Stefano Bury 🇺🇸

Prev. @McKinsey @INSEAD || Follows / RTs ≠ Endorsement

[Stefano Bury 🇺🇸 @stefanobury]

Claude's new font is infinitely better

[Stefano Bury 🇺🇸 @stefanobury]

Genuinely the only feature I liked about Instinct. For now, my workflow is getting Grok Bot to send emails to my wife

[Josh Mohrer @joshmohrer]

Do any of the majors (Grok, Muse, Dots) have two-player mode to share with a spouse yet?

[Stefano Bury 🇺🇸 @stefanobury]

The most underrated line here is that a large share of runs are scheduled, not invoked. In enterprises, that's where the ROI shows up: agents that do the work before anyone asks, not chatbots waiting for a prompt

[Harrison Chase @hwchase17]

https://t.co/1RVHfb9BXX

[Stefano Bury 🇺🇸 @stefanobury]

Partly agree. Better visibility for leaders is real, but having seen this as both a consultant and a client, the tech is only half of it. If the operating model doesn't change too (who decides, how teams are shaped, what gets measured, what new capabilities do you need), you end up with a better dashboard sitting on the same old org

[Ethan Mollick @emollick]

A thing I see happening at large organizations as a result of AI is the expansion of the role and span of control of senior management who can directly monitor the whole firm. It vastly changes company operations. It also is straightforward to implement compared to firm redesign.

[Stefano Bury 🇺🇸 @stefanobury]

After two days of scanning every brick and mortar and online retailer, one of my agents was able to buy a Needoh Squishy Advent Calendar.

If you have a pre-teen, you know this is a huge deal.

Agents are turning moms and dads into parents of the year.

[Stefano Bury 🇺🇸 @stefanobury]

@theohopkinson @FundamentEdge I rushed to the comments looking for this one

[Stefano Bury 🇺🇸 @stefanobury]

Useful framework. For applied AI companies, the honest answer is usually distribution plus accumulated workflow know-how, which is sublinear on its own. The ones that turn every deployment into reusable playbooks and evals might get closer to linear. Superlinear is still mostly a pitch-deck word

[martin_casado @martin_casado]

I think I've gotten to the bottom of the issue with the "network effects" discourse. The problem is the shape of a moat is often conflated with the mechanism. And there is remarkably little consensus on what a network effect is, or a flywheel.

Here's a very simple, very rough vocabulary that I think works with current working definitions. Seriously, learn these few rules so your boardroom conversation about defensibility isn't a waste of time:

Shape of Moat: This is what happens to the advantage as the company scales

Static : Doesn't change
Sublinear : Advantage compounds but at a diminishing rate as the company scales (this is not a long term durable moat against challengers)
Linear: advantage is fixed relative to a competitor as company scales
Superlinear : Advantage compounds at an increasing rate as the company scale (this is the white wale everyone pretends they have, but almost no one does)

Mechanism of Moat: The property, business process that produces the advantage.

Network : more users creates more interaction between users. Hopefully superlinear.
Scale : greater volume produces better unit economics. normally sublinear
Data-driven : new users means more data for new users (or to get new users). normally sublinear
Platform : growth on one side increases value to the other. sublinear to superlinear depending on size
Distribution : scale improves reach or channel economics. normally sublinear
... etc.

Flywheels: Flywheels are also conflated with the above. The most useful working definition of a flywheel is the aggregate advantage when multiple mechanisms are self reinforcing, something in a chain (A->B->C). Imagine multiple forces pushing a flywheel to make it go faster. The hope is this complex interaction produces a superlinear advantage. They are incredibly hard to reason about. And most discussion around them are based on vibes.

