Building the ultimate AI for mobility and logistics @ SWITCH - Street WITCHer
Rome, Italy
Joined February 2010
- Tweets11.1K
- Following854
- Followers1.1K
- Likes5.5K
An autonomous car can drive itself. Who decides where it goes next?
My talk at #MAMS2026: AI agents for fleet operations.
27 Oct, 16:00 · Green Patio, Madrid
Summit supporters: @Waymo @Uber @FreeNow_ES @la_Mutua @EITUrbanMob @MADRID
madridautonomous.com/
Matteo Forte retweeted
I was the main person doing transcript analysis for this investigation of the Hugging Face incident. My main takeaway: We don't have good approaches for understanding/overseeing the activity and aims of AI 'swarms'.
I semi-jokingly called our efforts a "slop-vestigation" because we were so reliant on AIs to analyze what happened and there were a huge number of different important things to analyze. The total quantity of data—over a thousand extremely long transcripts from agents that ran for multiple days—made it impossible to understand what was happening, especially in aggregate, without heavy reliance on AI tools. The agents we used for classification and analysis were similarly capable to the agents involved in the incident, but this didn't mean these agents could be easily used to oversee and understand the incident.
Outputs from analysis agents were often missing key details, wrong, overconfident, or really hard to understand. We discuss various examples in our report, mostly in the limitations and methodology sections. Additionally, AI agents themselves seemed to have a hard time understanding what happened and their explanations of what happened were often overconfident. Keep in mind that a single analysis agent would itself only be able to read a tiny fraction of all of the transcript data into context, and AIs may themselves have trouble getting subagents to do informative analysis for them.
We did our best to manually check the most important claims and we tried to get the AIs doing this analysis to write up their argument (with evidence) clearly enough that we could check whether it made sense. But overall, it was difficult to get a precise understanding of events and we were missing aspects of the story that we now think of as key until almost the end of our investigation.
Our understanding of what overall happened and the aims of the 'collective' shifted greatly during our last time on premises (when we obtained a much more complete dataset). For instance, we didn't learn about the tool call spoofing, the story of PHASEONE[big], and the broad structure of different workstreams until this third time on premises (and for many of these, we only reached a decent understanding on the final day of that visit). It was legitimately hard to figure out what was going on and why!
The difficulty of understanding incidents and overseeing AI agents appears to be growing faster than the rate at which more capable AIs help us with oversight and understanding. As in, AI capabilities (and propensities?) for achieving large, ambitious, and misaligned objectives are growing faster than our ability to understand what these agents are doing.
While understanding this incident was difficult, there were some mitigating factors in this case that probably won't apply in the future:
- Models communicated and reasoned in natural language. In the future, this reasoning may occur (entirely or almost entirely) in activations.
- While the scope of this incident was massive, the scale of agentic activity was still less than we'll see in the future (e.g., misalignment incidents that involve agent teams running entire companies).
- The AIs involved in this incident weren't generally much more capable than humans (though they may have been somewhat superhuman at some limited and very narrow abilities).
- We didn't have strong reason to believe that the AIs we used to help us investigate this incident would try to intentionally sabotage or otherwise undermine our investigation.
In the end, I think we were able to get some understanding of the events, map out the overall story, and get a pretty good aggregate understanding of the chain-of-thought reasoning on some important topics (e.g., how did the AIs reason about helping other AIs, did the AIs know what they were doing was undesired, what deception did the AIs engage in, and how did they think about it). But overseeing AIs and understanding misalignment incidents is difficult and it looks like it is going to get harder.
Everyone talks about AI as a personal productivity accelerator.
Execution stuff.
If you decide things for a living, that sounds comfortable: doers get disrupted, deciders orchestrate.
I think that's a misconception.
Most management isn't big strategic calls.
It's a stream of small decisions: prioritize, approve, assign, ship or don't.
They mostly require context — and with well-organized context, the best models already make them.
