I hadn’t heard of Dan Guido till a number of months in the past, after I got here throughout the video of a chat he gave at [un]prompted, an AI safety practitioners’ convention. Dan is the CEO and cofounder of Path of Bits, a software program safety analysis and improvement agency that works with corporations in tech, protection, and finance. However Dan wasn’t speaking about safety. He was speaking about what it takes to make an organization AI native, which is near the middle of the bullseye for many people proper now.
We’ve been attempting to determine how to do this at O’Reilly, however till I got here throughout Dan’s discuss, we didn’t have a structured course of. We’ve been constructing alongside the traces he laid out ever since. So for this episode of Dwell with Tim I requested Dan to reprise the discuss earlier than we acquired to the dialog. He was purported to take twenty minutes, like his unique convention discuss, however he took thirty-five, and I needed to reduce him off barely earlier than the top to make room for questions. That was a troublesome selection, since every thing he needed to say was golden.
Dan opened by reminding us of the present state of play in enterprise AI adoption. In February, Fortune reported on a Nationwide Bureau of Financial Analysis research wherein practically 90% of some 6,000 executives stated AI had produced no measurable change in employment or productiveness at their corporations over three years. Individuals began calling it the brand new Solow paradox, after Robert Solow’s 1987 line that “you’ll be able to see the pc age in every single place besides within the productiveness statistics.”
Dan’s perception is that this isn’t proof that AI doesn’t work. It’s proof that almost all corporations are deploying AI improper. They hand out ChatGPT and Claude licenses, after which management waits for the magic to occur. It doesn’t.
Dan began out by describing three ranges of AI adoption.
- AI assisted is the place everybody begins: “You give folks entry to ChatGPT, it drafts emails, it summarizes paperwork. It’s only a productiveness device, and your group doesn’t change. Your workflows are the very same as they had been earlier than. You simply have a little bit buddy that helps you with a few duties.”
- AI augmented is the place you begin redesigning workflows, in order that AI does the primary move on a code evaluation and a human does the second.
- AI native is structural: “That’s the place you’ve redesigned the corporate and its workflows from the bottom up, assuming the AI goes to be there and that it’s a core participant. That’s not likely a device. That’s extra desirous about AI as teammates.”
In his framing, the primary of the three is a device and the final is an working system. For Path of Bits, he stated that “working system” has a selected goal:
“I need our safety experience to compound as code. Each engagement we do, all the talents, the workflows, every thing that we construct makes the following engagement quicker and higher.”
Worker resistance is the primary downside
Dan confessed how laborious it was to get began on the ladder from AI Assisted to AI Native:
“After I introduced final 12 months that we had been all in on AI, that we had been going to be utilizing it throughout all of our workflows and redesigning the way in which the corporate operates, I’d say solely about 5% of the corporate was with me. 95% was resistant.” About 20% was actively resisting. The opposite 75% had been resisting extra passively. “They’ll associate with it in public, however in course of they’ll sabotage it. They’ll hope that in the event that they hold their head low, it will move over them, and that three months from now administration’s focus will change and it received’t be an issue anymore, and we will get again to doing what we had been doing. That’s the place the vast majority of folks land when these initiatives occur.”
Slightly than argue together with his workers, Dan studied the literature on why folks reject new know-how and determined he wanted to deal with 4 biases towards AI: self-enhancing bias, identification menace, opacity, and intolerance for imperfection.
Self-enhancing bias is the behavior of crediting your wins to your individual judgment and your losses to circumstance, which is a specific downside for senior people who find themselves strongly hooked up to the years of expertise and instinct that acquired them to their current place. Opacity isn’t having the ability to see how a choice acquired made. Dan’s remark is that you simply don’t perceive your physician’s reasoning both, however someway you belief the physician however get suspicious of the machine. Dan didn’t point out this work particularly, however intolerance for imperfection appears to consult with Dietvorst, Simmons, and Massey’s work on algorithm aversion, which discovered that individuals abandon an algorithm after watching it err as soon as, even when it outperforms the human different. Their follow-up paper discovered that giving folks even a slight means to change the algorithm’s output is sufficient to overcome the aversion.
Dan spent essentially the most time on identification menace. He described a research wherein the identical kitchen equipment was marketed in two methods: “On one hand, it does the cooking for you. However, it helps you prepare dinner higher. It’s the identical gadget. The individuals who recognized as cooks rejected the primary model and accepted the second.”
Most data work, Dan argued, and safety auditing specifically, is what he referred to as symbolic fairly than instrumental. That’s, it carries that means about who you’re. “So I’ve to border AI as one thing that makes you a extra harmful auditor,” he stated. “Not that it does the audit for you.”
