Eliza Times

Industry News

• OpenAI raises $122B at $852B valuation: OpenAI closed a massive funding round, reaching $2B monthly revenue. The scale suggests investors see near-term returns despite current subsidized pricing. link

• Anthropic's Claude Code leaked, then DMCA'd: A misconfigured .map file exposed Claude Code's full TypeScript source via npm. Anthropic is actively taking down GitHub backups—open sourcing clearly not planned. link

• Okta pivots to low-growth profitability model: After three years of layoffs, Okta is managing like Cisco—prioritizing buybacks over growth. The AI thesis remains generic with no clear competitive moat for the next five years. link

Tips & Techniques

• Structured outputs are underrated for agents: Using structured outputs with hooks/middleware after agent actions creates more reliable workflows than fuzzy execution. Especially useful with subagents running in isolated contexts. link

• Use a "council of LLMs" for better outputs: Running multiple models with different strengths and having them debate/vote on outputs is strangely effective. Different models catch different issues. link

• InferScale: inference-time optimization for LLMs: Documentation shows how to systematically improve open-source LLM outputs using multiple responses, ranking, and selection—without increasing infrastructure costs. link

New Tools & Releases

• PrismML launches 1-bit Bonsai models: Apache-licensed 1.7B, 4B, and 8B models with true 1-bit parameters. Running at 584 tok/s on consumer hardware with 14x compression ratio that scales. link

• TRL v1.0 ships stable post-training APIs: Hugging Face democratizes alignment with stable APIs for SFT, reward modeling, DPO, and GRPO across the entire post-training stack. link

• Wan2.7-Image: unified generation and editing: Alibaba's new model generates, edits, and understands images in one unified architecture. Single model for multiple vision tasks. link

• CaP-X: coding agents for robotics benchmark: Open-source framework where agents write code for robot perception/control, execute in sim/real, and iteratively improve. Agents score 32% vs 88% for human experts—gap is real. link

• Lunel: dev environment on your phone: Scan a QR code and get your entire dev environment in your pocket. No login, no signup required. link

Research & Papers

• AI predicts your next research paper topic: Fine-tuned LLaMA-2-13B on 221k materials science abstracts builds concept graph with 137k nodes. Best model (GNN + embeddings) reaches 94.3% AUC, with 26% of suggestions rated novel by scientists. link

• GrandCode achieves 1st place on Codeforces: Model reaches gold medal level on competitive programming contests, surpassing superhuman performance on this benchmark. link

• MosaicMem combines long-horizon navigation with memory: New demo shows multi-scene concatenation with promotable events for robot navigation across extended horizons. link

• Real-3DQA exposes 3D reasoning shortcuts: Current benchmarks show models perform nearly as well without 3D input. New benchmark increases recall at distance 3 from 5.9% to 35%—targeting genuinely novel combinations. link

--- *Curated from 582 tweets across 15 lists*

---

Emerging Trends

✨ Hermes Agent Launch (28 mentions) - NEW NousResearch's Hermes Agent is gaining adoption as an alternative coding agent platform, with users migrating from OpenClaw and praising its multi-profile support and flexibility.

🔥 Claude Code Source Leak (145 mentions) - RISING Claude Code's entire source code was accidentally leaked via a map file in their npm registry, leading to widespread discussion, analysis, and even forks. The community is examining everything from safety systems to bugs in the codebase.

🔥 Agent-First GTM Strategy (35 mentions) - RISING Companies are recognizing that AI coding agents (not humans) are now choosing which tools and APIs to use, requiring a shift to agent-first go-to-market strategies focused on API accessibility, clean documentation, and structured discoverability.

📊 Claude Code Cache Bugs (42 mentions) - CONTINUING Users discovered cache-invalidation bugs in Claude Code through reverse engineering, causing tokens not to be cached and dramatically increasing costs. Issues include string replacement bugs in bun and --resume breaking cache.

📊 Pazi AI Team Launch (22 mentions) - CONTINUING Pazi launched as an AI team that works 24/7 in Slack and other platforms, executing across entire tech stacks including Sentry, Linear, and GitHub. Now in beta with OpenClaw integration.