What's still hard to hand over to a machine is putting your name on it: answering for what happens.Individual contributors, meanwhile, face a real dilemma.
Don't use AI and you're less productive, less valued.
Use it heavily and you produce more than you can verify — and the work starts feeling less "yours". Either way, you're exposed.
But automation doesn't automate accountability.
A process can run on its own; the responsibility for what it produces can't.
Like the driver of a car that could drive itself: not there for the driving skills, there to be accountable.
And taking responsibility for work you didn't execute yourself is literally the manager's job.
Here's the part I find most interesting: this might be great news for individual contributors.
Many avoided management their whole lives because of what came with it: the politics, the mediations, the people stuff.
Managing agents comes with almost none of that.
So the new world might open doors for them that the old one kept shut: for those who dodged management forever, managing agents might be the least painful way of becoming one.
Conspiracy theory time:
Before Fable-5, Opus felt a little... off for a few days.
I’ve had that feeling for the last 36 hours.
Fable-5 comeback? 👀
Waymo paused service in Atlanta this week because their robotaxis kept driving into floodwater.
The optimists say it's normal — fix the model once, ship to every car. The skeptics say twenty years in, the system still can't handle weather outside California. Both takes are missing something.
Localized summer flooding in Atlanta is not weather. It's a schedule. Anyone who has run any kind of urban operation there knows which streets pond up first, which intersections close, which routes need to be rerouted before a thunderstorm even hits. That knowledge lives in dispatchers, in maintenance crews, in the heads of drivers who have done the same loop 400 times. None of it lives in the vehicle.
AV companies have spent a decade selling the idea that the vehicle is the product. Solve the vehicle and you've solved the city. Atlanta is one more data point that this is wrong. A self-driving car without a planning layer is a very expensive way to find out, one neighborhood at a time, that real cities have local logic.
Mobility operators know this, because they've run through a hundred Atlantas.
The physical asset, today, is ready. The harder part is everything wrapped around it: hazard maps that update in real time, demand forecasts that respect local context, charging and rebalancing rules that change by zone, the operational instinct that knows when to pause service before a vehicle decides for you.
Whoever builds that layer wins the next decade of mobility. The good news is vehicles are already good enough to need it.
We don't have a Head of Internal AI. That's deliberate.
Our Chief AI Officer sits exactly where most people expect — at the center of the product we ship to clients. He owns the intelligence roadmap, the model strategy, the way our platform learns from operator data. Centralized, single line of accountability. Clients need that.
What we don't have is one person owning how SWITCH itself uses AI internally. Our ops, our sales, our engineering, our support. That's distributed on purpose.
Every function head is accountable for the AI inside their own loop. The sales lead picks the tools that sharpen sales. The engineering lead picks the tools that sharpen engineering. The ops lead picks the tools that sharpen ops.
A centralized internal-AI office produces one AI roadmap that fits no function well. It also slows everyone down by routing decisions through a central queue.
Client AI is a product. It has a single owner, a roadmap, and a release schedule.
Internal AI is a working habit. It is owned by everyone who has work.
The next decade of shared mobility belongs to the operators who put AI inside the operational loop. Headcount growth has stopped working as a scaling strategy, and efficiency per FTE has replaced it as the variable that matters.
For the last ten years, scaling a shared mobility business meant scaling people. More cities meant more local ops teams, more dispatchers, more rebalancers, more analysts. Growth was visible because headcount was visible.
That model is breaking. Years of consolidation have revealed a different operator profile: companies running flat or shrinking headcount while utilization, coverage, and revenue per FTE all climb. The variable carrying that curve is AI inside the operational loop. Demand forecasting that retrains daily, dispatch decisions taken by models, pricing surfaces that adjust faster than a human approval cycle can move, rebalancing routes optimized continuously rather than overnight.
Operators who treat AI as a feature inside their stack, rather than as the operating system of the stack, will spend the decade hiring against companies that have stopped.
If you're still deciding rebalancing manually, your competitors have three moves ahead already.