In his work at Path of Bits, he intentionally constructed a countermeasure for every bias.
- Self-enhancing bias is addressed by “an AI maturity matrix” with seen ranges, as a result of you’ll be able to’t declare you’re already adequate when there’s a broadcast ladder that identifies a distinct set of abilities as crucial.
- Identification menace will get abilities repositories, the place an engineer who writes a tough plugin will get credit score for encoding their experience. Hackathons additionally change the dynamic from resistance to exploration. I’m placing phrases in Dan’s mouth right here, however I feel he’d agree that when skilled builders are referred to as on as mentors in a hackathon, that additionally reduces their expertise of AI as an identification menace.
- Intolerance for imperfection will get a curated market, sandboxing, and hardened defaults, so everybody’s first expertise of AI isn’t a catastrophe.
- Opacity will get a written AI handbook that clarifies the utilization coverage and the danger mannequin fairly than simply saying “belief us.”
Right here’s Dan’s slide on “the treatments that truly labored”:
Returning to one in all my interest horses, this can be a type of mechanism design. In my latest piece on the lacking mechanisms of the agentic financial system, I argued that we have to begin with desired outcomes and ask ourselves what mechanisms will assist to provide them. Dan’s strategy appears to be actually good at this. Most enterprises are treating AI adoption as a procurement downside or a communications downside. Dan handled it as a query of what incentives, defaults, and standing ladders produce the habits you need, given how folks really reply.
The final treatment on Dan’s checklist is that the CEO has to steer by instance. He famous, “I used to be the primary particular person via the door. My voice because the CEO issues much more than folks assume. The passive 50% of the corporate that isn’t certain if this initiative goes to achieve success, they’re watching to see what management really does, not what it says.”
A ladder, not a mandate
Path of Bits already tracked about 50 engineering abilities for efficiency evaluation, issues like Python, git, Rust, and numerous safety auditing capabilities. Dan pulled AI abilities out into their very own matrix, with 4 ranges, from not engaged via succesful and adoptive to transformative. Every of those ranges is detailed individually and extra particularly for assurance, engineering, gross sales, and challenge administration.
He famous that “The best stage of the maturity matrix isn’t someone who makes use of AI essentially the most. It’s someone who invents new methods to work and builds instruments with AI. So the identification of the skilled shifts from ‘I don’t want AI’ to ‘I’m the one who makes AI helpful for the corporate.’” This was his first essential design selection.
The second is what stage zero means. He stated “In the event you’re at stage zero, in case you’re not engaged, which means you’re preventing again towards the corporate. In the event you dismiss AI as hype, in case you refuse to make use of AI for safety work, this can be a disagreement on rules, not on abilities. For individuals who had been caught within the not engaged class, we had laborious conversations, and there have been individuals who left the corporate.” Ranges one via three are a ability subject, and the treatment is time with the instruments.
Whereas the slide describing the potential matrix is proven within the previous video clip, right here’s the place you will discover the total deck so you’ll be able to research it in additional element.
Driving adoption and abilities with hackathons
Among the best methods Path of Bits developed to maneuver folks up the ladder was to carry a hackathon each two months. Dan runs them with clear targets fairly than as a free-for-all. The main focus space and studying targets are outlined prematurely and introduced per week forward, with separate directions for engineers and non-engineers. Individuals work in pairs so every thing will get reviewed. There’s a demo session on the finish, after which follow-through. (It’s an essential a part of Dan’s massive thought, that it’s a must to construct a system by which, in his phrases, organizational data and functionality compounds.) He famous that “Within the days afterward we hold one or two folks round, they usually accumulate all of the reusable artifacts, construction them, and put them into the locations they should be.”
I requested what folks outdoors of product and engineering really work on, for the reason that reply for an accountant at a hackathon was not apparent. Dan’s response is that the hackathon isn’t measured in artifacts shipped however in the place folks sit on the potential ladder the next week. Primarily, he’s operating a coaching program that occurs to provide helpful output, fairly than a manufacturing dash that occurs to show folks one thing.
The primary hackathon, he advised me, was the equal of a seashore cleanup: “It’s like these corporations that ship everyone to the seashore with an enormous stick and say, let’s go decide up a bunch of trash and put it away, and you then get the massive staff picture after with all of the contractor luggage of rubbish. That’s what we did with our public supply code repositories.”
