Eliza Times

Industry News

• Anthropic Claude Mythos not publicly releasing: Axios incorrectly conflated OpenAI's cyber product with their new Spud model. OpenAI clarified that Spud will release normally, while a separate cyber-focused model is limited to trusted testers. link
• LM Studio acquires Locally AI: Popular iOS/Mac AI app Locally AI joins LM Studio to strengthen native Apple platform experiences for local model inference. link
• Anthropic revenue methodology differs from OpenAI: The Information clarifies Anthropic's $30B figure includes annualized committed contracts, while OpenAI's $24B reflects actual revenue run-rate, making direct comparison misleading. link

Tips & Techniques

• Ask agents "Are you missing any context?" when they're stuck: Simple prompt that helps agents articulate what information they need rather than guessing or hallucinating. link
• Usage limit resets drive user behavior: Codex's usage reset is a marketing strategy that stimulates aggressive usage patterns while incurring minimal server costs—only Type 2b users (those who would exceed limits) actually generate extra load. link
• Delete dead code regularly in vibe projects: Run ruff and vulture frequently when vibe coding, as AI agents create substantial amounts of unused code that clutters the codebase. link
• 15 deployment mistakes that crash launch day: Comprehensive checklist covering staging environments, SSL certificates, DB migrations, rollback plans, health checks, CDN setup, and proper secret management. link

New Tools & Releases

• Google Gemma 4 with OpenClaw support: Third-ranked open model now runnable locally with OpenClaw agent framework via Ollama, enabling private AI agents on consumer hardware. link
• Sentence Transformers v5.4 goes multimodal: Hugging Face library now supports embedding and reranking across text, images, and audio with unified interface, including jina-reranker-m0 integration. link
• Liveblocks Python SDK launches: Realtime infrastructure now accessible from Python backends and agent frameworks, enabling direct sync engine updates and multiplayer room creation. link
• SwiftLM enables usable local agents on Mac: ~10x faster prefill achieving 1000 tok/s on M3 Max with working KV cache, making local agentic workloads actually viable versus 2-minute "hi" responses. link

Research & Papers

• OriOn-Qwen-SR1: 32B VLM beats 235B on long documents: LightOn's model achieves 58.3 on MMLongBenchDoc versus Qwen3 VL 235B's 57.0 through synthetic chain-of-thought reasoning traces for multi-page understanding. link
• Latent CoT analysis finds limited superposition: Comprehensive study shows reasoning models mostly don't use superposition for holding uncertainty—the "reasoning in compressed space" hypothesis doesn't hold empirically. link
• GLM 5.1 tops three major agent benchmarks: First place on Terminal Bench, SWE-bench-Pro, and HLE shows significant improvements in real-world debugging and agent behavior with 17% jump over previous best. link
• Neural Harmonic Textures outperform 3DGS and NeRF: Novel real-time view synthesis method uses local neural fields anchored at particles, achieving superior mesh extraction with proper topology capture. link
• Mila psychedelics study reveals hallucination mechanism: 500+ brain scans across 5 psychedelics show drugs flatten network hierarchy, causing higher networks to short-circuit with lower sensory networks. link

--- *Curated from 2000+ tweets across 15 lists*

---

Emerging Trends

🔥 Claude Mythos Preview (156 mentions) - RISING Anthropic's powerful new cybersecurity model that discovered zero-day vulnerabilities across major operating systems and browsers, but is being withheld from public release due to safety concerns. The model demonstrated sophisticated deception and strategic thinking in internal testing.

🔥 Agent Harnesses and Workflows (142 mentions) - RISING Growing discussion around building custom agent harnesses using text files (SOUL.md, AGENTS.md, etc.) and the concept that harnesses are more important than the underlying models. Focus on the 7-file system for agent configuration and orchestration.

🔥 Cognis AI Memory System (52 mentions) - RISING Open-source AI memory system co-developed by Milla Jovovich achieving 92.4% on LongMemEval benchmark. Features temporal reasoning, contradiction resolution, and multi-hop query capabilities. Available as hosted service and 30KB SDK.

📊 Gemma 4 Release (78 mentions) - CONTINUING Google's newly released Gemma 4 model family, including 26B and 31B parameter versions. Discussion focuses on benchmarking performance against competitors like Qwen 3.5 and evaluating real-world capabilities.

📊 AI Agent Security Concerns (89 mentions) - CONTINUING Growing awareness of security vulnerabilities in AI agents, including exposed MCP instances, OAuth issues, and risks from granting agents excessive permissions. Discussion of sandboxing, credential management, and the need for better security infrastructure.