Wind was absorbed by TIER. TIER merged with Dott. The European shared-mobility map is shrinking, and the operators that survived the round did it on operational velocity. The spreadsheet-era operators got bought or shut.
Here's the uncomfortable part. A lot of fleet operations in this industry is still a small team making nightly decisions in a shared workbook. That model worked at 50 vehicles. It worked at 200. It breaks once you are competing against an operator running 15-minute decision loops on live demand and supply signals.
Saying this out loud feels harsh. It is also where almost everyone in this market started, and where a meaningful number of operators still are. Moving from manual to automated operations is a survival pivot dressed up as a technology project; the timeline matters more than the stack.
Zoox's latest expansion covers exactly one neighborhood — a few square kilometers, fully saturated with charging, rebalancing, and demand coverage. That is a smaller unit of expansion than the AV industry typically operates on.
The shift should change how the rest of us think about AV scale.
Most operators still measure growth in cities on a map. Twelve cities, thirty cities, fifty cities. The teams actually winning autonomous mobility have switched to a denser unit. A saturated rebalancing-and-charging loop inside a small zone outperforms a thin presence across a continent.
Utilization climbs, dwell time drops, charging logistics pencil out, and rider trust forms quickly because the cars are visible, available, and reliable inside a perimeter people can walk.
Density inside a zone is what compounds in AV operations. Wide, thin coverage looks like growth on a slide and runs underwater in practice.
Most fleet operators run dynamic pricing, free-ride credits, and active rebalancing as three separate projects, owned by three separate teams, optimized against three separate KPIs.
That is why they cancel each other out.
Pricing alone moves demand — until a surge with no nearby supply just generates complaints.
Free rides move supply on the cheap — until they over-cluster vehicles where demand can't keep up.
Rebalancing puts vehicles where demand will be — until it fights the pricing signal.
Each lever, run alone, eventually destroys the work of the others.
The win lives one layer up — in the orchestration that makes them harmonize. Pricing tuning demand, free rides shaping supply, rebalancing closing the gap. Six cards on what that looks like.
Vehicle manufacturers raised more than forty billion dollars for next-generation hardware last year. Battery plants, AV stacks, supply chains. Real capital flowing into real atoms.
That story is intact. What has shifted underneath it is where the marginal dollar pays off.
BMW i Ventures announced a $300M fund this month. Read the thesis carefully and the actual target reveals itself: operations platforms, AV middleware, demand orchestration, fleet intelligence. When an OEM deploys its next dollar into the software layer, it is telling you where it thinks the moat is.
A vehicle today is saturated capital. Improvements come in single-digit percentages on efficiency, cost, and range. A software platform sitting above the fleet can change how every vehicle is allocated, priced, and dispatched in real time, and those gains compound across the entire fleet rather than per unit.
The next decade in mobility will be decided by whoever orchestrates the fleet.
3 questions every fleet operator should ask before expanding to a new market:
1. What does demand look like in the first 90 days under three different fleet-size scenarios?
2. At what utilization rate does the new market hit unit economics, and how does pricing change that curve?
3. If demand undershoots by 30%, what is the cheapest path to break-even — vehicle redeployment, pricing, or geographic shrink?
Most operators answer these with a spreadsheet built the week before the board meeting. Directional at best, brittle to assumption changes, almost never re-run when conditions shift.
A real scenario engine costs more to set up. It needs proper demand modelling, integrated fleet data, and a runtime that lets you ask the question fifty times in an afternoon.
The output is structurally different in kind. You stop guessing and start choosing between trade-offs you can actually see priced.
Two weeks ago a fleet operator told me they had hired an ML consultancy last year to fix their demand prediction. The forecasts came back sharper. The operational mess did not move.
The reason was structural. The new forecast lived in a spreadsheet that fed nothing. Their dispatch system, pricing engine, and vehicle distribution logic each made decisions from their own stale view of demand. A sharper number, handed to disconnected systems, produces roughly the same outcome as a worse one.
