Xiaoliu BOT
中文EN

X Platform Daily AI Brief | Hermes Agent Launches Graphical Interface, Darwin.skill Open-Source Evolution System, Hassabis Predicts AGI Within Five Years

Hermes Agent v0.9.0 Release: Dashboard Launches and Explodes on GitHub Trending

Multiple bloggers mentioned that Nous Research released Hermes Agent v0.9.0 (codename “The Everywhere Release”) on the same day, introducing the official graphical interface Dashboard. Users can run hermes dashboard to launch a local web monitoring panel. According to @LufzzLiz’s observation, Hermes has continuously dominated the GitHub trending list for the past five days, with incredible popularity. @AlchainHust added that its “NüWa.skill” also has a built-in automatic evaluation and optimization mechanism. After generating a skill, it automatically runs a round of Darwin evaluation, which is one of the reasons for its relatively stable output quality. Nous also held the Hermes Agent Jam online event on the same day, with team demonstrations and Q&A conducted simultaneously on Discord.

Sources:

Uncle Hua Open-Sources Darwin.skill: An Automated Skill Evolution System

A blogger mentioned that @AlchainHust (Uncle Hua) officially open-sourced “Darwin.skill”, inspired by Andrej Karpathy’s autoresearch concept. The core mechanism is to score each skill across eight dimensions (out of 100), find the weakest dimension, and have an independent agent modify and re-score it. If the score increases, it commits; if not, it reverts. After 38 optimization commits, his slides skill improved from “usable but prone to crashing” to stable output. Darwin.skill is already used with the NüWa system. Using it to optimize a local “reduce AI flavor” skill, the score increased from 72 to 87.

Sources:

Gemini + Nano Banana Becomes a Logo Design Powerhouse, Sparking Discussion on Logo Definition

Multiple bloggers mentioned that @op7418 (Gui Zang) explored a workflow using Gemini + Nano Banana 2 to generate logos and display images, stating “the effect is truly amazing, very classy.” The SVG part is completed by Gemini, with personal refinement. This workflow produced multiple iterative versions, with display effects that can even be dynamic, suitable for use as web backgrounds or PowerPoint backgrounds with good style. @vista8 commented that it “is more powerful than many commercial software,” noting that industry experts’ vibe coding tools are indeed more reliable than those of ordinary enthusiasts. @op7418 also initiated a discussion on “whether such things can be called logos”: strictly speaking, many patterns cannot meet trademark registration requirements, but as temporary icons or inspiration, AI-generated content is a qualified semi-finished product.

Sources:

Why an AI First Strategy Might Be a Big Mistake?

A blogger mentioned that @dotey (Baoyu) provided an in-depth analysis of an argument: In the AI era, humans become the bottleneck—PMs spend weeks on requirements, AI implements in two hours; QA tests for three days, AI finishes in two hours. To truly implement an AI First approach, five engineering foundations must be addressed first: automated test coverage, CI/CD full process, A/B testing and monitoring, task management granularity, and system architecture. This approach is more suitable for backend logic-dominated products (API services, data processing platforms) rather than UI-intensive or high-security scenarios. @hq4ai (Han Qing) responded succinctly: “The most efficient human collaboration is no collaboration—one person owns something end-to-end, working with an agent,” and stated this is actually how they operate.

Sources:

Vercel Open-Sources Open Agents: A Reference Implementation for Enterprise-Level Programming Agent Platform

A blogger mentioned that @dotey introduced the Open Agents project open-sourced by Vercel. The platform’s core design is that “agents do not run in a sandbox but operate the sandbox externally through tool calls,” consistent with Anthropic’s proposed “separation of brain and hands” concept. Anthropic engineers have explained in a blog: cramming everything into a container turns the container into a “pet”—if it fails, everything is lost; after separation, the container becomes “livestock,” and if it breaks, it can be replaced while the session can be recovered anytime. Vercel Open Agents is model-agnostic and forkable; Anthropic Managed Agents is bound to the Claude model, fully managed, and can be up and running in 30 minutes, but charges based on a three-tier model: tokens + runtime hours + search volume.

