Agile InsightsBriefing
A black and white factory conveyor carrying machined metal parts, with one pair of hands checking a part, ringed in yellow and labelled Human in the Loop.

Episode 16 · Primer and discussion · About 13 minutes in two parts

When making something gets cheap, checking it becomes the real work.

A I is changing what work an organisation can afford to take on, not only how fast it does its usual work. What an investment firm’s scorpion toxin experiment shows about orthogonal work, the cost of checking, and where the outputs go.

#oversight#leadership

Part 1: Primer · About 6 minutes

Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.

Part 2: Should you encourage orthogonal work? · About 7 minutes

Download the MP3 · Linda and Leonard, AI voices, read a script written from the sources below.

In this episode

01Opportunity, not only efficiencyThe efficiency case is the same work done faster. The opportunity case is work you could never justify before. Nathaniel Whittemore of The A I Daily Brief calls one part of it orthogonal work.
02An investment firm designs proteinsTo test whether A I drug discovery was real, Citrini Research worked with Anthropic. In 64 hours a model generated more than 318,000 designs; two independent labs found 17 that latch onto a scorpion toxin.
03The cost moved, it did not vanishCitrini reports about US$50,000 on computing and just under US$150,000 on lab work over about two months. None of the 17 has yet been tested to see whether it stops the venom.
04New work leaks in new waysAs reported by Ars Technica, a security firm found more than 6,800 shared Meta Muse conversations listed in public internet records, including a loan application. Meta suspended sharing.
05What to doAdd a line for new work to your A I use case list, give each an owner, budget cap, pass marks set in advance and a stop, budget for checking from the start, and treat sharing as publishing.
Sources and links (5)
Claude designs binders against scorpion toxins · Read on Anthropic
Anthropic’s research post, 20 October 2026: the design campaign, the lab results, the pass criteria and what was not shown.
Citrini Therapeutics · Read on Citrini Research
Citrini Research’s own account: why an investment firm ran the experiment, the budget, the lab bill and the timeline.
The AI Daily Brief, 21 October 2026 · Listen on The AI Daily Brief
Nathaniel Whittemore’s episode that names orthogonal work and sets it against the efficiency case.
Meta Muse exposes users’ secrets when they share conversations · Read on Ars Technica
Ars Technica’s report, about 20 October 2026, of the security firm Claroc’s findings and Meta’s response.
Meta suspends Muse chat sharing after 6,800 links leak via DNS · Read on AI Weekly
AI Weekly’s summary of the Muse findings, including how the share addresses were built.

Written from public sources and checked on 23 October 2026. The experiment’s figures come from Anthropic’s and Citrini’s own accounts, which agree. The Muse findings were read through AI Weekly and The AI Daily Brief, not from Ars Technica or Claroc directly. Orthogonal work is commentary, not an established category. Check the original before quoting.

Our fact check · 11 claims: 4 confirmed, 5 reported, 1 disputed

Before we record, every claim the hosts make is checked against the sources (this table was last updated 23 October 2026). Confirmed means a primary source or two independent reports; Reported means one party says so; Partly right, Disputed, Unverified and Not yet known mean what they say, and the hosts say so on air. This is the table as we left it.

  1. ConfirmedCitrini is an investment firm and stays one; motive was an investment questionCitrini; Anthropic
  2. ConfirmedJoint effort with Anthropic’s life-science team and an internal research modelAnthropic; Citrini; AIDB’s “Citrini ran 300 sub-agents” compresses this
  3. Confirmed64 hours; 318,484 designs; 180 + 31 = 211 tested; 17 binders, each against one toxinAnthropic; Citrini
  4. ConfirmedNone tested for neutralisation; no antivenom claimedAnthropic; Citrini
  5. ReportedBudget US$200,000: compute US$49,997, lab just under US$150,000Citrini
  6. Reported300+ sub-agentsCitrini
  7. Reported“64 hours, 39 days, four to seven years”AIDB relaying Will DePue on Citrini’s charts; Charts not read
  8. Reported1.2m+ stings a year, 3,250+ deaths, mostly childrenAnthropic citing published estimates
  9. DisputedCritics: toxin binding shown 15+ years agoAIDB, unattributed; Anthropic: no de novo scorpion-toxin binders described before
  10. ReportedClaroc found 6,800+ shared Muse conversations; share hostnames carried the chat title and were listed in passive DNS; Meta suspended sharingArs Technica via AI Weekly and AIDB; AIDB’s “guessable” is partly right (listed, not guessed)
  11. “Orthogonal work” as a categoryNLW; Commentary

Want to know more?

Sign up for our podcast at agile-insights.com.au.