[Stefano Bury 🇺🇸 @stefanobury]

Enterprise agents have the opposite problem and the opposite advantage. Nobody needs them to be fun. They need to be boringly reliable at one job, and once they are, they become an email-and-calendar kind of dependency very fast

[signüll @signulll]

i've been talking to a lot of normal ppl outside of tech about personal agents & the convos have been incredibly fascinating, particularly w.r.t. to retention.

most consumer products typically have one of two types of retention, one is desire driven which is just pure dopamine (like insta, tiktok, games). they often don't even need notifications because you know there is a reward for your desire always waiting. the second is dependency driven, like emails, calendars, even credit cards & banking. most ppl have to do this, it's not something they necessarily desire (unless they are a sociopath).

the challenge with personal ai agents right now is that they occupy a weird third space where they aren't emotionally rewarding most of the time to become entertainment like the first category, & they're not reliable / indispensable enough to become pure infra. i.e. they fail, & there is a lot of friction. outside of pure research driven tasks, even one failure is often enough to not come back for the utility value prop.

which is why i think vertical agents that focus on making your life more interesting & entertaining are worth deep exploration still. & even very focused vertically integrated utility agents that work reliably will still be in demand. tons of ppl still think in jobs to be done when reaching for anything & with general purpose agents that's just not as crisp yet on what those critical jobs are.

[Stefano Bury 🇺🇸 @stefanobury]

Florida in the Fall

[Stefano Bury 🇺🇸 @stefanobury]

Seeing this in enterprise conversations. Once a model clears the bar for a workflow, buyers stop asking which is smartest and start asking about cost, control, latency and where the data lives. Procurement teams are about to get very opinionated about models.

[sarah guo @saranormous]

if an increasing set of economically valuable tasks is “intelligence saturated,” how is it not blindingly obvious that open/custom model inference is a market that rivals the size of closed intelligence today

given the option… of course customers want choice, control and perf

[Stefano Bury 🇺🇸 @stefanobury]

This is the part most enterprise pilots skip. They test on a clean dataset, go live, and then nobody owns the loop from flagged output to real failure to eval to fix. Whoever owns that loop is basically the agent's manager

[Harrison Chase @hwchase17]

the useful part of a judge on production traffic is what happens to the outputs it flags

calibrate it on human labels, flag low scorers, confirm the real failures, add them to evals, fix, rerun

prod finds the failures your offline dataset never had

https://t.co/7iUK4e1Fj1

[Stefano Bury 🇺🇸 @stefanobury]

The underrated line here is the last one: 'the right UX, implementation, and change management.' The models are ready for a lot of enterprise work. Most org charts, data access and approval chains aren't. That gap is where the next few years of value gets built

[Aaron Levie @levie]

Pretty cool that we clearly have an emerging form factor for what AI might look like for the foreseeable future as autonomous systems dedicated to specific tasks or domains.

This is a huge leap past the LLM wrapper days because the new expectation is that you can finally farm out more sophisticated work to the agent to tackle for you. And you can reliably trust that work can get accomplished properly (still with a human in the loop) because of far better tool and computer use, and the general intelligence of the models improving.

1) The tool and computer use capability increases in the models mean that you can provide the agent with far more granular access to external data systems and other software to operate. This means, in effect, the agent begins to execute workflows similar to how a person would have in that same function. And where there isn’t existing software, the agent just builds it itself. This is a huge unlock.

2) Next, the actual capability of the models are improving in just about every vertical domain. The models can make far better judgment calls in legal, do deeper bio research, scan for security issues in code, understand supply chain processes, and more. Obviously in many of these spaces then models still have worse intuition than people, but that will only get better.

3) Finally, the harnesses and surrounding infra — sandboxes, file systems, data pipelines, etc. — allows for longer running agents, agents to be able to work on arbitrarily complex or long tasks, process data at enormous scale, and more. Few of these components worked well for AI even a year ago, but we now have many of the core primitives necessary for more autonomous systems.

When combined with the right UX, implementation, and change management for the human oversight and coordination, it’s clear that these autonomous systems will become quite powerful in a wide variety of enterprise workflows.

We’re starting to see them in IT automation, cyber operations, drug discovery, manufacturing, AI research, and more spaces. You can almost imagine one or more of these systems for every role in the economy. Quite a bit of opportunity ahead.

[emmacui @emmacui]

1/ Everyone worries about runaway AI bills. A startup CFO I interviewed this weekend for my Stanford research on AI spend told me the budget is the easy part. Data is the hard part. 🧵

[Stefano Bury 🇺🇸 @stefanobury]

Sinofsky is right, and it's the most under-appreciated part of enterprise AI. Most of a company's real process lives in people's heads as exceptions and judgment calls, not in an SOP. The job isn't prompting the agent, it's capturing what the 20-year veteran knows before you automate around them

[a16z @a16z]

Steven Sinofsky on why it's so hard to diffuse AI into companies: most work runs on know-how that's hard to turn into step-by-step instructions for an agent.

