Add preferred source
At AI Engineer Paris 2026, three speakers converged on a single uncomfortable truth: the bottleneck in AI is no longer model capability but the slow, messy work of organizational redesign. An economic historian opened the event by showing that steam engines took over a century to move productivity statistics, while electricity needed factory layouts to be rebuilt from scratch. A solo developer known as “The Kitsa” then demonstrated what that redesign looks like at the individual level — a sprawling homemade stack of Proxmox sandboxes, load balancers, and anti-slop linters held together by someone who admits he hasn’t seen a config file in two years. Mistral’s VP of Engineering closed with the enterprise answer: a unified agent harness, dynamic permission systems, and a sovereign data center outside Paris. The through-line is that AI infrastructure investment is already at 1.8% of US GDP and forecast to hit 3% by 2028 — a boom bigger than the fiber overbuild of the late 1990s. Whether it produces durable productivity gains, the speakers suggest, depends less on the next model release and more on the unglamorous connective tissue: sandboxes, guardrails, connectors, and policies that turn raw capability into deployed work.
Key Elements
Four years into the generative AI era, the smartest people in the room are no longer arguing about whether AI will transform the economy. They’re arguing about when — and the answer keeps pointing at the least glamorous part of the stack. At the opening night of AI Engineer Paris 2026, speaker by speaker, the story stayed the same: the technology is arriving faster than the organizations designed to absorb it.
An economic historian set the frame with a chart that should humble anyone expecting a smooth productivity boom. The steam engine was invented in 1712 by Thomas Newcomen and improved by James Watt later that century. Productivity data didn’t budge until the 1830s — over a hundred years later. Not because the engine was useless, but because it needed machine tools, railways, steamships, and a newly skilled labor force before an entire economic system could form around it.
Then came a developer known on X as “The Kitsa,” who walked the audience through his own attempt to close that gap manually. His setup involves a Proxmox server spawning a fresh virtual machine for every coding task — he calls them “orbs” — each with its own database, seeds, and browser, so an agent “cannot break out, they cannot do something stupid — they just have mini computers within my computer.” He runs eight at a time before his hardware gives up. He pays for four $200-per-month AI subscriptions and built a load balancer to juggle them based on weekly limits. And when asked about his config, his answer was pure 2026: “I don’t fucking know. I haven’t seen a config in two years. I just wish for things and the things magically happen.”
Then came Mistral’s VP of Engineering — the company’s third employee — to argue that enterprises can’t and shouldn’t live like The Kitsa. They need a unified harness, dynamic safety policies, and infrastructure they actually control. Mistral, he said, is now 1,300 people, raised a €3 billion Series D last week, and is building its own 10-megawatt data center outside Paris.
The Two Clocks of Technology
The economist’s core framework is simple enough to explain at a dinner party and devastating enough to reshape an investment thesis. Technology moves at two speeds: the fast clock of invention and prototypes, and the slow clock of deployment — when a technology gets deeply integrated into how firms and economies actually work.
Electricity is the cleanest example. Thomas Edison flipped the switch on Pearl Street Station in 1882. Factories didn’t meaningfully electrify for another 30 to 40 years, not because the power wasn’t there, but because factories had been built around steam engines. Machines were arranged according to where belts and shafts could reach, often over multiple stories, all driven by one power source. Electricity’s real advantage — that you can generate power in one place and consume it in another — only mattered after factories were completely retooled to use a “unit drive system” where every machine had its own motor. By 1930, 80% of factory power was electric.
Computers followed the same pattern. Robert Solow’s 1987 quip — “you can see the computer age everywhere but in the productivity statistics” — captured the lag perfectly. PCs were everywhere, but productivity didn’t respond until the mid-to-late 1990s, when enterprise software, database technology, and supply chains had finally been redesigned around them.
