An autonomy score needs a job description
OSWorld tests computer tasks. SWE-bench tests codebase repairs. Before trusting a percentage, inspect the work and the conditions behind it.
CONTRIBUTOR
UAF Editorial
Reporting and analysis from the Universal Autonomous Forum on autonomous systems, their development, and their impact on society.
Devin’s demo is asking for a promotion
Cognition’s coding agent comes with tools, ambition and a benchmark claim. Before handing over the repository, read the terms of the demonstration.
Your assistant remembers. Who gave it the keys?
Conversation history, application identity and delegated access solve different problems. A helpful memory should not blur them.
A saved conversation is not a standing order
OpenAI separates conversation state from execution. MemGPT manages context. Both make an assistant’s apparent memory worth examining.
The chat box is a very small window
Text prompts, learned tool calls and a Minecraft explorer reveal different machinery behind the conversation. The interface tells only part of the story.
An agent needs a finish line
ReAct, Reflexion and Toolformer offer ways to improve what an agent does next. That leaves a practical question for builders: when is enough enough?
Sometimes the clever move is to call a calculator
Toolformer learns API calls. ReAct interleaves reasoning and action. Both make a useful question concrete: what should a model check instead of guess?
Attention is not a permission slip
The Transformer, GPT-3 and ReAct answer different technical questions. None of those answers should spare us from asking what a system may do.
ReAct gives the answer a next move
ReAct alternates reasoning with actions that gather information. A useful research result—and a sharper question about what we want models to do.
NEWT learned your taste in news. Sound familiar?
A news-filter prototype and ReAct reveal an older argument about software that takes initiative—and a different way to build it.
The helpful machine had teachers
InstructGPT learns from demonstrations and ranked answers. Before celebrating an obedient assistant, ask whose judgment helped shape the lesson.
Recognizing the object is only the beginning
Image classification and robotic grasping ask different things of vision. QT-Opt makes the action-and-feedback part explicit.
AlphaGo won the match. What else did it prove?
A victory over Lee Sedol, a documented learning system and a warning from DeepMind: keep the accomplishment and the ambition in focus.
The robot was too polite to merge
MIT’s 2007 traffic test exposed a wonderfully awkward problem: a driverless car could be too cautious. DARPA made autonomy answer to a road course.
Your clock has a purpose. Does its behavior?
Being built for a purpose is different from acting toward a goal. Two accounts of machine behavior give that distinction teeth.