He picked it as a result of open supply upkeep is the a part of the job that seems like a grind. No new options, simply closing points and off dependencies on public code the place nothing was in danger. “As an open supply maintainer, you simply get crushed down by the general public. This doesn’t work, I can’t use it, this factor sucks. Dozens of points mentioning flaws you already knew about. It feels burdensome. We needed folks to see that adopting AI would relieve burden.”
The second hackathon was about transport impactful product updates, but it surely was additionally designed to maneuver everybody up the potential ladder by giving up management. Engineers needed to run Claude Code in bypass permissions mode, absolutely autonomous, on public repositories, inside sandboxes the corporate had ready prematurely. The one they’re operating now could be about persistent background brokers that may be handed a job throughout an audit and are available again with a proof of idea exploit or a draft discovering.
Right here’s a have a look at Dan’s slack message saying the hackathon:
The whole lot the hackathons produce will get harvested into artifacts.
Path of Bits runs three abilities repositories: an inner one for firm workflows, a public one that anybody can use, and a curated one that vets third-party abilities earlier than they’re allowed in.
Publishing abilities to the general public repository is not only a advertising and marketing train. “It retains us trustworthy, and it forces us to put in writing issues that different folks can use, not simply folks outdoors the corporate however inside too,” Dan stated. “It actually helps us take into consideration the tribal data that’s baked into the device.”
The curated repository exists as a result of Path of Bits is aware of how unhealthy the provision chain is. They’ve printed analysis on easy methods to write malicious abilities, and so Dan isn’t going to inform 130 workers to start out downloading code from strangers and operating it on their laptops. “If you need adoption, you want a secure provide chain.”
Turning scar tissue into infrastructure
Maybe much more essential than the talents repository is, as Dan put it, “turning scar tissue into infrastructure.”
“Each single time Claude Code didn’t do one thing we needed, we’d bake it right into a set of world, copy-pasteable defaults. Recognized good settings, beneficial patterns. I name it scar tissue. If I rent someone new tomorrow, I don’t need them to need to undergo your entire discovery technique of the final 12 months of Path of Bits to determine easy methods to use the device.”
The configuration repository, claude-code-config, is the place the amassed classes stay.
Dan constructed the primary model himself after which opened it to drag requests from the entire firm, assigning somebody after every hackathon to go accumulate what folks hadn’t contributed on their very own. “It’s simpler to place out one thing that’s unpolished than it’s to get it good on the primary attempt.”
Briefly, an enormous a part of the Path of Bits “enterprise AI working system” strategy is a set of standardized instruments and hardened defaults. Standardization isn’t a straitjacket. It’s a basis.
On sandboxing, Path of Bits intentionally didn’t decide a single most well-liked resolution. There’s a devcontainer for builders, dropkit for disposable DigitalOcean droplets, COOP for remoted VMs, and the sandboxing now constructed into Claude Code for informal customers. “The purpose isn’t that everyone makes use of the identical sandbox,” Dan stated. “The purpose is that everybody has a secure sandbox to make use of, and that it’s straightforward for them to do it.”
One other of the hardened defaults is procedural. Path of Bits enforces a seven day cooldown on each bundle their builders set up:
“There are dozens of safety corporations scanning the web looking for a brand new cool weblog publish they’ll write about malicious code hiding on PyPI or npm, they usually often determine on the market’s a provide chain subject inside hours. So we simply delay all of the packages that Path of Bits makes use of. Usually the malicious stuff will get picked up earlier than we ever get an opportunity to run it.”
That’s free-riding on a aggressive marketplace for safety analysis, and given the pace of as we speak’s market, it’s a chic resolution. There’s an entire class of defenses like this ready to be discovered, the place the mechanism isn’t a technical system however a well-chosen delay.
Information, and DJ Patil’s “Tidy Home”
The issue we run into most frequently as we construct AI workflows at O’Reilly isn’t the mannequin or the tooling. It’s information. Who has entry to which system, which system does that information stay in, and who do I ask? In a 500 particular person firm that’s annoying. I’m wondering what it’s like at an organization with 50,000 workers.
I advised Dan about DJ Patil’s Tidy Home framing. He agreed that information entry for AI is an enormous downside. His reply begins with permissions:
“The permissions debt is invisible till an agent hits it. Making information agent legible is a pressured permission audit. It’s a must to really undergo and work out who can entry what…. It additionally raises the stakes for permissions errors. In the event you overshare info, now an agent inside your organization goes to search out it immediately. There are quite a lot of these technical debt form of issues the place, with brokers, all of it’s turning into due on the identical time.”
