Eliza Times

Industry News

• Anthropic launches Claude Code Security: Limited research preview scans entire codebases for vulnerabilities and suggests targeted fixes, reportedly finding 500+ zero-days in heavily-fuzzed open source code. Wall Street reacted immediately with cybersecurity stocks dropping 8-9%. link

• Emad Mostaque's new book released: The former Stability AI CEO has published a new work, though details remain sparse in the tweets. link

• Germany's AI congress criticized as "boomer daycare": Attendees describe the event as lacking technical depth, raising concerns about Europe's competitiveness in AI development. link

Tips & Techniques

• Parallel Codex workflow with git worktrees: Developers are running 10-20 VS Code windows across multiple worktrees and branches simultaneously, with some considering color-coding systems to manage cognitive load. The key is learning to context-switch effectively across parallel agent sessions. link

• Security audit your OpenClaw setup manually: Despite agent capabilities, manually reviewing configuration files gives you better understanding of controls and prevents configuration interaction issues that agents might miss. link

• Codex plan mode reasoning levels now configurable: Users can set separate reasoning levels for planning vs implementation directly in config files (e.g., plan_mode_reasoning = "high"), allowing Plus users to save usage limits by planning with high reasoning and implementing with medium. link

• Agent PR review at scale with parallel analysis: Running 50 Codex instances in parallel to analyze PRs, generating JSON reports with vision-based comparison, intent signals, and risk assessment. The reports can then be queried, deduplicated, auto-closed, or merged as needed without requiring a vector database. link

• "Taste-driven development" as the new bottleneck: With code generation becoming abundant, the constraint shifts from typing speed to judgment—your internal reward model trained over years determines which diffs you accept and how you structure systems. Agents increasingly condition on your priors and approximate your aesthetic. link

New Tools & Releases

• Taalas inference chip delivers 8x speed boost: Hardware running Llama 3.1 8B at 15.4k tokens/sec (8x faster than Cerebras), with frontier open-weight models promised this year. If real, this represents a 2.5 OOM improvement that could make timelines incomprehensible. link

• H100 GPUs collapse 85% on secondary market: Once $40,000, now selling for ~$6,000 on eBay as inference shifts away from traditional GPU clusters toward specialized chips and architectures. link

• simdjson shows 30% additional performance gains: New optimizations demonstrate further headroom in SIMD-based JSON parsing, continuing the library's track record of squeezing performance from modern CPU instructions. link

Research & Papers

• Bytedance releases "Diffusion as Chemistry" framing: New paper argues LLM reasoning failures resemble chemical reaction failures—when complexity increases, models "flinch" measurably with 55.6% shutdown probability and 2/3 response length drop. Challenges claims that models have "no awareness of error." link

• Chomsky's Universal Grammar formalized mathematically: Turns out it can be expressed via quantale weakness and logical semantics, connecting cleanly to vector embedding models and offering new perspective on transformer architecture. link

• MIT releases practical diffusion models course: Free video lectures for IAP 6.S183 covering foundational to applied diffusion techniques, publicly available for self-study. link

--- *Curated from 650+ tweets across developer and AI research communities*

---

Emerging Trends

🔥 Codex vs Claude Code Competition (85 mentions) - RISING OpenAI's Codex CLI is rapidly improving and competing directly with Anthropic's Claude Code for AI-powered coding, with polls showing close competition (54-46) and users debating which is superior. Version 0.150.0 brought major quality-of-life improvements including configurable reasoning levels.

🔥 Gemini 3.1 Pro Release (195 mentions) - RISING Google's Gemini 3.1 Pro launched with strong performance on image-to-code tasks, skeuomorphic UI generation, and SVG work. Users testing it across investment strategies and creative workflows, with particular praise for visual design capabilities and AlgoTune benchmark score of 2.02x.

🔥 Claw Ecosystem & Alternatives (165 mentions) - RISING The term "Claw" becoming standard for OpenClaw-like agent systems. Karpathy's endorsement of alternatives like NanoClaw (4000 lines vs 400k), which uses skills for configuration instead of config files. Multiple alternatives emerging: nanobot, zeroclaw, ironclaw, picoclaw. Discussion of "skills as one-time execution" pattern for agent self-modification.

📊 OpenClaw Security Crisis (280 mentions) - CONTINUING Continued security concerns around OpenClaw with reports of malware in ClawHub (1,184 malicious packages), RCE vulnerabilities, and exposed instances. Karpathy and others express reluctance to run it due to security nightmares, while alternatives like NanoClaw gain attention for their smaller, more auditable codebases.

📊 Llama.cpp Team Joins Hugging Face (88 mentions) - CONTINUING Major announcement that the llama.cpp team is joining Hugging Face, described as a "match made in heaven" by the community. Significant development for open-source AI infrastructure and local model deployment.