"Algorithmic thinking is really, really, really hard for the vast majority of people who have jobs."

"If you were to go to any person and ask them to create a flow chart for a particular thing that they have to go do, they would probably fail at producing that flow chart."

"Say doing a marketing plan and there's 50 marketing people working on a giant product line. One person probably understands and could document the flow chart."

"So if you put one of these agents in front of people to create these things, their ability to explain to it what to do is really, really limited."

"You're basically just developing the next abstraction layer for how people interact. And developing an abstraction layer has historically, at each level of the abstraction layer, been a highly skilled, very specific individual within an organization."

@stevesi

[Stefano Bury 🇺🇸 @stefanobury]

"revisit every 3-6 months and remove cruft" 💯

This is the real enterprise problem. Most companies still treat an AI deployment like a 2010s ERP rollout: ship it, freeze it, never touch it again

[Harrison Chase @hwchase17]

had a lot of fun talking with @HarryStebbings about all things ai

one line of questioning we went done - vertical vs horizontal agents, and how to build them

im bullish on vertical agents, and i think you can create differentiation (from horizontal agents) in three ways:

1/ context: prompts, skills, tools - all of this is specific to your domain, pretty easy to provide, and gives domain specific context to the agent

2/ harness: harder to tweak than context, but can alter fundemental behavior of the agent

3/ model: hardest to tweak of all, but obviously changes behavior greatly

i think most agent engineers these days are still tweaking 1 and 2 (in that order)

nb: regardless of what layer you tweak at, as the base models change you should revisit your context/harness/model every 3-6 months (and generally remove cruft that is no longer needed)

https://t.co/gmA8oyfxiX

[Stefano Bury 🇺🇸 @stefanobury]

100%

Tech deployment is easy. But once agents start to run for days across 6 systems and need someone to redesign the workflow around them, domain expertise stops being a nice-to-have and becomes the enabler to successfully adopting AI with bottom-line impact

[Aaron Levie @levie]

The big bitter lesson question in the applied AI layer is that as models get better and better, what is the best way to diffuse that intelligence into an enterprise.

Ironically, I’d argue that as models are getting more capable, and can handle longer running tasks and better tool use, the ability to differentiate on a domain by domain basis goes up instead of down.

When models were quite limited in what they could do, ironically all you could be literally was a “wrapper” because you were effectively delivering tokens back to the user with a custom prompt or some limited RAG.

But in a world of agents using tools, tasks that run on for hours or days, workflows that kick off agents, agent swarms that need to be coordinated against a goal, data sources that are domain specific, model routing that requires domain-specific routing, implementation that requires real change management of the workflow instead of “just type into this chat box”, it’s very very possible that specialization of these systems goes up more.

Obviously this doesn’t apply to everything - the closer you are to the labs core use-cases the tougher this play gets, but I’d argue we’re going to see autonomous systems for a wide range of job functions that weren’t practical to build out before. We’ll see this in security, life sciences, healthcare, finance, manufacturing, and many other fields. Either way going to be a fun and wild couple of years.

[Stefano Bury 🇺🇸 @stefanobury]

@levie The swarm math works once verification is cheap. in most enterprises the bottleneck is the one person who has to sign off on what 200 agents did overnight

[Stefano Bury 🇺🇸 @stefanobury]

The most underrated job in enterprise ai right now is basically "agent HR" or basically agent management. Identity, permissions, performance reviews, offboarding. The org chart and operating model is always the hard part

[Satya Nadella @satyanadella]

I recently wrote about why we need to treat models as insider risks, with security controls sitting outside the model itself.

Today, we're putting that principle into practice by expanding Project Perception with new capabilities to help continuously assess agent risk, adapt permissions, and contain threats.
https://t.co/ec3Z2B260G

[Stefano Bury 🇺🇸 @stefanobury]

@gotravelyourway Nay Pyi Taw. Despite being modern and brand new, BPs were all done by hand