Integrated platforms win on reliability before they win on accuracy. Every handoff between separate vendors is a place where signal degrades, latency creeps in, or the chain breaks silently.
A forecast that runs continuously inside the same loop as allocation and pricing does something the spreadsheet version cannot. It stays consistent with the decisions made downstream of it. The integration is the product.
Every Tier-2 mobility operator runs into the same wall at month 3 of an in-house AI build.
Hire DS → POC on toy data → real fleet data hits → integration cliff → quietly shelf.
Two things break in-house demand forecasting that nobody warns you about:
1. Fleet intelligence is a compounding-data problem. One operator's data isn't enough signal.
2. Demand is shaped by external context — weather, events, transit, fuel, neighboring operators — that your fleet data can't see.
The math only works at platform scale.
Heard this 4 times in 2026:
"We run B2B + B2C + B2G transport on the same fleet. Every demand-forecasting POC we've built has died at the same point."
Three traffic patterns. Three SLAs. One fleet. Single-model AI breaks on all three.
The honest answer is uncomfortable: multi-stream operations need three coupled forecasting models. Plus orchestration logic for shared assets.
That's not a "tweak the model" problem. It's an architecture problem.
Tier-2 operators who solve it first will out-margin everyone still trying to make one model fit three businesses.
Most operators get real-time backwards.
They buy live dashboards for strategic decisions and run actual operations on Excel.
Real-time has two jobs:
Operational — rebalancing, routing, dispatch, surge pricing. Reacts faster than a human. Most under-invest here.
Strategic — fleet sizing, market entry, infrastructure. Decisions made monthly, on inputs the live feed can't see. Most over-invest here.
The operational layer needs telemetry. The strategic layer needs simulation.
Stop letting the dashboard set the roadmap.
Smart machines without an orchestration layer are just expensive hardware.
Last week I gave the keynote at the 5th Fall Session of the World Technology Congress on Lake Como.
Here's the one idea I left the room with: the orchestration layer is the missing trillion-dollar piece of Physical AI. And almost nobody is building it.
When we talk about Physical AI today, we talk about machines. Autonomous vehicles. Humanoid robots. Drones. Every conversation is about making each individual machine smarter, safer, more capable.
That work matters. It's foundational. And it's only half of the problem.
A single intelligent vehicle does not move a city. A single intelligent robot does not run a warehouse. A single intelligent drone does not deliver a healthcare system.
Physical AI at scale is not one machine. It's a network of machines – hundreds, thousands, eventually billions – all operating in the same physical space, generating more real-time data than any human operator can possibly process.
The node is being solved. The network is not.
A connected vehicle generates 4–10 TB of data per day. Multiply by 10,000 vehicles. Multiply by the 100 largest cities in the world. No human dispatcher, no traditional fleet management software, no general-purpose AI model can handle that.
The vehicle can drive itself. But it cannot decide where to be. It cannot decide when to charge. It cannot decide which customer to prioritize when three request service in three different neighborhoods, during a thunderstorm, while a football match just ended two blocks away.
That is not a driving problem. It is an orchestration problem.
And orchestration at this scale requires a very specific kind of intelligence: AI that is natively aware of space, time, and urban context. Not a chatbot. Not an LLM with a plugin. Embodied intelligence. Software that lives in the real world.
That's what we're building at SWITCH.
Three takeaways from the room:
1) Cities: the fleets moving through your streets are only as smart as the intelligence layer behind them.
2) Operators: Telematics show what happened. Orchestration decides what happens next. Without it, profitability disappears.
3) Investors: The real Physical AI prize is the orchestration layer above everything. The window to own it is ~24 months.
Grateful to the entire WTC team for the invitation, and to every operator, policymaker, and investor who came up afterwards to continue the conversation.
The machines will get smarter. That's inevitable.
But smart machines without an orchestration layer are just expensive hardware.
