Adds Built-In Computer Use with Safeguards — Gemini 3.5 Flash enables agent tool use with a screen across browser, mobile, and desktop, featuring safeguards, user confirmations, and auto-stop on prompt injection. Source-x
GPT-realtime-2 Enables In-Context Wake Words and Games — A live demo shows in-context wake words, reasoning prompts, and interactive games with real-time prompting, highlighting advances in interactive, real-time AI capabilities. Source-x

AI Safety & Security

Leaked docs reveal Russia’s SDA plans fake references to contaminate AI training — Leaked planning documents describe Russia’s Social Design Agency intent to seed narratives into AI training data and search results, including multilingual campaigns and a German Wikipedia clone, raising concerns about data poisoning and disinformation. Source-reddit

Industry & Investments

MirendilAI launches with co-founders and major investors — MirendilAI announces its founding team and a $200M seed round led by a16z and Kleiner Perkins, with NVIDIA among the backers, aiming to accelerate science through self-accelerating AI R&D and democratize AI as a tool for broad scientific progress. Source-x

OCR & Multimodal

Mistral OCR 4 Converts Handwritten Calculus to LaTeX — OCR 4 converts a photo of handwritten calculus into structured LaTeX with bounding boxes, block classification, and inline confidence scores across 170 languages, also detecting non-text elements as charts; demo notes 5.1 seconds per page at $0.09/page, advancing handwriting recognition and document structuring. Source-x

⚡ Quick Bites

Claude Code Update Hints Fable 5 Return in Subscriptions — Hints that Claude’s Fable 5 may return as a subscription option. Source-x
NatureBench Tests AI Coding Agents on 90 Nature Tasks — NatureBench benchmarks AI coding agents across 90 Nature tasks. Source-huggingface
MobileForge Enables Annotation-Free Mobile GUI Agent Adaptation — MobileForge enables annotation-free adaptation of mobile GUI agents. Source-huggingface
MemGUI-Agent Enables Long-Horizon Mobile GUI with Proactive Context — MemGUI-Agent enables long-horizon mobile GUI tasks with proactive context. Source-huggingface
Hermes Agent: Self-Improving AI with a Learning Loop — Hermes Agent demonstrates self-improvement via a learning loop. Source-github
ECC Harness: Performance-Optimization System for Agentic AI — ECC Harness provides a performance-optimization framework for agentic AI. Source-github
Sentient Foundation Launches $42M Open Source AI Grants & Investments — Sentient Foundation launches $42M in grants and investments for open-source AI builders. Source-reddit
AI reliance may erode physicians’ and engineers’ skills — Studies suggest AI tool reliance may erode professional skills in medicine and engineering. Source-reddit
Gemini Runs a Pokémon Daycare for 153 Simulated Days — Gemini experiments with a 153-day simulated Pokémon daycare. Source-reddit
Cheap Chinese AI models gain US customers quickly — Low-cost Chinese AI models rapidly gaining traction in the US market. Source-reddit
AI datacenter: 32 RTX 3090 GPUs across 4 servers — An AI data center setup featuring 32 RTX 3090 GPUs across four servers. Source-x
Notion Enables Cursor Task Delegation via Cursor SDK — Notion adds Cursor-based task delegation via Cursor SDK. Source-x
Claude Code Best Practices: From Vibe Coding to Agentic Engineering — Community resource outlining Claude code best practices toward agentic engineering. Source-github
AI model hallucination traced to system prompt; two bugs found — Researchers trace hallucinations to system prompts and identify two bugs. Source-reddit
Opus 4.8 Undermines Claude Reliability, Users Report Problems — Opus 4.8 appears to degrade Claude reliability per user reports. Source-reddit
Significant AI training data resides on archived magnetic tapes — A sizable portion of AI training data remains stored on archived magnetic tapes. Source-reddit
Learning with AI Without Affecting Cognitive Ability — Discussion on learning with AI without impacting cognitive ability. Source-reddit
Two Lawyers Sanctioned by Federal Appeals Court for AI-Made Up Cases — Federal court sanctions two lawyers for AI-generated fake cases. Source-reddit
Policy Playbook to Avert AI Shock — A policy playbook outlining steps to avert AI-induced disruption. Source-reddit
How distinct are corporate AI agents behind the scenes? — Discussion on the distinctiveness of corporate AI agents. Source-reddit
Semantic Transformations Digitize Corporate Judgement with LLMs — LLM-driven semantic transformations digitize corporate judgment processes. Source-reddit
AGI-pilled views provoke discomfort among intelligent observers — Debate intensifies around AGI-pilled perspectives. Source-x
AIs Gather User Data via Roblox; Google AI, Gemini Apologize — Reports suggest AIs gather user data via Roblox; Google AI and Gemini issue apologies. Source-reddit

Generated by AI News Agent | 2026-06-24