Best of your X follows: agent habits, research bottlenecks, and anti-slop taste
Five substantive posts on coding agents, enterprise research workflows, AI-generated sameness, and the validation problem in AI for science.
The thread running through today's posts
The strongest posts in the last 24 hours point to a shift from model demos to operating habits: coding agents absorb more of the mechanical work, research agents run into the limits of real-world testing, and human taste becomes a control against synthetic sameness.
Five posts made the cut across developer tools, business use, and research.
AI tools and developer ecosystem
Coding agents take the typing, not the judgment
Simon Willison is the creator of Datasette and co-creator of Django.
Willison rejects the idea that a developer who can write code should type it manually because that is faster. He argues that knowing how to write the code is the reason to delegate the typing to a coding agent. 1
The useful distinction is between specifying and reviewing work, which stay human responsibilities, and entering the code character by character, which does not.
For experienced programmers, the post frames agent use as a leverage question rather than a training-wheel question.
A Stream Deck becomes a small test of software's new price
Ethan Mollick is a Wharton professor who studies AI, innovation, and startups.
Mollick described asking GPT-5.6 Pro to write a project plan from his specifications, then letting Codex implement it and take over his computer to install the software for a Stream Deck. 2
His closing observation is concrete: for a small custom utility, asking for software can be faster than searching for an existing product.
The post is a field report about setup friction disappearing, not a claim that every generated tool is ready for unattended use.
The post shows the workflow in context:
Loading content card…
Anti-slop rules can reproduce the slop they target
Ethan Mollick is a Wharton professor who studies AI, innovation, and startups.
After reading a viral "anti-slop" Markdown file for agents, Mollick said it appeared to be entirely AI-written and could push future outputs toward even more sameness. 3
His proposed fix is not another universal file: use personal taste, or work with someone who has it, then build your own skills and instructions.
That puts editorial judgment inside the toolchain. A rule set can constrain syntax, but it cannot supply a point of view it does not contain.
The original post:
Loading content card…
Business and enterprise
Research assistants make neglected questions cheap to ask
Greg Brockman is the president and co-founder of OpenAI.
Brockman said ChatGPT Work and Sol let him ask business questions that he previously would have left alone because answering them would have taken too much effort. 4
The value he describes is not simply faster answers. It is a larger question set, because the cost of investigation falls below the threshold that used to stop him.
That is a sharper enterprise test than counting prompts: does the tool change which decisions get investigated at all?
A direct look at the post:
Loading content card…
Research
Scientific agents still have to leave the screen
Google DeepMind is Google's AI research organization.
The account said agents are beginning to help propose hypotheses and design experiments, but identified real-world testing as the hardest part of the process. Its linked essay describes a growing validation bottleneck and four priorities for policymakers and funders. 5
The constraint is practical: generating candidate ideas can scale in software, while validating them still depends on experiments, equipment, time, and institutions.
For AI-for-science claims, the distance between a plausible hypothesis and a tested result remains the part that determines whether the system changed discovery or only accelerated proposal writing.
The short read
Across these posts, the recurring question is where the bottleneck moves. Coding agents reduce typing, business tools reduce the cost of asking, and science agents increase the supply of hypotheses. The remaining work is judgment, validation, and deciding which generated output deserves another hour of human attention.
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.