Asana Acquires No-Code Agent Builder Stack AI
According to TechCrunch, project management platform Asana has acquired Stack AI, a no-code tool for building AI agents. Asana reportedly plans to integrate the acquired Stack AI technology into its existing suite of AI workflow tools.
Building agents with no-code tools has the advantage of letting even teams with limited engineering resources incorporate AI agents into their operations. By offering agent-building capabilities as an extension of its core strength in project management, Asana could take workflow automation a step further. As "who builds the agents" becomes a key challenge in enterprise AI adoption, no-code approaches are likely to gain prominence.
1T-Parameter Model "Mimo 2.5 Pro" Hits 40 t/s on a Consumer GPU Cluster
A report gaining traction on Reddit's LocalLLaMA community claims that Mimo 2.5 Pro (a 1-trillion-parameter model) achieved 40 tokens per second when run on a cluster of eight NVIDIA Spark/GB10 units.
According to the report, the results were 40 t/s at 1K context, 32 t/s at 30K context, 25 t/s at 125K context, and 17 t/s at 250K context. With parallel requests, it reportedly reached 60 t/s with 2 concurrent requests and 83 t/s with 4. The fact that a giant 1T-class model can deliver this level of throughput on consumer hardware shows that local LLMs are rapidly becoming practical.
NVIDIA's GB10 (Spark) platform is offered at a relatively accessible price point, and the cost of running large models locally is steadily coming down. The fact that it maintains practical speeds even with long contexts is particularly noteworthy.
"LLM Smells" — Analyzing Anti-Patterns in LLM Usage
An article titled "Various LLM Smells" has been drawing attention after collecting 26 points on Hacker News. Drawing an analogy to "code smells" in software engineering, it lays out common anti-patterns and inappropriate ways of using LLMs.
Now that LLMs are in wide use, discerning what they should not be used for is becoming just as important as knowing what they can do. This kind of accumulated practical insight should help safeguard quality when putting LLMs to work in production environments.
Self-Improving AI Research "SIA" Appears on arXiv
A paper titled "SIA: Self Improving AI with Harness and Weight Updates" has been published on arXiv. As the title suggests, it proposes a mechanism by which AI models improve themselves through harness (evaluation framework) and weight updates.
Self-improving AI is a concept in which models improve their own performance without additional external data — a research direction that pushes AI autonomy to a new level. It is still at the research stage, but if it becomes practical, it could automate the model update process itself. That said, the stability and safety of self-improvement will likely require further verification.
New Intelligence Agency to Monitor Critics of AI Data Centers
According to reporting by Ken Klippenstein, a new intelligence agency has been established with critics of AI data centers as targets of surveillance. As concerns grow over the environmental impact and effects on local communities accompanying the expansion of AI infrastructure, the move to step up surveillance of the critical side deserves attention.
AI data centers have drawn backlash in many regions over electricity consumption and water resource issues, and increased surveillance could spark debate from the perspectives of free speech and civil oversight. How to balance technological progress with social acceptance is a key theme for the sustainability of the AI industry.
---