Magnificence and story are key to my workflow harness

I believe my harness-workflow is taking form. I believe I’ve gotten to some core rules, what motivates them, and the place I’m going. Thanks once more for being a subscriber. I admire everybody’s assist.

A handful of rules are defining my new workflow.

The primary one is an assumption. I assume that every day, when utilizing brokers, I’ll begin my work from a resting state of amnesia. I cannot have a dependable reminiscence of the work I did earlier than. That is partly as a result of I work utilizing dialogue, reflection, and soliciting opinions from Claude Code. Conversations don’t go away a document a lot as they go away a sense, and as such I can not simply recall what we did. However amnesia can also be the pure byproduct of the separation of verification from manufacturing, and the way that cleavage impacts the feeling of understanding the work nicely.

And so I deal with amnesia as a everlasting characteristic going ahead defining my new resting state. It’s the new natutal state, and subsequently it’s the first presupposition of the workflow, and no matter I construct, will probably be with structural amnesia in thoughts.

Secondly, on a degree that bears repeating advert infinitum, AI brokers have “flattened the isoquant” for the manufacturing of cognitive output. For simplicity, assume a manufacturing perform primarily based on machine hours and people hours that offers linear isoquants. This implies one can produce a manuscript that might be submitted to a journal utilizing solely machine time. It can’t be printed per se, however it may be submitted.

Quantity three. I’m extra productive with AI holding fixed human time use. That means, although I might substitute machine time for human time, there’s some sense through which if all I did was preserve the identical human time, H*, at cognitive duties, however change to integrating with AI brokers, I’d know extra. I’d produce “extra” in no matter sense. Extra papers, higher papers. For this I don’t actually care how we particular it; extra or higher is ok.

However, as a result of it’s a linear isoquant, and the chance price of my time is optimistic, and the marginal price of utilizing AI brokers is definitely zero (we pay by the month, not by the tokens, as much as some level), the pure tendency might be for my time to fall. Does it fall to the nook answer the place my cognitive output is solely and solely produced with machine time? Does it should? That may be the Social Catalyst Lab’s APE — automated program analysis undertaking. However let’s don’t go to the nook simply but. As a result of look, it’s potential to scale back human time and nonetheless study extra.

I name this the secure zone. It’s secure as a result of when H falls between H-bar and H*, it’s good as a result of I’ve reclaimed a few of my invaluable time by lowering human time beneath that authentic level, H*, however I’m nonetheless studying extra I’m nonetheless making extra, I’m doing higher work.

As long as as H-bar.

However what if my time, H, falls beneath H-bar? What then?

I name this the depreciation entice. Why? Nicely for one I’m producing cognitive output however not understanding it, and my experiments have proven that we’re nowhere close to a spot the place we are able to take our fingers off the steering wheel for quantitive empirical social scientific utilized econometrics you identify it. I’ve an experiment the place I can present Claude Code is specification looking when given even the tiniest discretion. I’ve to complete up the revision and I’m presenting twice within the subsequent two weeks, so I ought to be capable to share quickly. However for now, simply take me at my phrase.

However why is that this harmful? As a result of, as human time falls, human capital falls, as a result of human capital is growing in attentive time.

Within the previous days, cognitive output trusted human capital, which trusted prior consideration, which trusted human time. However, for the reason that isoquants have flattened, I can produce cognitive output with out consideration, with out human capital, and thus my capacity to confirm the output is severely hampered.

So, is it optimum to have zero human time on analysis? Most likely not. It’s most likely not optimum for anyone to set human time, H, beneath their very own private H-bar. That’s their (and theirs solely) “reservation time”. You shouldn’t produce analysis with attentive time beneath your reservation time, H-bar.

Is that an absolute quantity? I want. It’s decidedly not and so nobody can say to another person the place it’s. It’s primarily based on the Cobb-Douglas, multiplicative inputs, manufacturing know-how that defines the manufacturing of cognitive work (borrowed from Shockley 1957):

Think about that these 8 inputs, F1 to F8, when multiplied collectively deterministically produce a bit of cognitive output — no, let me rephrase that. A chunk of verified cognitive output.