The economist’s data table makes the lag unmistakable:
| Technology | First Deployment | Productivity Shows Up | Lag | What Was Required |
|---|---|---|---|---|
| Steam engine | Newcomen, 1712; Watt later | 1830s at earliest | 100+ years | High-pressure engines, machine tools, railways, skilled labor |
| Electricity | Edison’s Pearl Street Station, 1882 | 1920s | ~30–40 years | Factory redesign — unit drive replacing group drive |
| Computers | Widespread by 1987 | Mid-to-late 1990s | ~10 years | Enterprise software, databases, supply chain redesign |
The pattern holds: the bottleneck is never just the core technology. It’s the complements — the organizational changes, the skills, the new processes — that turn possibility into output.
The Capital Markets Are Running Ahead
If the history lesson has a warning, it’s this: capital markets move on the fast clock. Investors price in transformation long before it arrives in productivity data.
The economist offered two precedents. In 1840s Britain, Parliament authorized hundreds of railway lines in a rush of cheap capital and the belief that every town needed a connection. Railway investment hit 5–7% of GDP in a single year — one of the largest booms in history. In the late 1990s, the rush to lay long-haul fiber for the commercial internet hit about 1–1.5% of GDP in 1999.
AI infrastructure spending is already bigger than the fiber boom. It reached about 1.8% of US GDP in 2026 and is forecast to hit roughly 3% by 2028. The economist declined to call it a bubble but noted a historical pattern that should give investors a framework: even when private returns disappoint, the infrastructure can outlast the mania. Surplus railway tracks and “dark fiber” from the dot-com era later enabled cheaper transportation, streaming, and the mobile era.
The question he left hanging: is the current buildout overbuilding that will produce durable productive infrastructure, or capital destruction that will take a decade to absorb?
One Solo Developer’s Answer: The Slopware Factory
The Kitsa’s talk was the most jarring of the night because it showed what closing the gap looks like without institutional support: chaotic, fragile, and held together by one person’s accumulated hacks.
He walked through the evolution of AI coding from Copilot tab-completion to what he calls “YOLO mode” — agents running with complete permissions. “Take my production databases, customer data, GDPR, whatever, man. Take everything, just finish the job.”
He offered a taxonomy of the current workforce that landed with uncomfortable laughter. Slop grenade throwers — the majority — prompt agents, generate output they don’t read, and throw it at colleagues with “good luck, hopefully it doesn’t explode production.” Meat proxies sit between a manager and an LLM, taking the manager’s prompt, feeding it to the model, and barely reviewing the result before passing it back. Many companies are now “layers and layers of meat proxies talking to each other.” The third group — actual engineers — are a small minority trying to build systems around their processes.
His prediction for next year: “Many people are gonna wake up to the fact, they’re gonna go to the mirror and be like oh fuck, and they’re gonna get fired like it’s inevitable.”
The Orchestrator Problem
Three years into the orchestrator boom — Cursor, Conductor, and a dozen others — The Kitsa’s frustration was blunt: “No one is even trying.” Every orchestrator is a copy-paste of the same layout: chats on the left, chat in the middle, git diff on the right. “We just raised 37 million for a new orchestrator and the screenshot is like, is that Codex? Is it Cursor? This is the same shit always.”
His own journey led him to Proxmox — a higher-level operating system that runs multiple Linux VMs — plus Traefik and PM2 so every process is accessible at a nice URL. “Anytime you want to see how your landing page looks like in dev, you can go to dev.myURL.whatever and you’re gonna see it and you’re gonna remember it and you’re gonna bookmark it, it’s not gonna be a random IP address.”
The Anti-Slop Stack
The key insight from The Kitsa’s experience: without external guardrails, agents will disable your rules to get the job done. “I just disabled five of your configuration lines,” as he describes agent behavior. The only solution is enforcement that lives outside the agent’s reach.
| Tool | Function |
|---|---|
| OXLint | Linting |
| Ultrasight | Additional checking |
| Anti-slop (by Dylan Monroy) | Rule enforcement |
| Police CLI | Custom rules lint can’t express — e.g., “never use your own buttons, always use Chakra UI buttons” |
| Jev (Jev Rabbit) | Checks entire PRs against plain-English rules — “you can literally express yourself in the English language” |
He also built Skillbox, an open-ur skills while a coding agent gets twenty — and they don’t need to know about each other. “My setup is portable. I can just take it and plug into something else and it works.”
