Eliza Times

Industry News

• OpenAI pivots to enterprise: WSJ reports OpenAI cutting side projects to focus on coding + enterprise, with potential $10B JV talks with TPG, Brookfield, and Bain. Meanwhile, NVIDIA makes inference 10x cheaper with Vera Rubin chips. link

• Crusoe AI raises at $10.4B valuation: Series E led by Valor Equity Partners and Mubadala to build AI infrastructure, with impressive investor lineup including NVIDIA, Fidelity, and Founders Fund. link

• NVIDIA GTC highlights: Jensen Huang declares "the inference inflection has arrived" and states "every company needs an OpenClaw strategy." Announced partnerships with Disney for character robotics and expanded AlphaFold database with protein complexes. link

Tips & Techniques

• Debug Claude Code faster: Add this to claude.md to skip the "wait, actually..." loop: when stuck on a bug, gather real data first via logging instead of re-theorizing without evidence. Form one hypothesis, test it, reassess. link

• Fix missing agent context: If an agent isn't doing what you want, simply ask it: "Are you missing any context?" This actionable prompt helps agents self-correct. link

• Cursor stop hook for cleaner code: Configure Cursor to automatically check its work after each edit. This follow-up verification step catches dead code and improves implementations without manual review. link

New Tools & Releases

• Mistral Small 4 released: 128 experts, 119B total parameters, 256k context window with configurable routing. One model to do it all, with strong multilingual capabilities. link

• Grok Text-to-Speech API now available: Natural voices with expressive controls for building voice-enabled applications. link

• Context Hub by Andrew Ng: Stack Overflow for AI coding agents to share learnings. CLI tool allows agents to access real-time API documentation and best practices. link

• FastMCP 3.1.1 ships: Migration completed with working Hugging Face OAuth, token passthrough, and elicitations support. link

Research & Papers

• Attention Residuals from Kimi: Novel architecture applies attention across layers (not just tokens), improving feature usage and training stability. Works better on structured data vs random/memorized data. link

• Seoul World Model: First world model that renders actual cities, not imagined places. Enables realistic urban simulation at scale. link

• xLSTM Distillation: Near-lossless distillation of quadratic Transformer LLMs into linear xLSTM architectures. Significant efficiency gains while maintaining performance. link

• Meta Omnilingual MT: First MT system benchmarked on 1,600+ languages. Prioritizes proteins important for human health including WHO bacterial priority pathogens. link

Contrarian Takes

• LLMs may not boost net productivity: Zeeg argues LLMs remove barriers to starting but create maintenance debt, especially with incremental complexity. After months of shipping with LLMs, sometimes writing code manually is faster. link

• Code != bottleneck: "If you can now build anything and decide to just copy slop and wonder why you don't make money, maybe code was never the bottleneck." Building is easy; knowing what to build is hard. link

--- *Curated from 500+ tweets across tech/AI communities*

---

Emerging Trends

✨ Crusoe Energy Series E (476 mentions) - NEW Crusoe Energy (CrusoeAI) announced a Series E funding round at $10.4B valuation, led by Valor Equity Partners and Mubadala Capital, to build AI infrastructure. Major investors include Nvidia, Fidelity, Founders Fund, and others.

✨ AlphaFold Database Expansion (313 mentions) - NEW Google DeepMind expanded the AlphaFold Database with millions of AI-predicted protein complex structures, in collaboration with EMBL-EBI, Nvidia, and Seoul National University, prioritizing proteins important for human health and disease.

✨ Orchid Proactive Agent (233 mentions) - NEW Orchid launched in public beta as a proactive delegation layer/agent that requires no configuration, powered by Convex and Cloudflare. Multiple tweets discuss it as a new approach to AI agents for modern work.

✨ ScienceClaw AI Agent Swarm (27 mentions) - NEW MIT's LAMM team released ScienceClaw × Infinite, an open-source AI agent swarm platform for crowdsourced scientific discovery. The swarm self-coordinates across institutions to solve problems like designing peptide binders, discovering materials, and uncovering cross-domain structures.

✨ VisMatch Image Matching (265 mentions) - NEW VisMatch released on PyPI as a wrapper for 50+ image matching models including LightGlue, RoMa-v2, MASt3R, and LoFTR, making it simple to run image matching with a single command. Already has thousands of users.