Discover the linkage within the manufacturing perform between manufacturing (provide) and verification (demand). His output actually is an equilibrium of provide (manufacturing) and demand (verification) as a result of discover there’s the researcher actions taken — consider a superb drawback, work on the issue — and there are the actions which are clearly being taken by the journal and others. Others are providing criticism; the actor should select whether or not or to not revenue from it. The journal’s editors and referees select to provide revision requests or rejections. The actor should select whether or not to persist in it. So he’s collapsing just a few issues surrounding the equilibrium into the actions taken by the cognitive employee, however put that apart for a second as a result of that is actually the employee’s facet of the equation. It’s simply the employee in battle and cooperation with the demand facet — i.e., the journal.

The issue is we don’t know the place H-bar is. And even when H-bar my hunch is many individuals expertise confusion as a result of their very own manufacturing has turn into some fully “un-bundled” that the psychology of verification is gone. They don’t have any skilled the elements of analysis that used to create the psychology of understanding. And so the sensation round what you’ve finished and never finished is gone. And my hunch is that as human time falls however manufacturing from machine time will increase, that feeling approaches zero.

So what can I do? What ought to I do?

Nicely, let me begin by saying with what I’m not going to do. I’m not going again to the way in which issues have been. There isn’t a state of affairs wanting human civilization failing that I’m going again to my type of labor earlier than Claude Code. In order that’s not taking place.

Secondly, I cannot tolerate a manufacturing perform that creates errors. And since errors are created below uncertainty, not deliberately, I’ve to create a manufacturing perform that reintroduces understanding, not only one that produces. How do I do this? By making zero errors the constraint, not the objective.

1. Zero errors because the endogenous alternative variables is when zero error is a objective. When it’s a objective, we act to reduce errors topic to some mounted constraint, most likely treating the manuscript because the isoquant itself. We do analysis and interact in “constrained error minimization”. Constrained error minimization can discover non-zero errors are optimum, and so it received’t work for me.

2. Zero errors because the constraint can not discover non-zero errors to be optimum and even potential. Zero errors because the constraint dictates the form of analysis itself, what groups to make use of, the place to take a position ones talent, and what new abilities to develop, and what previous abilities to desert. When maximizing analysis topic to the constraint that there be zero errors within the work, the analysis quantity and high quality varies in unknown methods, and will even be bigger within the longrun given errors can jeopardize one’s profession.

My newest experimentation is to create a workflow that’s combines just a few rules guiding analysis right into a dashboard and a guidelines that helps me visualize the undertaking and shops my undertaking’s reminiscence. The primary two pillars are:

1. I all the time get up with amnesia

2. Love all the time remembers

In “50 First Dates”, Drew Barrymore has no capacity to generate new recollections because of a traumatic mind damage a few years earlier. She’s in any other case wholesome, and has retained all of her recollections earlier than the accident, however she can not create new ones as soon as she goes to sleep.

However Adam Sandler and her fall in love, repeatedly. They’ve 50 “first dates”, and more and more they fall in love extra rapidly, although not all the time.

Sooner or later, she realizes he and her household have been hiding the reality about her amnesia from her. She admits herself to a residential facility in order that she will be able to paint in peace, and minimize herself off from the Groundhog Day like cycle of loving, forgetting, and deception.

However at some point, Adam Sandler decides to go discover her on the facility. And there he finds her, surrounded by work, drawings and sculpture of an individual she is aware of and can’t appear to see. He’s in each image she’s created virtually like a ghost haunting her.

The story means that love has its personal type of reminiscence. Love is saved some place else than the place we retailer the traditional move of life. And at the very least on this parable’s case, it’s not saved within the place the place her quick time period recollections are created and saved.

And on the finish, we study that Sandler, her father and brother have found out a method to assist her bear in mind that doesn’t require eradicating her amnesia. Their methodology makes use of footage, dwelling movies, and a every day observe of solitude earlier than the beginning of every single day. The observe helps her get re-acclimated to her life. This strategy helps her to recollect love. Not fall in love — she is already in love. Reasonably, the observe helps her bear in mind she is already in love.

