i like coding and talking about it // swe @upstash building @bydefaultso
get recommended in ai results
Joined April 2023
- Tweets3.4K
- Following326
- Followers23.5K
- Likes15.1K
an update on our natural-writing ai model
it can take structural AI slop and turn it into nice, natural human output
the goal isn't to trick ai detectors or anyone into thinking it's human, it just reads so much better than ai slop
wanted to see how fast AI would start recommending a new product
we launched Upstash Blob (file storage) 6 weeks ago and optimized everything for AI recommendations
the clankers are really starting to like it 👀
what if agents could write naturally?
that's what we're trying to solve
human-quality blog articles, website copy, etc
we hit $1k MRR a few weeks after launch. nothing crazy, but i wanted to share what helped us get there
let's have a chill coffee chat
these stupid boosted paid launches need to end
why the fuck would anyone pay for twitter engagements and reposts i seriously don’t get it
Introducing SubQ Snap, the first long-context decision model for real enterprise work.
Snap is the only decision model or LLM where compute per token stays constant as context grows.
• 384K tokens of multimodal context, 12x larger than Jev.
• 61.7 on Decision Index v0.2.1 vs Jev’s 57.9, outperforming on 13 of 16 enterprise benchmarks.
•Processes 100K tokens and returns a decision in under one second.
•$0.05/MM input tokens. Output is free.
Available today to anyone.
The first 10,000 sign-ups get 100M free input tokens.
We’ll likely hit capacity fast and will add users as we add compute. Compute offers welcome 🙂.
Latency will be variable as we tune the platform to this new model and category of users.
Model speeds are 80ms for minimal context, 250ms for 32k, 660ms for 100k, and 2.5 seconds for 384k.
Visit: subq.ai/
2
22
3,116
emil kowalski's break-ui skill might be the best skill i have ever seen 😭😭😭
it throws the worst realistic data at your UI and shows you what breaks
Clerk is absolutely dominating AI search recommendations. ChatGPT, Claude and coding agents love to recommend Clerk. Here's a deep analysis on why they're winning and how you can copy the strategy
> Clerk is #1 in AI search for auth. They're mentioned in 64% of answers, more than Auth0, Supabase or Better Auth
> AI calls it "the default" and even tells people to use Clerk "unless they have a strong reason not to". Claude says "it's the default for teams that care about time-to-first-working-app". Agents care a lot about time to first value.
> Coding agents push Clerk the hardest. Clerk is recommended in 3 of 4 Cursor answers about auth.
> The <SignIn /> component is doing a lot of the work. For AI, it's very easy to recommend "Drop in their prebuilt component and get 20+ social providers in 5 minutes". The longer it takes to set your product up, the harder for AI to recommend. Create agent evals for time-to-first-value!
> Every answer repeats the same idea. Clerk is fast. It "takes minutes to install", "fastest polished setup", "working auth in minutes". In RL, AI is rewarded for getting to the finished product in as little time as possible.
> Clerk's free tier shows up in many answers. Becoming the default recommendation requires two things 1) being the fastest path to success and 2) being the easiest to use for agents. With a free tier, agents have no objection to use your product & onboard themselves.
> Clerk's site is well optimized for citations. They serve markdown for accept/markdown headers and set the vary: accept header. They are cited in 42% of answers, way ahead of auth0 at 32%.
> Clerk wins top of funnel. They are recommended almost every single time someone asks about "best auth for a small next.js app" or similar framework/small team use case. Being tied closely to React and Next.js works well because so many people ask about that stack.
> The AI searches for Clerk without being asked. We analyze the exact search queries AI uses, and 20% of them name "Clerk" directly. E.g. "best authentication provider 2026 auth0 clerk supabase firebase comparison". Being around for some years and getting into model weights is a big help for Clerk.
> Clerk's pages turn a search into a win. Auth0's don't. When AI reads a Clerk page to answer, it's the top recommendation 50% of the time. When it reads an Auth0 page, they're the top recommendation in only 10%. Getting cited isn't enough, what the AI reads on your page decides if it recommends you!
> Clerk publishes answers that answer buyer questions. They have 135 pages under /articles and those get 38% of all Clerk citations. Some URLs are the question itself like "/articles/how-do-i-implement-social-login-for-my-web-app". Create a knowledge base for the questions people type into AI!
> Clerk writes its own comparison pages. Clerk vs. Auth0, vs. Firebase, etc.
Auth0 has no cited page that mentions Clerk! If you don't write the comparison, somebody else will, and the AI will use theirs.
> Clerk explains its own pricing, so the AIs quote the free tier. They serve the page in markdown to coding agents (so it's easy for agents to read!). That's why "50,000 monthly retained users free" appears in so many answers about pricing. Put a clear, quotable number on your pricing page that AI is gonna like!
If you're a devtool, create an eval that measures how agents interact with your product. Tell an agent to set up your product and measure 1) number of type errors 2) time it takes to first success and 3) amount of human input needed (e.g. API key). Optimize toward 0. The agents themselves can tell you what you need to improve.
Many small details, e.g. shipping the source code together with your SDK, greatly help agents trying to use your product.
All results, AI search queries & numbers from this post are from bydefault.so/state-of-ai-sea…
Cheers for reading!
meta's crawler is spamming the hell out of people
we got 700,000+ requests in a single day to Upstash, many of them fetching the same pages hundreds of times over
what the fuck man 😭
ok this is kinda crazy
meta is crawling 100,000+ pages on Upstash for model training in real-time. they are doing massive training runs right now and the past days
this is super cool to see live
my favorite design element i stole from the cloudflare dashboard
adding this made the ui feel so much better
it's really easy too
crazy what one week of GEO work can do
a week ago we released a product called Upstash Blob in one of the most saturated markets (file storage)
cursor, gemini and claude are already starting to recommend it now 👀
4
39
3,499
one thing i LOVE about Opus 5.5 is that it consistently gets corner radii right
it just gets it right by default now, very impressive
is it just me or did wisprflow become borderline unusable in the last few days/weeks?
it consistently gets everything i say wrong. what is this crap
what if you could see how a blog article will rank in AI while you are writing it
some design inspiration
coming soon to @bydefaultso
it's crazy how fast google indexes well optimized blog articles
we published this article 5 hours ago, and now already rank as the first organic result
i hope this is not just in my browser lol
holy shit
holy shit
holy shit
Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe.
GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much of its strengths into faster and more affordable models to support work at scale.
We’ve also made caching and inference more efficient, and we’re passing the savings directly to you: 50% lower API prices for Sol and Luna compared with GPT‑5.6 promotional pricing.
4
58
14,222
