Bringing you the best information on AI so you can stay updated. Clear, practical breakdowns of models, agents and tools. Stay informed. Stay ahead.
United States
Joined November 2022
- Tweets2.6K
- Following406
- Followers661
- Likes902
The AI Engine Room retweeted
18 models. One leaderboard. The biggest names in AI go head-to-head inside Hermes.
The Hermes Index now compares 18 models across four benchmark suites, helping you see how they perform on real agent tasks.
And a reminder: Hermes works with most of your AI subscriptions and providers, including Nous Portal, ChatGPT, Claude, OpenRouter, OpenCode Go, and many more.
Connect your preferred providers to Hermes, then use the Index to help decide which models to run.
Hermes Index has been updated to include GPT 6.1 Sol and several other models!
Check it out here:
portal.nousresearch.com/benc…
The AI Engine Room retweeted
OMG… the Tesla Roadster just went upside down on a half pipe and landed!
The AI Engine Room retweeted
Announcing Grok 4.5, our first model trained specifically for coding and agents. It was trained with Cursor and offers frontier intelligence at leading speeds and cost efficiency.
x.ai/news/grok-4-5
The AI Engine Room retweeted
Customers are not an abstraction for us: we exist to help enterprises, public institutions, and industries build their own intelligence, so the value created from their data, workflows, feedback, and models accrues to them rather than to model providers.
The AI Engine Room retweeted
🐱 LongCat-2.0 is now fully open-source — MIT licensed, no restrictions.
Since our launch a few days ago, the response from the community has been incredible.
Thank you for all the feedback, discussions, and interest.
Today, we’re releasing the model weights and inference code to everyone.
◆ 1.6T MoE · ~48B active · 1M token context
◆ Agent-native: Integrates directly with Claude Code, OpenClaw, and Hermes Agent
◆ Deployment: Support both GPU and NPU platforms— verified on large-scale domestic clusters
📑 Tech Blog: longcat.ai/blog/longcat-2.0/
🤗 HuggingFace: huggingface.co/meituan-longc…
💻 GitHub: github.com/meituan-longcat/L…
🪄 ModelScope: modelscope.ai/collections/me…
👇 Inference Code
GPU: github.com/sgl-project/sglan…
NPU: github.com/meituan-longcat/S…
Introducing LongCat-2.0 🐱
1.6T parameters · MoE with ~48B active · 1M context
The full model behind Owl Alpha on @OpenRouter — now available.
Built for agentic coding from the ground up:
◆ LongCat Sparse Attention (LSA) — scales efficiently for 1M-context tokens
◆ Zero-Compute Experts — dynamic activation 33B–56B per token, zero wasted compute
◆ MOPD — three specialized expert groups (Agent / Reasoning / Interaction), gate-routed per task
How it stacks up:
→ Terminal-Bench 2.1: 70.8
→ SWE-bench Pro: 59.5 (GPT-5.5: 58.6)
→ SWE-bench Multilingual: 77.3
→ FORTE: 73.2 · RWSearch: 78.8 · BrowseComp: 79.9
📖 Tech Blog: longcat.chat/blog/longcat-2.…
Try it across different scenarios 🧵👇
The AI Engine Room retweeted
The fastest Physical AI growth may be inside labs, not factories.
IFR World Robotics 2025 shows medical robots had one of the sharpest jumps in 2024:
• Medical robots overall: 16,700 units sold
• Growth: +91% over 2023
• Surgery robots: +41%
• Rehabilitation and non-invasive therapy: +106%
• Diagnostics and medical laboratory analysis: +610%
That last number is the real signal.
Labs are controlled environments. The workflows are repetitive. The value is measurable. The need is rising because hospitals and medical labs face labor pressure and growing demand from aging populations.
This is Physical AI with fewer distractions than humanoid demos.
No factory floor hype.
No vague autonomy claim.
Just robots handling structured medical workflows where precision and throughput matter.
Source: IFR World Robotics 2025
You're unable to view this Tweet because it's only available to the Subscribers of the account owner.
The AI Engine Room retweeted
A paper on open model adoption and finds Chinese models, led by Qwen, now dominate.
China passed the U.S. in open model downloads in summer 2025.
By March 2026, Chinese models had 1.15B tracked downloads, while U.S. models had 723M.
Most of that shift comes from Qwen, which became the default base family for many builders.
Qwen’s lead is not just about 1 famous model; it comes from having many useful models across many sizes, especially small ones that people can run cheaply and reuse often.
DeepSeek matters in a different way: it does not dominate the whole open model ecosystem like Qwen, but it leads the very large model category, especially models above 250B param
After adjusting for model size and age, some U.S. models like GPT-OSS 120B and Nemotron Super 120B still show very strong adoption momentum, even though China leads overall.
----
Link – arxiv. org/abs/2604.07190
Title: "The ATOM Report: Measuring the Open Language Model Ecosystem"
The AI Engine Room retweeted
Open-source memory for AI agents just reached parity with the best published results. Alice is a local-first memory + continuity layer, MIT-licensed. 64.6% on LongMemEval with the receipts: per-question evidence and a reproduction script committed to the repo. 🧵
The AI Engine Room retweeted
David Soria Parra (MCP co-creator @ Anthropic) called this out:
"People blame MCP for context bloat. But we already know the fix — progressive discovery."