The Factory View
The Kitsa previewed Bench, his own orchestrator, which he called “the first beautiful agent orchestrator that you see in your life.” The core idea: a factory view showing every task moving through stages — Solbot coding, Solfast reviewing, a tester clicking around — until a task is fully reviewed and shipped. Tasks move from to-do to ready, through the factory process, to review. “You actually feel like you’re coordinating a factory and not just talking to a bloated orchestrator app.”
His closing pitch was that he can run the whole thing from his phone: “I don’t use CLIs, I don’t use Codex apps, I don’t use anything. I can open it on my phone, I can open it anywhere, and just from any underpowered device I can orchestrate my entire slop factory of things.”
Mistral’s Bet: Sovereignty, Safety, and the Unified Harness
If The Kitsa represents the individual frontier of the diffusion gap, Mistral’s VP of Engineering represented the institutional answer. His framing: Mistral “is not looking for AGI. We’re looking for usefulness.”
The company has been on an acquisition streak that maps directly to the diffusion problem. The Koyeb acquisition was about agentic workloads and sandboxes — “the unit of compute that you need to actually run workloads at scale.” The Emmi AI acquisition, announced in May 2026, is aimed at physical AI and simulation — “providing the foundational intelligence required to design and build the next generation of aircraft, vehicles, and semiconductors.” There’s also a $150 million partnership with Orbis and a €3 billion Series D, “the biggest round from a technical company in Europe.”
The Unified Harness
The Mistral VP introduced “Steven,” a finance person who wants to ask simple questions but also build applications — pulling data from a lake, spreadsheets, and Notion, wrapping it into an app to share with coworkers. He operates under corporate data guidelines. Mistral’s answer is a unified version of Vibe, its end-user product, that adapts to the user’s role without changing the underlying harness.
“You can start a session locally on your laptop, teleport it to the web, execute it in sandboxes, check it on your phone, steer it, push to GitHub, check CI, and iterate.”
The platform compounds with use: someone creates a skill, someone creates an app, and they share them with the team. He gave a concrete example: an on-call engineer paged at 3 a.m. gets an asynchronous agent investigation that plugs into Grafana, Sentry, and GitHub to understand what could have regressed, then posts findings to Slack so the engineer starts from somewhere instead of cold. “You used to need to do a lot of plumbing to do this. The unified Vibe version means the user just enters a prompt like ‘when interview preview feedback is missing, send a personalized reminder to the interviewer who forgot to do it’ — and Vibe understands the intention, spawns agents in the background, builds code, runs tests, deploys, hosts, and iterates for minutes or hours.”
After a week of execution, you can see what worked and what didn’t. Maybe the agent called a Grafana MCP incorrectly because it didn’t understand an undocumented option. Looking at traces, the skill can self-correct — more token-efficient and faster next time.
The Agentic Safety Dilemma
The safety section was the most technically interesting of the Mistral talk because it confronts a real tension head-on. He framed it as a dilemma: “When you add all those security guardrails everywhere, you basically break all the use cases because you had such a powerful agent, such powerful tools, but then now you’re restricting everything because you fear everything.”
The threat model is real. Agents have privileged access to many connectors and data. They’re non-deterministic — a guardrail that works today might not work tomorrow. And attacker-controlled inputs are getting more sophisticated: every model has a jailbreak, and as soon as data is fetched from outside, prompt injection is a constant threat.
Mistral’s answer is dynamic collaboration — granting the right amount of permissions at the right moment, with good reason, and filtering which operations should be allowed and why. In the demo, Vibe CLI reads a GitHub issue containing a prompt injection. The agent posts a comment exfiltrating a secret key. Mistral’s solution: prompt injection rules that describe classifiers, define sources and thresholds, and get attached to enterprise policies. The action is stopped at the network level without the user configuring anything. “That’s the power of enterprise controls — the user never opted in to get protected. This is a policy that has been defined and pushed at the enterprise level.”
