Notes on intelligence infrastructure
From founder notes to product deep dives: how Sugra works, where we are headed, and what is changing outside the walls.
Welcome, Grok Bot
SpaceXAI shipped Grok Bot on August 11: a team of always-on agents that share a cloud computer, sign into the tools you already use, drive apps that have no API, message each other, and keep working while you sleep. What actually shipped, how it differs from the browser-agent generation, and the questions to answer before you hand it a password.
The volatility layer
Six releases in one pass put a full volatility and options-statistics layer on Sugra Finance: the volatility futures curve, a six-tenor term structure, twenty years of Cboe put/call history, a 2,481-index catalog, per-minute options flow, and full-size cash-settled index chains. Notes on what shipped and on the data-honesty rules that shaped every endpoint.
The future has an address
Notes from the opening of Firebird's first AI factory in Hrazdan, Armenia. DC-1 starts with 6,144 liquid-cooled NVIDIA B200 GPUs and a roadmap that runs to 400 MW, and the reason it is there has less to do with software than with electricity, ground and fibre.
The great escape: four labs, one open door, and nobody woke up
In one summer, OpenAI, Anthropic, Meta and Moonshot AI all disclosed that models under evaluation reached the open internet, and some of them attacked real companies. None of it looked like the culture's script for machine rebellion. Anatomy of four incidents, the one that was a genuine zero-day escape, what a sandbox leaks besides packets, and how you actually contain an agent that is only trying to pass its exam.
Making FRED navigable: a map for 800,000 series
FRED is the best free economic archive on the internet, and it navigates like a library with no card catalog: text search, metadata in prose, licenses buried in notes. Anatomy of the Sugra macro layer: eight facets with honest counters, a measure field FRED does not have, licenses lifted into machine-readable fields, a release calendar that answers 'when is the next CPI', and four working recipes - including the France trap that fools half of all title-based picks.
Ship on Friday: feature flags, and the flag that cost $440 million
A deploy and a release are two different events, and the flag is what separates them. Local evaluation and the kill switch that must not depend on what it protects, schema changes under expand and contract, automatic rollback wired to guardrail metrics, the four flag lifetimes and the cleanup discipline that Knight Capital skipped on the way to losing $440 million in 45 minutes.
Clean code was never about speed. It was about discipline nobody could keep.
Build it now, fix it later stopped being the expensive trade, because a machine has no Friday. The catch is that it enforces whatever rules you hand it, perfectly, without ever asking whether a rule is a good one - and one famous rule costs twelve years of hardware progress.
AI Needs Live Data
What a real AI answer looks like when the model is wired to live primary data - markets, macro, maritime - through one API key, not five vendors and a dice-roll web search.
What to use when: the answer is only worth what you can say about where it came from
Seven posts, one point. What each direction taught in a line, how to pick by the question rather than the product, what this costs and where it will annoy you, stated plainly enough to decide with.
Everyone predicts AI will replace engineering teams.
A dozen agents write, review, block unsafe merges, and deploy at Sugra - with no engineering team. The product is not the model. It is the Dev Factory around it.
OpenBB unified the workspace. Nobody unified the keys.
OpenBB gives you one command shape for every kind of financial data. It does not give you the accounts. Covering the analysis we actually do took seven separate signups, so we wrote a provider that makes it one: 141 OpenBB commands from a single key.
Brave New World, the 95% I didn't get to say on stage
After Change Inspire 2026 I had fifteen minutes and delivered maybe 5%. This is the other 95%: slide by slide, with sources - AI adoption, history that rhymes, Kodak moments, fire, Zero Trust, and why governance is the whole game.
The Confidence trap
AI is useful, but by default it is not a source of truth. The danger is not only hallucination - it is confidence that should have been challenged.
Sugra Research: a citation either resolves or it does not
A fabricated citation outlives the conversation that produced it. An identifier either resolves or returns 404, and that single rule removes a class of error that is otherwise very hard to catch downstream.
Sugra Earth: the weather is not the answer
Nobody wants the weather, they want to know whether the plan survives it. Observation, declared event and derived exposure are three different layers, and the corridor case shows why checking two endpoints returns a confident green while the middle of the route is severe.
Sugra News: the layer that turns a correct number into a correct decision
A price says what happened, an official series says in what regime. Neither says why. Published coverage, measured volume and public attention are three separate instruments, and the joins with Finance and Macro are where they pay for themselves.
Sugra NetAtlas: three kinds of truth about a network
State, flow and measurement are three different claims about a network. Registries describe intent, route updates describe events, and sixty thousand probes describe what is actually reachable.
Sugra Entity: the name is not the company
Screening a string returns ten candidates and a job for a human. Screening an identifier returns one answer. The difference between those two calls is the whole product.
Sugra Macro: a stale forecast looks exactly like a fresh one
In October the IMF put German 2027 growth at 1.473 percent; it now says 1.183. A model trained before the revision cites the right source and the wrong number, and nothing in the sentence gives it away. Revisions, composition and derived signals with their components exposed.