Read the transcript · Part 1: Primer

Linda Here is the idea for today. A I is changing what work an organisation can afford to take on, not only how fast it does its usual work. And when making something gets cheap, checking it becomes the real work. By the end of this part you will know the difference between the efficiency case and the opportunity case. You will know what people have started calling orthogonal work. And you will know why the cost of checking is where the queue now forms.

Leonard Most of what I hear about A I at work is about speed. Drafting emails faster, summarising meetings, writing code in half the time. That is real, and it is what most business cases count, because hours saved are easy to put in a spreadsheet. So when you say A I changes what work we take on, I am not sure what that adds. Is that not just more of the same work, done quicker?

Linda That is step one, and it comes down to two kinds of case. The efficiency case is the same work, done faster or cheaper. The report that took a day now takes an hour. The opportunity case is work you would never have done at all, because it cost too much to justify. Nathaniel Whittemore, who hosts The A I Daily Brief, has long argued that the opportunity side is bigger and harder to see. On the twenty first of October he gave one part of it a name, orthogonal work. Orthogonal means at right angles. It is work off to the side of your core business, done only because it makes the core business better.

Leonard So what does that look like in a real firm?

Linda Step two is the story behind the name. Citrini Research is an investment research firm. Its analysts thought life sciences stocks looked cheap, and they kept hearing that A I would transform drug discovery. But the experts they spoke to disagreed wildly. Some said A I had never produced an important drug, and others said it would cure all disease within five years. So the firm decided to find out for itself. Five people, three of them with no science background at all, worked with the life science team at Anthropic. They set a budget of two hundred thousand U S dollars. And they picked a real problem, designing proteins that latch onto scorpion toxins.

Leonard As you do.

Linda Bear with me.

Leonard Scorpions are a long way from picking stocks for a living. Why would an investment firm choose that problem, and what did the team actually do with it?

Linda Because it is a real need that almost nobody pays to solve. Anthropic says scorpions sting more than one point two million people a year, and more than three thousand two hundred die, mostly children far from good medical care. New treatments are unlikely to earn back what they cost. So an A I model, working mostly on its own for sixty four hours, generated more than three hundred and eighteen thousand protein designs. Citrini says the model ran more than three hundred helper agents at once. The team sent two hundred and eleven of the designs to two independent laboratories. Seventeen of them passed strict tests for latching onto a toxin.

Leonard Seventeen working designs from less than three days of computer time. From the outside that sounds like an investment firm made a scorpion antivenom on the side. I can already picture the headline...

Linda And that headline is the common misunderstanding, which brings us to step three. Latching onto a toxin is not the same as stopping it. None of the seventeen has yet been tested to see whether it neutralises the venom. Anthropic says plainly that it does not claim an antivenom. Now look at where the money went. Citrini reports about fifty thousand dollars on computing, spent in sixty four hours. The laboratory bill was just under one hundred and fifty thousand, and the results took about two months. A real medicine would take years. Citrini wrote that its digital invention was now set to create work for human scientists. Generation got cheap. Verification did not.

Leonard Less exciting.

Linda Much less exciting. But much more useful.

Leonard So the cheap part was the guessing, and the expensive part was finding out whether the guesses were any good. I can see the same thing in an office. A team that can draft ten proposals a day still needs someone senior to read every one of them before it goes to a client. If nobody plans for that reading time, the drafts just pile up on one desk. The bottleneck has moved, but it has not gone away.

Linda We have all seen that desk.

Leonard We have.

Linda Exactly, and step four is what this means for governing A I. First, a business case that counts only hours saved will never find orthogonal work, so the list of A I uses an organisation keeps needs a line for new work as well as faster work. Second, the Citrini experiment was well run, and that is the real lesson. It had a fixed budget, pass marks set before the results came in, independent labs paid to check, and a plain statement of what had not been shown. Third, budget for checking from the start. That means review time, testing and evidence, because that is where A I generated work now waits.

Leonard So the takeaway from part one is simple enough. Count the new work as well as the faster work, and pay for the checking before you celebrate the output. In part two, we argue about whether an organisation should encourage this kind of work at all, and what can go wrong when it does.

Linda See you in part two.

Read the transcript · Part 2: Should you encourage orthogonal work?

Linda Here is the idea for this part. A I lets an organisation take on work it could never justify before, and the hard part is no longer making that work. The hard part is checking it and controlling where it goes. By the end of this discussion you will have two views on whether to encourage this kind of work, where the evidence settles, and four things an organisation can do about it this quarter.

Leonard A quick recap first, so this part works on its own. Orthogonal work is a name from Nathaniel Whittemore of The A I Daily Brief. It means work off to the side of your core business, done because it makes the core business better. The example is Citrini Research, an investment firm. To test whether A I drug discovery was real, it worked with Anthropic to design proteins against scorpion toxins. The design took sixty four hours. Two labs then confirmed seventeen designs that latch onto a toxin, but none has yet been shown to stop the venom. Citrini is still an investment firm.