Industry News

• NVIDIA announces NemoClaw at GTC: Jensen Huang positioned OpenClaw as fundamental infrastructure, comparing it to Linux and HTML. NVIDIA released NemoClaw as an enterprise-ready reference platform with built-in security and network boundaries. link

• Mistral's Nemotron Coalition: Mistral joins NVIDIA's Nemotron Coalition to train the first base models together, marking significant open source AI collaboration. link

• Joe Kent resigns from NCTC over Iran policy: Trump's Director of National Counterterrorism Center resigned citing concerns about Israeli pressure driving U.S. toward war with Iran, calling the policy direction "confused and incoherent." link

• Granola backtracks on data access: After community backlash over locking down local cache access, Granola announced they'll ship a public API next week and better local agent support. The MCP alone wasn't sufficient for data export and local workflows. link

Tips & Techniques

• Ask agents about missing context: If an agent isn't performing as expected, simply ask it: "Are you missing any context?" This simple prompt can reveal what information the agent needs to improve. link

• Security audit prompt for API routes: Before shipping, run this in Cursor: "Find every route that reads or writes user data, payments, or PII. For each, list: 1) is there auth? 2) does it check the user can access this resource? 3) are inputs validated with a schema?" Save results to api_audit.md and fix every "no". link

• Typing variables aids learning with AI: One developer realized that letting Claude write everything meant they stopped learning. Typing variable names forces you to memorize them, which means learning how the code actually works. Don't let AI make you stupid. link

New Tools & Releases

• Mistral Small 4 released: New 128-expert MoE model with 119B total parameters, 256k context window, and configurable reasoning. Features 3x speed improvement over Mistral-3 with Apache 2.0 license. link

• Gamma major launch: Gamma announced three big features: Gamma Imagine (AI-native visual creation tool), AI-Native Templates (modify entire decks with single prompts), and Gamma Connectors (integration with ChatGPT, Claude, and major work apps). link

• HuggingFace datasets JSON support: Datasets v4.7+ now has full support for JSON datasets including tool-calling with varying arguments, auto-detecting mixed fields and types. Also added Storage Buckets support for cheaper/faster raw data storage. link

Research & Papers

• LLM gradient bottleneck discovered: New paper shows the output layer of LLMs destroys 95-99% of training signal during backpropagation, significantly slowing pretraining. The LM head acts as a gradient bottleneck. link

• Test-time compute can hurt reasoning: Research shows beam search in LLMs can degrade performance after a certain point. More test-time compute doesn't always help—the quality of the signal guiding the search matters more than raw compute. link

• X-Cell virtual cell model from Xaira: Foundation model that predicts how gene expression changes under perturbations. Trained on diverse biological data to simulate cellular responses, representing a step toward computational biology at scale. link

• Meta Omnilingual Machine Translation: Encoder-decoder system translating from 1,600+ languages to 250. Compact 3B size (1.5B encoder + 1.5B decoder) achieves SOTA performance with fast inference and multi-target translation. Shows encoder-decoder isn't dead. link

--- *Curated from 500+ tweets across 8 tech-focused lists*

---

Emerging Trends

✨ NemoClaw/OpenClaw Enterprise (45 mentions) - NEW NVIDIA announced NemoClaw at GTC, described as a reference platform for enterprise-grade secure OpenClaw with built-in security features. Jensen Huang framed OpenClaw as one of the most important open source releases, emphasizing every software company needs an OpenClaw strategy.

🔥 ScienceClaw AI Agent Swarm (41 mentions) - RISING MIT's LAMM team released ScienceClaw × Infinite, an open-source AI agent swarm platform for crowdsourced scientific discovery where agents self-coordinate across institutions to solve problems like designing peptide binders, discovering ceramics, and uncovering hidden structures.

📊 Crusoe Energy Series E (476 mentions) - CONTINUING Crusoe Energy closed a Series E round at $10.4B valuation led by Valor Equity Partners and Mubadala Capital to build AI infrastructure. The round included major investors like NVIDIA, Fidelity, Founders Fund, and Tiger Global.

📊 AlphaFold Database Expansion (313 mentions) - CONTINUING Google DeepMind expanded the AlphaFold Database with millions of AI-predicted protein complex structures in collaboration with EMBL-EBI, NVIDIA, and Seoul National University, prioritizing proteins important for human health and disease.

📊 VisMatch Image Matching (265 mentions) - CONTINUING VisMatch launched on PyPI as a wrapper for 50+ image matching models including LightGlue, RoMa-v2, MASt3R, and LoFTR, making it simple to run image matching on any two images with a single command.