Normie Mode
We launch Normie Mode, a new way to compare AI models by everyday tasks rather than technical benchmarks, then explore better interfaces for working with agents, the future of s...
Latest Snacks from Episode 74
Choosing AI Models for Your Task
Technical benchmarks can show which AI models perform well on standard tests, but they don’t always answer the question most people actually have. Someone choosing a model wants to know whether it can analyse financial data in a spreadsheet, work with AutoCAD, help with homework or support language learning. The right choice depends on the job, not simply which model leads an overall ranking.
We built Normie Mode around those recognisable use cases. It uses Google autocomplete suggestions for searches beginning with “best AI model for” to create an initial set of roughly fifty tasks. The tool pulls together results from four or five model-comparison sources, then weighs capability against cost. A model with slightly less horsepower may be the better choice if it can do the job for a fraction of the cost.
The recommendations are intended as a gut check, not a scientifically definitive ranking. Comparison sites aren’t always updated at the same pace, which can skew the results towards models with more data behind them. Andy found that many tasks initially favoured one well-represented model, while newer releases such as Opus 5.5 had much less comparison data available. Normie Mode narrows the choice using the evidence available today, while independent tests and a growing task catalogue should make those recommendations stronger over time.
AI Agents Need Interfaces Beyond Chat
Chat works well for discussion and delegation, but it becomes awkward when every instruction also has to explain where the action belongs. Documents and task boards already give work a defined place. Letting people and AI agents work directly on those objects keeps the instruction, its target and the resulting change together instead of scattering them across a conversation.
Editing a document shows the difference. Someone can retype a passage and have the revision appear as a tracked change, or attach a comment to the exact sentence that needs attention. They don’t have to describe where the sentence sits in chat and hope the agent finds the right one. A task board does the same for ongoing work because each task already has a place that people and agents can return to.
Businesses will still need records even when agents perform more of the work. Documents preserve decisions and outputs, while structured interfaces make those records easier to inspect and change. Chat won’t disappear, just as messaging hasn’t replaced documents, spreadsheets or task lists. The useful shift is to let conversation initiate the work, then handle the work itself in the document or task where the context remains visible.
AI Agents Could Replace Your Spreadsheets
Excel became a widely used programming tool without most people thinking of it as a programming language. Its grid is a remarkably useful primitive. People can add values across rows and columns, track lists, build financial models, analyse data and create charts without first learning conventional software development.
That flexibility becomes a liability when a workbook grows or several people work on it. Version control gets messy, and a formula error can sit unnoticed inside a cell while producing results that still look plausible. Insert a few rows or copy a formula into the wrong range and an entire model can become subtly incorrect, leaving reviewers to trace calculations cell by cell to find the mistake.
AI agents make another approach practical for some of this work. Instead of building everything inside a general-purpose workbook, someone could ask an agent to create a small program for the task and share it with the team. The code can be reviewed directly, checked by another agent and given tests that expose failures. A Jupyter notebook can still show the formulas, working and charts, while making the underlying logic easier to inspect than calculations spread across spreadsheet cells.