The Sovereignty Stack
Sovereignty, as the VP defined it, has three layers. Data sovereignty — where data is stored, who has access, which jurisdiction can read it. Intelligence sovereignty — where model behavior analysis and fine-tuning happens. Compute sovereignty — where inference transits, where audit logs live.
Mistral’s answer to all three is Les Ulysses, a 10-megawatt data center near Paris running NVIDIA B300 hardware — “one of the largest data centers with this kind of cutting-edge hardware from NVIDIA in Europe.” It serves critical internal workloads, hosts Le Chat’s next model training, and runs the most sensitive customer workloads. “This is how we can have from end to end a full sovereignty path that we present to the customers, including agentic safety, but also including inference and training.”
Specialists, Not Just Generalists
Mistral’s model strategy has a clear answer to the diffusion problem: generalist models aren’t enough. Some specialist models serve broad purposes — OCR 4.1, announced in summer 2026, can transcribe handwriting from scanned 1960s documents, distinguishing captions, titles, tables, and images. It powers Mistral’s document library within Vibe, where users upload terabytes of unstructured documents and query them at runtime. A live demo showed real-time transcription and translation with accurate speaker identification — diarization — in a noisy meeting.
Other specialist models are fine-tuned for specific customers with rare languages or proprietary programming languages “that nobody has heard of. By default, coding models are not good at it. So how you can actually sit down with a company and make a specialized model makes an amazing difference.”
The banking testimonial made the ROI concrete: a KYC process for corporate clients in Belgium went from 80% of incomplete files sent to the middle office down to 10% with two agents. The audio family, Voxtral, includes text-to-speech with voice cloning so convincing that when he played two samples and asked the audience to identify which was him, “the second sample was not me at all.”
The Diffusion Gap Is the Investment Opportunity
The three talks, taken together, describe one problem from three angles: the gap between what AI can do and what organizations are structured to capture. The economist’s answer is all of the above — core technology, adoption, complements, reskilling — over a longer horizon than anyone in the room would like. The Kitsa’s answer is complements and tooling: sandboxes, guardrails, load balancers, skill bundles. Mistral’s answer is adoption and safety: unified harness, enterprise policies, sovereign compute.
The unresolved question is whether the current infrastructure boom — 1.8% of US GDP, heading for 3% — will follow the railway and fiber pattern, where overbuilding produces durable productive infrastructure even if private returns disappoint. The economist’s history suggests the infrastructure outlasts the bubble. The Kitsa’s experience suggests the tooling layer is still pre-standardization — every orchestrator looks the same, every developer is building their own factory, and the meat proxy problem is endemic. Mistral’s bet is that enterprises will pay for a unified, safe, sovereign stack rather than assembling one from parts.
What to watch: whether Bench and similar orchestrators standardize the software factory layer, whether Mistral’s sovereign data center and agentic safety framework win enterprise deals, and whether AI infrastructure investment at 3% of GDP produces measurable productivity gains before the next investment cycle turns. The economist posed the question without an answer. The other two speakers spent their time building answers in real time — one a fragile personal stack, the other an institutional product. Both are betting the bottleneck isn’t the model. It’s everything around it.
Full content available at:AI Engineer Paris 2026 Opening Keynotes: Mistral, Langfuse & Sizzy | Day 1
- Candid Holds ‘AI Cheating Day’ in South Korea, Bringing Together 100 AI Practitioners to Swap Workflows
- AI Startup Plaud Doubles Singapore Investment to $15.7 Million With New APAC Hub
- Samsung Electronics Unveils AI and Robotics Blueprint in Silicon Valley, Aiming to Become “AX Leader”
- AI Model Release Cycle Shrinks from 125 Days to 44 Days—Self-Improving AI Raises Control Concerns
- Meta’s Muse Agent Fuels 17% Arm Surge as Wall Street Bets on a CPU Supercycle
Once added, BigGo Finance appears first in Google Search Top Stories, so you get the broadest, most up-to-the-minute, and most comprehensive global financial news first.