MCP supports it. MCPZERO ships it. Your context window will thank you.
github.com/mcpzero/mcpzero
The AI Engine Room retweeted
Stop feeding whole repos to AI agents.
Frigg gives Claude, Codex, Cursor, and other MCP clients code context they can actually navigate: search, symbols, definitions, references, efficient bounded reads.
Try the OSS local and CI workflow context engine:
github.com/bnomei/frigg
The AI Engine Room retweeted
Everyone wants AI agents.
Almost nobody wants to maintain them.
The demo takes days.
Production takes months.
- Monitoring.
- Retries.
- Prompt versioning.
- Fallbacks.
- Human approvals.
That's where the real engineering starts.
What's been the hardest production AI problem you've faced so far?
The AI Engine Room retweeted
The more I learn about LLM evals, the more I realize accuracy is only the first layer.
The real questions are:
can I reproduce the result,
explain the failure,
track drift,
and know when the system should not answer?
The AI Engine Room retweeted
Agentjacking: Hacking AI Coding Agents gentoro.com/blog/agentjackin…
The AI Engine Room retweeted
HIC Mouse targets the messy middle of AI coding: coordinate-based file edits, staged changes, and atomic rollback so agents leave auditable files. Pricing starts at $15/mo.
tinyurl.com/e8uxepe5
The AI Engine Room retweeted
Security for AI agents continues after deployment. Explore operational practices for data protection, audit-grade observability, regression testing, change governance, and multi-agent trust boundaries. social.ora.cl/6010BDCk3j
The AI Engine Room retweeted
AI agents don't need better models. They need better rails. Permissioned spending, verifiable identity, on-chain settlement. That's the unlock.
The AI Engine Room retweeted
litellm-labs/litellm-agent-control-plane: 1 place to call all your agents - OpenCode, Hermes, Claude Managed Agents, Cursor Agents API, DeepAgents. github.com/LiteLLM-Labs/lite…
The AI Engine Room retweeted
Give agents machine-readable business context with Fabric IQ in Microsoft Fabric. See how. youtu.be/vi1JZvA21FE Build enterprise-ready AI agents that scale without sacrificing security or control using Microsoft Azure. Establish a shared Governance Hub to centralize model access, MCP catalogs, and policy enforcement, then give every agent a traceable identity, runtime protection, and data governance through Agent 365. #AgentArchitecture #microsoftazure, #ai, #azure, #cloudcomputing, #generativeai
The AI Engine Room retweeted
The MCP ecosystem is maturing fast enough that it now has a software architecture paper.
A new arxiv preprint (June 29, 2026) catalogued five recurring patterns across fifteen independently-developed MCP servers in production: Resource Gateway, Tool Orchestrator, Stateful Session Server, Proxy Aggregator, and Domain-Specific Adapter. The authors drew from Celabe's production servers plus ten from the official MCP registry.
The most actionable finding: tool-selection accuracy drops below 90% between 10 and 15 tools per context for Claude Haiku 4.5, and between 20 and 30 tools for Claude Sonnet 4. Beyond that range, the LLM is guessing more than selecting.
This is not a model limitation. It is an API design problem. The same issue appeared when REST APIs got too coarse: you bundle too many actions into one endpoint and the client starts making wrong assumptions. The fix was decomposition, scoped resources, and clear naming conventions. The fix for MCP is the same: Domain-Specific Adapters and Proxy Aggregators that partition tools into coherent namespaces rather than one giant tool registry.
My read on the broader trend: this paper is a signal that the MCP ecosystem is about one year behind REST in maturity. REST had informal patterns from 2005 to 2008, then Roy Fielding's dissertation and Richardson's Maturity Model gave the community a shared vocabulary. Swagger (now OpenAPI) followed in 2011. The MCP equivalent of OpenAPI does not exist yet. Someone will build it.
The four anti-patterns flagged (authentication gaps, versioning failures, unobservable tools, and missing error contracts) read like the exact failures that broke early microservice architectures. The solutions are also the same: explicit contracts, scoped permissions, observable state.
If you are building MCP servers, this is the first paper worth treating as a design reference.
teapawt.org/TheTuringPost/status/2…
One of the most practical MCP papers we've ever seen
It's useful for anyone building MCP servers who's unsure how to structure them well.
Here are the main tips:
1. Don't use too much tools, because tool-selection accuracy drops below 90% after 10–30 tools.
2. See MCP servers as a software architecture problem.
The researchers documented 5 emerging architecture patterns:
• Resource Gateway – exposes existing data sources through MCP
• Tool Orchestrator – coordinates multiple tools behind a simpler interface
• Stateful Session Server – maintains conversation or workflow state across requests
• Proxy Aggregator – combines capabilities from multiple MCP servers into one endpoint
• Domain-Specific Adapter – wraps a complex external API into LLM-friendly tools
It's also interesting that MCP itself adds almost no latency (0.01 ms via stdio, 0.39 ms via HTTP). So real-world performance is mostly limited by the network.
The AI Engine Room retweeted
X shipping a hosted MCP is the boring but important part of agents.
The moat is not another chat box. It is permissioned access to real workflows: auth, scopes, rate limits, auditability.
Agents get useful when tools become inspectable surfaces.
techcrunch.com/2026/06/30/x-…
