I’m either prompting or yapping. Forever need more tokens
Joined July 2012
- Tweets13.5K
- Following588
- Followers1.1K
- Likes21.1K
Jon retweeted
Qwen4 (Kerria on Arena) vs Claude Sonnet 5.5
Same prompt for both: a giant jack-o'-lantern in a witch hat with a full Dracula castle inside.
Sonnet actually built a small castle in there, with towers and furnished rooms. Qwen4 got the outside right, but inside it's just empty floors stacked on top of each other, more parking garage than vampire castle
Happy Halloween! 🎃
Jon retweeted
Introducing Pirate ZK Inference
You can now use DeepSeek-V4.1-Flash-1017 and other sovereign AI models anonymously:
Connect your agent, choose a model, and Pirate ZK Inference anonymizes your personal details before they reach your inference provider like OpenRouter.
Featuring zkAPI + a TEE enclaved scrubber model. This is Pirate Face's first feature in its upcoming privacy ladder for AI.
Try it below ↓
Jon retweeted
Retard, all code is opensource. intelligence is too cheap to meter.
Likely
Likely
Your gonna post all this and your making assumptions
You didn’t even dig into this????
Your gonna post all this but not lift a finger and just regurgitate info
Likely?? Seriously why spread stuff with likely’s and assumptions
First you call it a detail, then you say fix already existed to then saying likely 2x
No ai wrote this bullshit your dented smooth brain did
I’ll eat my words if wrong but shut up first
Jon retweeted
So I migrated my SaaS from @Vercel to @Cloudflare…
…and about a week later, it went down for ~40 hours before I realized it.
To be fair, the migration itself was a pretty big one. TruckFull is a Next.js app with authentication, geolocation, OCR/VLM usage, payments, webhooks, data pipelines, background processing, and a bunch of other external services.
I tried OpenNext first, but it fought me on enough compatibility issues that I switched to Cloudflare Containers. That allowed me to keep my environment setup instead of maybe having to rewrite half the app.
Before moving production, I tested the lite container with 50 requests at once. All 50 returned 200. Peak memory was around 165MiB, with no OOM. Staging passed the usual route, auth, webhook, image, sleep/resume, and rollback checks.
I thought I was good
Then I got an email from Facebook at 2:30 in the morning saying they couldn't find my privacy policy
"Uh-oh" I thought
Went to TruckFull dot food, the production container stopped accepting connections.
It didn’t 'crash'. It was still ~running. It just stopped serving my website.
While investigating, I found ShapBot, Parallel’s (@p0) AI search crawler, hitting the site heavily. Thousands of requests in one hour, including relatively expensive pages.
That looks like what pushed the small container over the edge. I’m still checking the exact timeline, but it was enough to make me block ShapBot while I investigate.
Restarted the container so the site is back online now. And the container has more room. I'm working on setting up a little 'canary box' that will notify me when production goes down, as well as give AI the ability to study what's going on and provide me with analyses.
I guess just in summary maybe just a warning to other devs thinking about migrating from Vercel to CloudFlare - consider simulate crawler loads or pick a stronger config when first migrating
I still believe the migration was the right choice and a sincere thank you to @noisemakerjon who pushed me to migrate sooner than I felt comfortable, because that's how we grow
Jon retweeted
Im currently creating a game to learn Inference Engineering
You play as an engineer who deploys and optimizes open-source models in data centers, at client sites, and so on
You'll complete missions, advance from level to level, and earn rewards. Hope you will enjoy it!
Jon retweeted
Brother the answer is in your screenshot. Orange cloud time.
@vercel_support plz help my site (cubicles.lol) just stopped working because of a DNS issue - just as we're in the middle of going viral on Product Hunt.
apparently we have a serverHold with the .XYZ registrar and AI is saying I need to ask my registrar, which is you.
Jon retweeted
For anyone that wants to try jev or don’t understand it
“A new type of model called jev came out. In certain scenarios it can boost accuracy by 6x, speed almost 100x and cost 1/15 of a llm. It doesn’t have the same capabilities but its special and unique in its own way. Look at all my projects and repos that you can find or have access to and see if this model can replace or substitute anything we’re currently doing poorly. Come back to me with data and specs on what can be changed and why. (Use vercel ai gateway to call the model here is my key I will rotate soon don’t worry ____(tell user to get api key and where if this is still blank, but continue and wait for key till last moment or when we need data on what good it will bring in the end))
typesafe.ai/blog/introducing…”
Jon retweeted
hi everyone, i’m looking for remote work opportunities and i’m based in Sydney
i took a break from working for a few months to build web apps/mobile apps for fun but i don’t think it’s for me, i want to build something bigger as part of a team
i’ve worked for 3 years as a full stack software engineer building products for startups in SF remotely
for one startup i was one of the first engineers and helped them grow from seed to series B
my main experience is with Nextjs/tRPC/Prisma but i’ve used other frameworks like Tanstack/Drizzle/Convex for side projects
feel free to DM! i’d also appreciate any comments/reposts from my network, thanks guys
Jon retweeted
Context7 is now an OpenCode plugin.
opencode plugin @upstash/context7-opencode
MCP server + a skill that auto-triggers doc lookups. One command, no config.
context7.com/docs/clients/op…