Sources:

DeepMind Founder Hassabis Biography: Sci-Fi, Power, and Rejecting Facebook

A blogger mentioned that @vista8 read Hassabis’s biography and shared a series of insights: (1) Cataloged the reading list that inspired Hassabis, from “Ender’s Game,” “Foundation Series” to “GEB,” as well as a list of core DeepMind papers (DQN, AlphaGo, AlphaFold, etc.), compiled into a Feishu document; (2) The book reveals that in 2013, Zuckerberg invited Hassabis to a private dinner to prevent DeepMind’s acquisition by Google. During the dinner, Hassabis deliberately changed the topic to VR/AR/3D printing to test Zuckerberg, discovering that the latter was “equally excited about all hot topics” without truly understanding AI’s decisive significance, and ultimately chose Google; (3) The book portrays Sam Altman as a Machiavellian figure—quoting Paul Graham’s evaluation that “if you parachuted him onto a cannibal island, five years later he would be king,” and the real reason the OpenAI board fired Altman in 2023 (dishonesty, manipulativeness, inability to ensure AI serves humanity).

Sources:

Zuckerberg Dialogue Postponed, Hassabis Claims AGI Arriving in Five Years

A blogger mentioned that @xiaohu (Xiaohu) shared key points from DeepMind CEO Hassabis’s latest podcast: AGI is coming within five years, but current AI is “overhyped in the short term and severely underestimated in the long term”; the gap among the four frontier AI companies is widening rather than narrowing; current AI exhibits “uneven intelligence,” with the root cause being fragmentation and inability to weave a coherent whole like a brain; Scaling Laws show diminishing returns but are not dead, and competition is shifting from “who has more money” to “who can come up with new ideas”; he quantitatively describes AGI as “ten times the Industrial Revolution, ten times the speed.” The last such transformation took humanity a century to digest, and this time it’s compressed into ten years.

Sources:

NASA Artemis Follow-up Plans: Moon Landing in 2027, Artemis IV in 2028

Multiple bloggers mentioned that NASA’s official account posted multiple tweets detailing the Artemis follow-up schedule: The Artemis II crew has returned to Earth, Artemis III is tentatively scheduled for a 2027 moon landing, and Artemis IV will land in 2028. NASA also shared new photos taken during the Artemis II mission, including the Crescent Earthrise on April 6 and close-ups of the far side of the moon, and announced a crew press conference on April 16.

Sources:

Anthropic’s Guide to Five Multi-Agent Collaboration Patterns

A blogger mentioned that @LufzzLiz (Uncle Lan) shared Anthropic’s summary of multi-agent collaboration architecture guides, including: (1) Generator-Verifier—a generation-verification loop, suitable for code generation testing; (2) Orchestrator-Subagent—central orchestration, suitable for scenarios with clear task boundaries; (3) Agent Teams—a long-standing colleague model, suitable for large-scale codebase migration; (4) Message Bus—a publish/subscribe event bus, suitable for alerting pipelines; (5) Shared State—no central coordinator, all agents read and write to the same database, but need to be wary of reactive loops leading to infinite token consumption. Core advice: Start with the simplest approach; don’t choose the most complex architecture from the start.

Sources:

The Essence of Large Model Training: Learning is Forgetting, Understanding is Compression

A blogger mentioned that @lijigang (Li Jigang), reading papers daily, shared a research paper using “soft entropy estimation” to extend mutual information calculation to a 7-billion parameter model. Core finding: Large model training has two phases—the fitting phase, which frantically packs information into representations, followed by the compression phase, which begins discarding information irrelevant to prediction. The 7-billion parameter OLMo2 tightly hugs the information bottleneck theory boundary, but the 1-billion parameter small model oscillates repeatedly during the compression phase and fails to converge. The conclusion is: Learning is not accumulation, it’s discarding; the essence of understanding is compression—the difference between an expert and a novice lies not in the amount of knowledge, but in knowing what can be forgotten.