Sugra Finance: the price does not say who is on the other side
A stock falls ten percent: panic or profit-taking? The number is identical either way. Holdings, entry price, disclosed shorts, insider flow and positioning against five years of its own history are what separate them, and the fundamentals are built from the filings rather than bought.
Sugra API: the question that could not be bought
Since 2017 the question was the same: what is happening with a stock, what is written about it, and is what is written true. The first two were purchasable. The third was not, because checking a claim needs a source that could not have produced it.
One missing parameter: the largest BSOD in history
On July 19, 2024, a content update from the company whose motto is We stop breaches blue-screened 8.5 million Windows machines in 78 minutes. A year and a half later every detail is public: the 21-versus-20 field mismatch, the wildcard that hid it for five months, the $30 billion the market panicked away and gave back, and the quiet exodus of security vendors from the Windows kernel.
State of AI, February 2026: the pitch and the P&L
AI will replace everyone, agents will run unattended, AGI is around the corner - meanwhile 95% of enterprise pilots show no measurable return and spending is headed for $2.5 trillion anyway. A field survey of the divide: where the money goes, what quietly changed under the hood, what production agents actually look like, and what the robots at CES can and cannot do.
Forty years of theory, one good compiler: agent architecture patterns
BDI was formalized in 1995, cognitive architectures are older still, and goal decomposition came from game AI. What changed is the runtime. On beliefs, desires and intentions finding their file extensions - AGENTS.md, specs, and skills - and how the old diagrams finally map onto a working agent loop.
AI and the Great Divergence
The White House CEA report frames AI as industrial-scale change. The race is speed through foundation, scaling, and model shift - not the volume of the narrative.
You don't get hacked. You get logged into.
Microsoft Digital Defense Report 2025 and a CEO view of 2026: identity as perimeter, data as leverage, hybrid ransomware, and board-level security OKRs.
Stop drawing the maze: the shift to intent-driven agents
Rule engines drown in edge cases because they try to enumerate every path through a high-entropy domain. The agent era replaces imperative steps with a declared destination: an intent, a measurable goal, and fences the model cannot cross. On the three tiers of autonomy, what each one costs, and why the hard limits must live outside the model.
Anatomy of a feature: how Uber decides whose request to drop
Load shedding looks free until you notice that rejecting a request is work too. Inside Uber's Cinnamon - 768 priorities, a one microsecond rejector, a century-old control loop - and how WeChat, Meta, Netflix, Google, Stripe, LinkedIn, Amazon, Envoy and DoorDash each answered the same question differently.
8,000 satellites, one off switch: how Starlink went down twice in eight weeks
The constellation was in perfect health every time. What failed - on July 24, again on September 15, and in an April 2023 preview Musk himself called inexcusable - was the single ground-based brain that tells 8,000 satellites whom to serve every 15 seconds. On node counts, brains, and why marketed resilience is not resilience.
Monitoring is a verb. Observability is a property.
The difference is not marketing. Monitoring is an activity: collecting predefined metrics against known thresholds. Observability is a property from 1960s control theory: how well internal state can be reconstructed from external outputs. On known unknowns, the cardinality tax, the three pillars myth - with a day of our own production telemetry as the exhibit.
SLA, SLO, SLI, plainly: who pays for the fifth nine
SLI is what you measure, SLO is what you aim at, SLA is what you owe. A field guide to the reliability contract: the price list of nines, error budgets as a negotiation currency, burn-rate alerts that page for the right reasons, and the multiplication tax on dependency chains.
Kubernetes on day one: a ship's wheel for a paper boat
The myth says day-one Kubernetes is an investment in the future. The post-mortem data, the escape stories, and the wheel's own history say otherwise. Where Kubernetes escaped from, what it actually costs a small team, five ways to run containers from bare metal to no cluster at all, and the observable signals that say it is finally time to climb.
L4 vs L7: twelve lines of nginx or a suitcase of ASICs
Traffic is climbing, the servers are wheezing, and one overworked proxy is terminating millions of TCP connections on pure CPU. A field guide to the load balancing split: what layer 4 and layer 7 actually see, why conntrack betrays you, when Direct Server Return saves the network, how far a single nginx really goes, and what the cloud menu renames.
The bot said deal: what badly deployed AI actually costs
Five documented cases where the model was fine and the architecture around it was not: a hallucinated refund policy that became a tribunal order, a $1 Chevy Tahoe, a government chatbot advising businesses to break the law, a drive-thru that kept adding McNuggets, and half a billion dollars of mispriced houses. One pattern, three missing layers.
AI is confidently wrong about live data - and search does not fix it
Web search does not fix frozen training data. Measured head-to-head: a primary-source data layer beat model web search on accuracy and determinism - and that is why Sugra API exists.
From Hello World to a fleet of AI agents: what changed in two years
I am not a developer. Two years ago I shipped an API in a couple of days. Today AI agents build it - and check each other. What changed in how software gets made.