Eliza Times

Industry News

• xAI acquires Cursor for $60B: SpaceX/xAI offered $10B upfront collaboration fee plus $60B acquisition rights, shelving Cursor's $2B fundraise. Microsoft had also evaluated but walked away. link
• Anthropic locks 3.5GW TPU capacity through 2027: Via Broadcom deal with Google, while Amazon commits $5B more to Anthropic ($13B total) with $100B+ AWS spend over decade. Infrastructure is now the moat. link
• UAE launching AI-driven government: 50% of government sectors will be AI-driven within two years under presidential directive, replacing traditional bureaucracy with autonomous systems. link
• GPT-5.5 and GPT-5.5 Pro spotted: Both models appeared on OpenRouter with date 2026-04-23, and OpenAI removed "latest frontier" label from GPT-5.4 in Codex model list. link
• Bitwarden CLI compromised in supply chain attack: Version 2026.4.0 included malicious code after attackers abused GitHub Actions token, part of ongoing Checkmarx campaign. link

Tips & Techniques

• Ask agents about missing context when stuck: When an agent isn't behaving as expected, directly ask "Are you missing any context?" - simple but effective prompt technique. link
• State-first programming for AI agents: Ground agents in modeling state rather than just chaining function calls - restores the lost abstraction boundary for building reliable complex software. link
• Configure Claude Code auto mode with $defaults: New v2.1.118 includes "$defaults" in autoMode.allow/soft_deny/environment settings - without it your custom rules replace built-in blocklist instead of adding to it. link

New Tools & Releases

• GStack from YC's Garry Tan: Open-source toolkit that turns Claude Code into an AI engineering team with multi-agent coordination capabilities. link
• NanoClaw v2 with Vercel: Complete rebuild enabling agent-to-agent communication, human-in-the-loop approvals, and support for 15 messaging platforms. link
• Tencent Hy3-preview released: 295B total params, 21B active per token, 256K context window with 40% efficiency gain - built for agents with code/search/reasoning focus. link
• Google agents-cli open-sourced: CLI and skills package to turn existing AI coding assistants into "agent engineers" specialized for building/evaluating/deploying agents on Google Cloud. link
• OpenAI releases 1.5B privacy filter model: Open-weights model for PII detection, ported to MLX runs 24-33x faster on Apple Silicon. Apache 2.0 licensed. link

Research & Papers

• Stateless Decision Memory for Enterprise AI Agents: Replaces stateful agent memory with immutable decision logs using event-sourcing - solves horizontal scaling bottleneck for production deployments. link
• Chinese LLMs running industrial-scale distillation: US evidence shows China using distillation campaigns to steal frontier model capabilities - Kimi K2.5 found 10 Chrome zero-days cheaper than Claude Opus 4.6. link
• Vision Banana from Google DeepMind: Unified generative model outperforms specialist models across dense prediction tasks - strong evidence image generators are generalist vision learners. link
• GiantsBench: Scientific discovery benchmark: Tests whether LLMs can generate insights building on prior research - evaluates scientific reasoning at research frontier. link

Notable Projects

• Flipbook browser from team building interface-less web: Every pixel streamed live from model with no HTML, layout engine, or code - just exactly what the model generates in real-time. link
• Almanac Health launches with $12M: Clinical AI evidence retrieval grounding LLM answers in medical publications, calculators, and databases - published methodology in NEJM AI. link

--- *Curated from 889 tweets across AI development, research, and industry news*

---

Emerging Trends

✨ NVIDIA Free API Models (35 mentions) - NEW NVIDIA is offering around 80 AI models via hosted APIs for free, including MiniMax M2.7, GLM 5.1, Kimi 2.5, DeepSeek 3.2, and others. These integrate directly with OpenClaude, OpenCode, Zed IDE, Hermes agent, and Cursor IDE.

✨ GPT-5.5 Release (48 mentions) - NEW OpenAI is announcing GPT-5.5 (and possibly GPT-5.5 Pro), with mentions of model updates on OpenRouter and speculation about a major release alongside GPT Image 2. The model appears to be removing "latest frontier" designation from GPT-5.4.

🔥 Codex 4M Users Milestone (72 mentions) - RISING Codex has reached 4 million weekly active users, adding over 1 million users in less than two weeks. Rate limits are being reset to celebrate the milestone, with strong enterprise adoption reported.

📊 ICLR 2026 Conference (94 mentions) - CONTINUING ICLR 2026 is happening in Rio de Janeiro, Brazil this week with numerous researchers presenting papers, hosting workshops, and networking. Multiple attendees are sharing their poster sessions and looking to connect.

📊 GPT Image 2 Release (89 mentions) - CONTINUING OpenAI released GPT Image 2 with significantly improved text rendering, typography, and detail quality. Users are testing it extensively for various use cases including pixel art, posters, mockups, and commercial design work.