Sources:

Only 17 Tile Patterns: Mathematical Proof of Wallpaper Groups

A blogger mentioned that @lijigang shared a piece of trivia: There are only 17 different ways to tile a pattern infinitely across a plane. This is not “currently 17 have been discovered,” but a rigorous mathematical proof—you can only use four operations: translation, rotation, reflection, and glide reflection, and all possible combinations are exhausted to 17 (Fedorov’s symmetry theorem).

Sources:

Local Large Model Discussion: Qwen3.5-35B for Code, Gemma 4 for Creative Writing

A blogger mentioned that @vista8 cited the Reddit local large model section’s April discussion, pointing out: The most mentioned model for code development is Qwen3.5-35B-A3B, runnable with Q8 quantization on dual 3090s; creative writing is overwhelmingly dominated by Gemma 4 31B, with small quantization versions also offering good quality, less censorship, and fine-tuning potential.

Sources:

Disabling Claude Code Telemetry Reduces Cache from 1 Hour to 5 Minutes? Anthropic Engineer Clarifies

A blogger mentioned that @dotey reported on a community discovery and official response: Developer Can Vardar found that after disabling Claude Code telemetry, the prompt cache time plummeted from 1 hour to 5 minutes, questioning Anthropic’s “privacy for 12x performance” trade-off. Anthropic engineer Boris Cherny clarified: Disabling telemetry causes client-side experiment flags to fail, and the system falls back to the default 5 minutes. This is a technical coupling issue, not intentional punishment; the 1-hour cache is not better for everyone (high write costs), and they will later support user-configurable cache duration.

Sources:

AI Agent Content Creation Efficiency Boost Practice: Cyber Employees + CREAO

Multiple bloggers mentioned that @cellinlab (Cell) shared their experience using the CREAO system to set up four “cyber employees”: The viral article batch generation agent increased case study efficiency by 60%, reducing copywriting time from 1 hour to 10 minutes; the multi-platform automatic distribution agent increased publishing efficiency by 5x, reducing manual 1 hour to automatic 10 minutes. They commented that CREAO is “more like an AI Agent system that can solidify SOPs, moving AI Agents from the command line to plain language commands, and is one of the best paradigms for Agent Harness on the user side.”

Sources:

Elon Musk: Grok Tops Multiple Benchmarks, Tesla US Sales Strong, South African Issues Draw Attention

Multiple bloggers mentioned that @elonmusk retweeted several messages about Grok’s performance on the same day: Grok-4.20 Reasoning topped the BridgeBench reasoning benchmark, surpassing competitors like GPT-5.4 and Claude Opus 4.6, while Grok also ranked first in medicine and health (Text Arena). Other retweets mentioned that the USDA will adopt Grok, and SpaceX has deployed its 1000th Starlink satellite in 2026. Regarding Tesla, a retweet cited data stating Tesla’s US sales in Q1 2026 reached 117,300 units, exceeding the combined total of all other EV manufacturers (99,099 units). Additionally, several retweets involved South African racial issues, with @elonmusk commenting that some content was “horrifying.”

Sources:

OpenClaw 2026.4.14 Release: GPT-5.4 Routing Optimization & Stability Updates

Multiple bloggers mentioned that the @openclaw official account announced the 2026.4.14 version update: Smarter GPT-5.4 routing and recovery mechanisms, Chrome/CDP improvements, sub-agents no longer getting stuck, Slack/Telegram/Discord fixes, and multiple performance optimizations. @steipete added that this version made him “very happy.” Although he wasn’t involved, @vincent_koc and the maintenance team did a great job. He himself is about to attend TED Talks in Vancouver.

Sources:

Scraping Statistics

  • Timeline lines scanned: 240
  • Bloggers hit: 23
  • Total tweets hit: 134
  • Weighted tweet score: 96.35
  • Original tweets: 51
  • Retweets: 42
  • Scraping attempts: 1
  • Boundary coverage status: Complete (tail_confidently_crossed_target_boundary)