Linda So the question is this. Should an organisation encourage orthogonal work, and how does it govern work that sits outside its usual lane? My view is yes, encourage it. Most A I business cases count hours saved, and the cost of a question you could never answer does not show up there. Citrini could not get a straight answer from experts, so it spent about two hundred thousand U S dollars to get evidence of its own. For an investment firm that is a research cost, not a moonshot. If the only uses an organisation records are the faster versions of its existing work, it will never see this kind of opportunity. Give it a small budget, a named sponsor and a deadline...

Leonard I am more cautious, and for two reasons. First, this is one case, with a lot of help most firms will not get. Citrini ran the work from Anthropic's offices, alongside Anthropic's own life science team. Anthropic itself says the computing used was much more than most researchers have access to. Second, work outside your lane is work outside your expertise. A team of investment analysts cannot judge whether a protein design is any good. A finance team running its own legal analysis cannot judge the law. So the danger is not that the A I fails loudly. It is that it produces something plausible, and nobody in the room can tell.

Linda That is fair, and I think the case answers it rather than undermines it. Citrini knew it could not judge the output, so it paid two independent labs to do the judging. It set strict pass marks before the results came in. A design only counted if it bound in every repeat, at both labs, and showed no sign of sticking to an unrelated protein. Thirty five design and toxin pairs showed some sign of binding, and only seventeen passed. The attempt to hit all five toxins with one design failed, and they said so. So I would put it this way. Encourage the work, but make the independent check the price of entry.

Leonard The price of entry. I like that.

Linda I thought you might.

Leonard Agreed, and my second worry is about where the work goes once it is made. In the same week, a security firm called Claroc reported that it had found more than six thousand eight hundred shared conversations from Meta's new assistant, Muse. As reported by Ars Technica, each share address carried the conversation's own title, and those addresses were listed in public internet records. Among what the researchers found were a small business owner's loan application and a marketing company's internal revenue review. Meta suspended the sharing feature, and Claroc said it had no evidence anyone had misused the links. But think about who those users were. They were doing real work in a consumer tool, and then they shared it.

Linda Real work, in a consumer app.

Leonard Exactly.

Linda And that is the governing point. A share link is a publication channel, even when it feels like passing a note to a colleague. When people use A I to do new kinds of work, the outputs travel further and faster than the old drafts did. So an organisation's content rules need three plain questions for every A I tool its people use. Is sharing switched on? Are the links hard to find, and do they expire? And can anyone see what has been shared, and by whom? If the answer to the last one is no, you do not know what has left the building.

Leonard So where does the evidence settle? The facts of the experiment are well supported, because Anthropic and Citrini both published their own accounts, and the numbers agree. Sixty four hours, two hundred and eleven designs tested, seventeen that passed, and no test yet of whether any of them stops the venom. Citrini's own figures put the computing at about fifty thousand dollars, against just under one hundred and fifty thousand for the labs. The Muse findings are reported, from the security firm through the press. What is not known is just as important. We do not know whether any of these proteins will ever become a treatment. And one well supported firm is not yet a pattern, so nobody can say how common orthogonal work will become.

Linda Then here is what an organisation can do. One, add a line for new work to your A I use case list, alongside the faster work, so opportunity cases are seen and funded on purpose. Two, give every one of them an owner, a budget cap, pass marks set in advance and a point at which it stops. Three, budget for checking from the start. If the work sits outside your expertise, buy the check from someone who has it, the way Citrini paid the labs. Four, treat sharing as publishing. Know which tools allow it, set links to expire, and keep a record of what has been shared.

Leonard The part that stays with me is the shape of the costs. Sixty four hours to design, and then weeks in the lab, and years before anything could reach a patient. A I made the first step almost free. Every step after it still needs people who know what they are looking at. That is not a reason to hold back. It is a reason to plan the checking before you start.

Linda If you'd like to know more, sign up for our podcast at agile-insights.com.au.

Leonard Thanks for listening.

More on #oversight

Making You Think 6 · Primer and discussionThe Unaccountability Machine, by Dan DaviesEpisode 1 · DiscussionStaying in charge of decisions in the age of AIEpisode 37 · PrimerWhat is an AI skill (and why you should care)Episode 36 · Primer and discussionChecking agents in production, even when you can’t read their workEpisode 35 · DiscussionWho watches the agents? Detect, stop, and let a person approve the rulesReference 13 · PrimerThe Commonwealth AI rulebook: the Policy, the technical standard and the APS AI Plan

Every episode tagged #oversight

All of the briefings (38) · Follow Briefing in Apple Podcasts or any podcast app

Everything from Agile Insights: briefings, Making You Think and the reference series