So, amnesia is there, however the love for the undertaking is there too. All of us have felt the love in our analysis. I believe the unhappiness with AI brokers isn’t that the machine is doing the work. Reasonably, it’s that we can not bear in mind the work, and we aren’t holding as simply our fingers to the love we really feel for it.

So a profitable harness will maintain the work inside attain in order that the discoveries are all the time there, and in order that the sensation of understanding could be deepened, regardless of not writing the code.

One factor I’ve working towards me is “aphanasia”. Aphanasia means once I shut my eyes, I’m unable to visualise something. If you happen to ask me to shut my eyes and visualize my childhood home from Mississippi, I can not although I’ll suppose freely about my childhood home in Mississippi. I simply don’t appear to see it in the identical method others see issues of their minds once they shut their eyes. My minds sight makes use of creativeness, narrative and feeling to think about my childhood home.

I believe perhaps that is related to why my discussions of econometrics is full of tales. I’m continually telling the story of Orley Ashenfelter, Dave Card, Bob LaLonde, Josh Angrist, Janet Currie, Jim Heckman, Guido Imbens, Don Rubin, Alberto Abadie, Esther Duflo. Continually. Why do I do that? Actually, I believe it’s the solely method I can perceive econometrics. I’ve to know the biography of econometrics to know the arithmetic of econometrics. I would like the story. Tales assist me bear in mind.

So the harness will even document the work by way of outlines and narratives. I’ll have a tab within the dashboard the place I’m constructing out interpretations of outcomes, and linking the fabric over time, with an eye fixed in direction of the story that I might be crafting.

One other pillar is magnificence. Magnificence is vital for analysis and science as a result of magnificence makes you look. It makes you cease and stare. It makes you keep on a web page method longer than anticipated. And since I’m needing to remain in a continuing knowledgeable state, I’ll make sure that the dashboard is gorgeous, the tables are superbly rendered, the figures might be lovely, and a fantastic deck primarily based on my “rhetoric of decks” philosophy plotting my current progress might be created routinely every single day. If I can cease and look, which magnificence makes me do, then it can assist me bear in mind what has been finished, and what has not been finished, as a continuing feeling of verification.

One of many predominant methods I’ll make zero errors the constraint is thru a formalized guidelines. The guidelines is a tab on the dashboard and it guides the initiatives steps.

Checklists have been discovered to scale back hurt to sufferers in surgical procedures, and crashes in aviation, when adopted constantly. Additionally they assist groups work collectively. They usually make replicable work that may be transferred later into the information appendix of the manuscript and the information repository the journal requests upon acceptance. These are important for a contemporary empirical workflow for a lot of causes, however one is as a result of Claude Code doesn’t suppose, it doesn’t cause, and it doesn’t bear in mind. It tends to float, and when one makes use of Claude code as a considering associate in fixed dialogue all through the method, they drift can occur even moreso since what you ask Claude to do throughout these conversations and what it does do will typically diverge.

So I’ve to create scaffolding in a workable workflow that ensures that that drift is manageable and noticeable and in the end resolvable. And that’s with a guidelines. I’ve a 9-10 level guidelines for the manufacturing of labor with Claude Code that’s within the dashboard. Every level accommodates figures, tables, html and descriptions of findings and checks. We don’t go to the subsequent stage in that guidelines till we have now completed fully that first stage. And nothing in every stage exists until it has been audited by /referee2 and confirmed that it’s replicable.

I’m near having one thing working. I’ve been engaged on it, like I mentioned, for 2 months, perhaps three. It was born from noticing time and again that one thing was getting away from me and I couldn’t fairly put my finger on what, and extra importantly, why. I’ll push it to my repo quickly in addition to present video stroll throughs for these trying to make one thing too. I don’t count on it to work for everybody, a lot as I hope that it conjures up to determine yours.