Kevin Brown (@mathsociety) header
Kevin Brown (@mathsociety) profile picture

Founder of @CloudWorldModel. Estimate what an infrastructure change does to cost, latency, and errors before you ship it. Ex-Oracle Cloud Engineer.

Stamford, CT
Joined February 2009
PSN slowed 33 minutes after the GTA VI unlock. Whatever Sony's cause was, this is the trap we see in models: autoscaling adds servers, and every new server opens more connections to the same database. You can test that before launch night.
We simulated it: a game-launch sign-in rush on an assumed topology, not Sony's real system. At 50x normal load, servers autoscaled 4→22, but 40.6% of sign-ins still failed. Most of that was the shared account DB running out of connections. p99 went from 188 ms to 1.9 s.
107
"An alert has no test" is the core problem. Simulation could give you one: inject a known fault (lock contention, failover surge, capacity loss) into a modeled system and you get incidents with ground-truth root causes, as many as you want. Could that work as an RL environment for SRE agents?
Introducing SRE-1 a 27B model that dominates the Pareto frontier of accuracy/cost on production debugging, beating frontier models in a coding-adjacent domain Built by 11 people on open weights models + domain expertise. Final run: 7.4 days on 512 GPUs
1
117
We simulated the Microsoft 365 outage pattern: a lock-contention cascade took an assumed 7-worker gateway from 6 up to 0 in 5 steps. Cutting traffic didn't recover it; stopping new work did. Assumed topology, not Microsoft's system.
1
70
Why did the Microsoft 365 outage drag on? In our sim, cutting traffic to 240 RPS wasn't enough. Restarted workers got swamped by queued work and seized again. Only stopping new work let 6 of 7 rejoin. Assumed topology, not Microsoft's. Their report notes the same queued-traffic swamping.
1
36
Microsoft's preliminary M365 report: a latent lock-contention bug, not just too much traffic, cascaded the Nov 12 outage after a load shift and ~14% of gateway capacity already out. We modeled the pattern on an assumed topology, not their real system.
We simulated a simplistic version of the Microsoft Outage: a lock-contention cascade, not just capacity exhaustion. Assumed 7-node gateway, not Microsoft's. 1 of 7 out. Nodes at 450 RPS (cap 500, lock 425) seized one by one. 5 steps later: 0 of 7 up, 0 of 2,700 served. Load the outage yourself: cloudworldmodel.ai/workspace…
1
2
98
If there was one person i’d be interested in speaking to it would be Madison Huang. Hey @boardyai introduce me to Madison. Thanks!!!
1
134
There’s merit to open source and to keeping your code closed. @floci_io is an open-source AWS emulator. AWS recently announced its own emulator. I don’t know if AWS looked at Floci, but it highlights a trade-off: anyone, including much larger companies, can study what you build. github.com/floci-io/floci
1
75
I’m here waiting on RCA for the Microsoft 365 (Teams, Outlook, etc) outage. I checked this morning and still haven't seen an official one. I keep looking for the RCA every few hours. Reviewing RCA’s is interesting to me. I’m aware that i’m strange. Hehe 🤷🏾‍♂️
1
407
I’m still checking, for the most part RCA publicly unknown. It was a bad one though. 70+ countries/territories
12
How accurate is cloud simulation? We load-tested it and published the misses too. AWS: 93.8% on /accuracy. GCP, one topology in us-central1 at 300–500 RPS: 5 of 6 P95 holdout runs within 20%, 1 within 35%, all 6 P50 within 20%. Measured P95 was slower than predicted at every rung. cloudworldmodel.ai/blog/aws-…
1
50
There are lots of startup products that feels like they works because it does the visible thing it promised but when you look at the outcome It just doesn’t work for anyone involved. We want the actual outcome to be real.
1
52
An example would be the show love is blind on Netflix. If you get to the reunion part of the show and no one is still married, can you say love is blind? Yes, no, maybe? Basically, Cloud World Model wants those that use our predictions to be married by the end of the show. I hope that makes sense.
1
60
As I write this I can now think of the one question I need to ask everyone that I show the infrastructure report to. Brainstorming out loud. Thanks
8
Nice. We are not an emulator. We are building a World Model of how the Cloud behaves.
Today we're launching AWS Emulator, a free local emulator of AWS services for building and testing your apps offline. Every line of its code was written by autonomous AI agents running on Amazon Bedrock AgentCore. Hundreds of agents, organized into three teams, write the tests, find the issues, and fix the code, each change passing through an automated merge gate that checks compilation, security, coding standards, and fidelity to real AWS. go.aws/3Xk2dMV
1
94
An emulator emulates services. We don’t emulate service. We leave that to LocalStack, Floci and others. We model behavior. Slight difference.
39
The part of the cloud bill that grows fastest with traffic usually isn't compute, it's egress. I modeled a small web stack at 3x traffic, and every dollar of the increase was data transfer out. Wrote up how AWS, GCP and DigitalOcean bill it, plus 5 ways to cut it 👇🏾
Egress quietly becomes one of the biggest lines on a small cloud bill. In our September 2026 pricing snapshot, tripling traffic (100→300 RPS) raised one small web stack's modeled cost ~48% on AWS, ~54% on GCP and ~12% on DigitalOcean, all from egress. 5 ways to cut it, with dated list prices: 
cloudworldmodel.ai/blog/clou…
1
85
I’m a big rap fan. The Cloud based startups are like rappers. Rappers are always beefing with each others. The rap “podcaster” though they are all on each others shows getting revenue together. I want CWM to be more like the rap “podcasters” than the rappers.
64
Built this so an agent can ask "what breaks if traffic triples?" before touching infra. In this run it was app CPU, not the database, and the post shows exactly where the model says it doesn't know.
What happens to your AWS stack if traffic triples? We pointed an MCP client at Cloud World Model, simulated ALB + 2× m5.large + db.r5.large MySQL, and went from 200 to 600 RPS: p99 87 → 422 ms
errors 0 → 10%, all app CPU
MySQL CPU flat at ~18%
modeled cost $0.76 → $1.42/hr Real tool calls, honest limits:
dev.to/canvascloudai
80
I give Palantir credit here’s why.
Welcome, Startup Fellowship Cohort 003, and thank you to all who applied.
1
208
We applied to lets say 10 programs this summer. All straight rejections out of those that have responded with exception of Palantir they actually gave us an interview though we didnt get in. Thank you @PalantirTech, appreciate the consideration 🙏🏾
1
82
Cool @ number one
Top 50 most talent-dense breeding grounds for founders. Palantir at #1, Stanford at #2 and surprisingly Yahoo at #3.
39