Eliza Times

Industry News

• Allbirds pivots from shoes to AI compute: The footwear company announced it's selling off its shoe business to become an AI infrastructure provider, with stock surging 200%+ on the news. A stark example of "AI pivot" mania in public markets. link

• Anthropic hits reliability issues as usage explodes: Claude Code experiencing daily outages as the company struggles to scale infrastructure after crossing $30B ARR. Status pages remain green while users face repeated 500 errors. link

• Uber CTO: AI coding already maxed 2026 budget by April: Uber burned through its entire year's AI coding tool budget in Q1, forcing mid-year renegotiations with Anthropic. A wake-up call on unpredictable AI costs for enterprises. link

Tips & Techniques

• When agents get stuck, ask: "Are you missing any context?": Simple prompt that forces the model to surface gaps in its understanding rather than hallucinating forward. Dramatically improves output quality in complex tasks. link

• Why on-policy distillation fails (and how to fix it): New research identifies two critical conditions for successful OPD: comparable teacher-student capacity and high-quality sampling. Without both, distillation actively hurts performance. link

• Multi-token prediction provably improves reasoning: UC Berkeley mathematically proves that predicting multiple tokens ahead enhances reasoning capabilities over standard next-token prediction. Explains why Mythos may use looped transformers. link

New Tools & Releases

• Cloudflare Think: Long-running agents with zero ops: New agent framework with durable execution, persistent sessions, sandboxed code execution, and sub-agents—all managed automatically. Agents survive restarts and cost nothing when idle. link

• Multica Autopilot: Open-source Claude Code alternative: Community-built local version of Routines that runs entirely on your machine. Full control, no cloud dependency, compatible with any model. link

• Gemma 4 via OpenClaw: Google's most powerful US-released open model now has native integration with the OpenClaw agent framework. Runs fast enough on M1 hardware for practical agent workflows. link

• Nucleus-Image: First sparse MoE diffusion model: 17B parameters but only 2B active during inference—10x more efficient than dense models. Apache 2.0 licensed and production-ready. link

Research & Papers

• AI alignment via outcome-gradable tasks: Anthropic formalized superhuman AI supervision into testable tasks where Claude substantially beat all human-optimized baselines in 5 days. New framework for scalable oversight. link

• Experience replay cuts LLM RL training compute by 40%: Reusing high-quality trajectories from earlier training reduces compute without hurting final accuracy—sometimes improving it. Challenges conventional on-policy assumptions. link

• Why looped transformers may power Claude Mythos: New paper shows looped LLMs achieve better scaling laws than standard transformers with identical parameter counts. Explains rumored architecture behind Anthropic's unreleased model. link

• AIRA₂: Automated AI research that addresses bottlenecks: Next-gen AI research agents from MILA that systematically identify and resolve key scaling obstacles in ML research. Focuses on the science of automated research itself. link

--- *Curated from 500+ tweets across AI research, engineering, and product lists*

---

Emerging Trends

✨ Anthropic Growth and Competitive Positioning (92 mentions) - NEW Discussion of Anthropic's unprecedented revenue scaling, board additions (Novartis CEO), and potential IPO. Concerns emerging about API reliability and potential competition with customers like Lovable. Reports of leaked competitive features targeting coding agent market.

🔥 Cloudflare Project Think and Agents Week (85 mentions) - RISING Cloudflare is launching Project Think as part of Agents Week, featuring durable execution, sub-agents, persistent sessions, sandboxed code execution with credential injection at the network layer, and programmable egress proxy for secure agent access to private services.

🔥 Vercel Open Agents Platform (72 mentions) - RISING Vercel open-sourced their reference platform for cloud coding agents with agentic infrastructure including Fluid for agent brain, Workflow for durability, Sandbox for code execution, and AI Gateway. Multiple companies are building internal AI software factories.

🔥 Claude Code Outages and Reliability Issues (145 mentions) - RISING Users experiencing frequent Claude Code outages and service disruptions, with multiple reports of downtime affecting developers mid-workflow. This is creating frustration despite the tool's popularity for vibe coding and agent development.

📊 Vibe Coding and AI Development Tools (165 mentions) - CONTINUING Continued discussion of vibe coding with Lovable, Replit, Claude Code, and Codex for rapid prototyping. Debate over maintainability, with enterprises seeking control through platforms like Superblocks 2.0 that offer IT oversight of AI-generated apps.