Each shortcut a company took with its information over the previous twenty years is being referred to as without delay, and the businesses that may run the audit, make quick choices about boundaries, after which really share their information are those that may get a pressure multiplier.
Dan is towards letting a thousand flowers bloom, as a result of uncoordinated groups create overlap fairly than compounding. He’d fairly have one centralized basis, with innovation taking place on prime of that. He prompt a helpful metric for making that work throughout staff boundaries is what fraction of your staff’s information did you make reusable for everybody else, and the way a lot of it’s being utilized by groups outdoors your individual.
What post-AI jobs appear like
Earlier than the primary hackathon, Path of Bits ran hands-on periods to show its operations and go-to-market workers the fundamentals of git and the command line. Not mastery, simply sufficient to be a shopper of the factor. Right here we’re fifty years into my profession and the Unix command line nonetheless issues. Dan’s non-technical workers principally work inside Claude Cowork or Codex Desktop now, however he thinks the command line expertise was value it as a result of they know what’s taking place beneath the hood.
What occurs to a job when the device can do quite a lot of what people used to do? Dan gave the instance of his personal technical editors. His editors used the hackathons to construct the instruments that acquired them out of line enhancing, together with one which turns a public presentation right into a weblog publish within the firm’s voice. What the writers do now could be seek the advice of on easy methods to body a narrative so it’s efficient with a specific viewers.
I agree. Human jobs aren’t going away any time quickly. This will get heard as optimism when it’s actually simply remark. AI goes to interchange quite a lot of what we used to do, however additionally it is going handy us a considerable amount of new work, and far of that work hasn’t been understood but. High quality assurance for agent programs is likely one of the new jobs. So is abilities product administration, which is a task that didn’t exist eighteen months in the past and now has a headcount at a 130 particular person safety agency.
I requested a query in direction of the top about how we’re going to know which abilities and brokers are any good. What Dan has thus far is telemetry pulled from builders’ dot recordsdata via the corporate’s gadget administration system, which tells him what will get used and what breaks, plus one AI programs engineer whose job is product administration for the talents repository, reviewing incoming pull requests and deprecating overlapping abilities.
What Dan thinks comes subsequent is analysis. He says: “When you make investments loads into these agent programs, you want proof that they do the job. The way in which you do that’s you give everyone a efficiency evaluation. You give them an analysis information set, a benchmark.”
Path of Bits is now constructing benchmarks for its core abilities. How properly can we discover bugs on this language? How properly can we write a press release of labor? Establishing these datasets is actual work, with constructive and unfavorable circumstances, and comparisons towards the algorithmic instruments that exist already.
Put the reps in
I requested Dan for the highest 5 errors he made. He stated there was just one. “It’s essential to allocate an applicable quantity of FAFO time. (That’s F Round and Discover Out.) A product comes out on Friday. There’s no documentation for it. There’s no coaching steerage for it. There’s no course on it. You possibly can’t wait till someone systematizes the data. You simply must do it.”
Then he gave an analogy to going to the gymnasium.
The recipe for achievement
Dan has a replicable recipe, which he summarized as follows:
- Standardize on one agent workflow which you can help.
- Write an AI handbook in order that threat choices aren’t advert hoc, and that everybody is enjoying the identical recreation.
- Create a functionality ladder that makes clear that enchancment is predicted.
- Run brief adoption sprints that pressure hands-on utilization.
- Seize every thing as reusable artifacts: abilities + configs + a curated provide chain.
- Make autonomous brokers secure with sandboxing + guardrails + hardened defaults.
The Path of Bits abilities repository is public. So is the curated market, the configuration repository, the devcontainer, dropkit, and COOP (Continuity of Operations planning). He wrote up the entire playbook on The Path of Bits Weblog and gave a model of it to tl;dr sec. He thinks publishing makes the work higher as a result of it forces the tribal data out into the open the place it may be checked.
Which brings me again to the Solow paradox, which appeared to vanish by the late 90s, when US mixture productiveness did lastly go up. That didn’t occur as a result of computer systems acquired quicker. It disappeared as a result of corporations found out easy methods to reorganize themselves round what computer systems might do, and finally these organizational recipes unfold broadly sufficient to point out up in mixture statistics. The identical has to occur as we speak. The present AI discourse is obsessive about mannequin functionality and largely bored with diffusion. The issue isn’t that the fashions are oversold. It’s that nearly no person has carried out the required organizational work, and the few who’ve are principally maintaining it to themselves.
If you wish to transcend the spotlight movies proven above, watch Dan’s complete discuss right here. His slide deck is right here. And make sure you try the Path of Bits Github repository.
