Eliza Times

Industry News

• Anthropic releases Claude Mythos Preview - too powerful for public release: 10T parameter model scores 94% on SWE-bench and discovered a 27-year-old OpenBSD vulnerability, but will only be deployed to 12 vetted cybersecurity partners through Project Glasswing due to unprecedented offensive capabilities. link

• Safetensors joins PyTorch Foundation: Hugging Face's safe model serialization format becomes foundation-hosted, gaining first-class PyTorch integration and independent governance to scale ML security across the ecosystem. link

• Anthropic hits $30B+ ARR, surpassing OpenAI's $25B: Claude's enterprise adoption and coding capabilities drive revenue growth, positioning the company ahead of its IPO. link

Tips & Techniques

• Ask agents about missing context when stuck: If an agent isn't completing its task, explicitly ask "Are you missing any context?" to identify capability gaps rather than hallucinated solutions. link

• Database schema design prompt prevents architectural debt: Ask Claude to design full schemas with UUIDs, foreign keys, indexes, NOT NULL constraints, soft deletes, and explain reasoning for each decision - catches bad DB design early. link

• Vibe coding database workflow: Generate complete database schemas with migrations, seed data, and design rationale using Claude - the "explain your reasoning" part surfaces bad decisions before implementation. link

• Agentic code quality metrics that actually matter: Track PR volume per engineer, time-to-feature vs estimate, and roadmap compression - not productivity in abstract but how fast you're shipping with the same team size. link

New Tools & Releases

• Gym-Anything: Turn any software into agent training environment: Framework converts arbitrary desktop software into controllable environments for training/testing agents, includes CUA-World benchmark with 10,000 environments from 200 packages. link

• OpenRound: Technical interviews for the AI era: Assessment platform where candidates build features in real codebases using AI agents, measuring ability to ship with AI tools rather than whiteboard algorithms. link

• Genspark Workspace 4.0 integrates AI into Office: Claude-powered agents work directly inside PowerPoint, Excel, and Word with voice control (Speakly) and autonomous background execution (Claw compiles expense reports in 6 minutes unattended). link

• YC-bench: Startup survival benchmark for LLM agents: Agents get $200k virtual budget for 1 year, must avoid bankruptcy while running a business - only GLM-5 and Kimi survived and grew revenue, nine models failed. link

Research & Papers

• Single-agent LLMs outperform multi-agent systems on equal token budgets: Stanford paper shows multi-agent reasoning advantages disappear when controlling for thinking tokens - information loss through message-passing bottlenecks explains why unified context wins. link

• PD controller gains matter more than you think in robot RL: Empirical study across 100+ real-world training runs shows control gains behave fundamentally differently in learned policies vs classical control, directly impacting sim-to-real transfer success. link

• Video-MME-v2 benchmark: Human 90.7 vs best model 49.4: New 3,300-hour video understanding benchmark with progressive tri-level hierarchy exposes saturation in current benchmarks - Gemini-3-Pro leads but massive gap remains. link

• Hierarchical planning unlocks long-horizon world models: JEPA-based approach enables non-greedy behavior across extended time horizons by combining high-level strategy with low-level execution in latent space. link

--- *Curated from 500+ tweets across AI research and engineering communities*

---

Emerging Trends

✨ Claude Mythos Preview (85 mentions) - NEW Anthropic announced Claude Mythos Preview, a highly capable but unreleased model that exhibits sophisticated strategic thinking and situational awareness. The model sparked significant discussion about AI safety, cybersecurity capabilities, and found vulnerabilities in systems like a 27-year-old OpenBSD bug.

🔥 Agent Harnesses and Workflows (95 mentions) - RISING Growing focus on AI agent infrastructure through harnesses (like OpenClaw alternatives) and workflow systems. Users are building custom harnesses with text files (SOUL.md, AGENTS.md, etc.) and companies are launching workflow products like Morphic Workflows with 72+ pre-built workflows.

🔥 AI Agent Security Concerns (68 mentions) - RISING Rising concerns about AI agent security including MCP protocol vulnerabilities with 30,000+ exposed instances, prompt injection attacks through images, and OAuth security issues. Discussion of need for better credential management and containment beyond .env files.

📊 Gemma 4 Release (142 mentions) - CONTINUING Google's Gemma 4 model release with 26B and 31B variants, featuring extensive ecosystem collaboration with HuggingFace, VLLM, Ollama and others. Community is evaluating performance and comparing against other models like Qwen 3.5.

📊 Cognis AI Memory System (38 mentions) - CONTINUING Milla Jovovich co-developed Cognis, an open-source AI memory system achieving 92.4% on LongMemEval benchmark with top rankings across LoCoMo categories. Features contradiction resolution, temporal reasoning, and multi-hop queries with both hosted and SDK options.