Eliza Times

Industry News

• Pentagon demands full control of Anthropic AI: Pentagon issued "best and final offer" to Anthropic for unrestricted access to Claude, with decision deadline by Friday. Raises major questions about AI company independence and defense partnerships. link

• Anthropic stock impact hits India: Indian IT stocks crashed 6% after Anthropic released enterprise AI tools that compete directly with traditional service providers. Legal plugin release triggered selloffs in Thomson Reuters (-21%), LegalZoom (-19%), CS Disco (-14%). link

• Vercel support agent handles 90% of queries: Vercel's AI support agent now resolves ~90% of customer inquiries autonomously, but team unable to upgrade from Claude Opus 4.5 to 4.6 without performance degradation. Shows "best model" is a moving target. link

• Nano Banana 2 launches in Gemini: Google's Gemini 3.1 Flash Image (Nano Banana 2) now live in Gemini apps, offering 50% cheaper image generation (~$0.16 per 4k image) with enhanced visual quality and text rendering vs original. link

Tips & Techniques

• Ask agents if they're missing context: When an agent isn't performing as expected, explicitly ask "Are you missing any context?" to surface overlooked requirements or constraints. Simple meta-prompt that improves output quality. link

• Think about tests first for agent development: Highly recommend thinking about the test first when building agent features. Huge unlock for quality and reliability. link

• Large learning rates improve generalization: Large learning rate at training start while annealing toward smaller one by end ALWAYS gives better empirical generalization. The "shoot the minima" fear is overblown. link

• Don't invest in LLM guardrails: Simple rule: Don't invest in guardrails for LLMs. Instead, invest in making everything flow easily into LLM context. Trust the magic of compounding. link

New Tools & Releases

• Zvec: Embedded vector database: Alibaba's Zvec is "SQLite for vector databases" - lightweight, zero-config, in-process vector DB with hybrid search and grouped search. Handles billion-scale vectors with millisecond latency. link

• Tamarind Bio raises $13.6M for molecular AI: Platform serving tens of thousands of scientists including 8 of top 20 pharma. Library of hundreds of AI models, supports open models strengthened with proprietary data, prioritizes scientist access. link

• Faster Qwen3TTS at 4x realtime: Realistic voice generation at 4x real time with same quality as Qwen's original model. Major speed improvement for voice applications. link

• Claude Code gets Figma integration: Starting today, generate Figma design files from Claude Code using Figma MCP server. Supports bidirectional workflow - UI to Figma canvas and back to code seamlessly. link

Research & Papers

• VLANeXt: Recipes for strong VLA models: Systematic study of what actually matters for vision-language-action performance: policy heads, meta-queries, action chunking, flow matching. Provides actionable guidance for robotics developers. link

• Truthfulness probes vary widely across domains: Multiple orthogonal truthfulness directions shared across LLM honesty evals. Some generalize, others are domain-dependent. Proposes Truthfulness Multivector framework for more robust analysis. link

• Mid-training bridges pre-training and supervised learning: New research shows data that "bridges" pre-training and supervised learning is most effective for mid-training. Reveals optimal ratios and timing that aren't well documented in public research. link

• POCO predicts neural activity before it happens: Can predict what one neuron will do next by knowing what entire brain is doing now. Tool paving way for adaptive neurotechnologies. link

--- *Curated from 400+ tweets across AI research, tools, and industry lists*

---

Emerging Trends

✨ Nano Banana 2 Release (12 mentions) - NEW Launch of Nano Banana 2 (NB2) image generation model in Gemini WebUI/App, with comparisons to Grok Imagine and discussions of its strengths in micro-level detail but weaknesses in macro-level world knowledge and temporal accuracy.

🔥 Cursor vs Claude Code vs Codex Competition (145 mentions) - RISING Developers cycling between Cursor, Claude Code, and Codex tools, with discussions about workflow preferences, brain-off mode concerns, and the emotional attachment to code quality when using different AI coding assistants.

🔥 Claude Code Anniversary & Features (98 mentions) - RISING Claude Code celebrating its 1-year anniversary with major feature releases including markdown table rendering, remote control capabilities, demo recording features, and Claude Cowork enterprise plugins marketplace for Google Workspace, Salesforce, Excel integration.

🔥 SaaS Disruption by AI Concerns (102 mentions) - RISING Discussions about AI tools like Anthropic's Claude Cowork threatening traditional SaaS businesses, with Indian IT stocks crashing 6% and debates about whether SaaS companies will need to dramatically lower margins or face extinction from AI automation.

📊 AI Job Displacement & Economic Impact (87 mentions) - CONTINUING Viral discussions around Citrini's scenario about AI-driven job displacement, fiscal policy responses, and debates about whether AI will lead to abundance/aristocracy for everyone or permanent underclass, with references to the Atlantic article generating strong reactions.