Computer scientist building the infra AI needs to scale around search engines, databases, and distributed systems.
Earth
Joined August 2012
- Tweets11.9K
- Following389
- Followers786
- Likes63.6K
Your email provider can refuse to send any of your app's emails, password resets included, while too many of them bounce.
For example, Resend can pause sending from any account whose bounce rate is above 4%. Every email that your app sends to a made-up test address or a mistyped sign-up address can count as one of those bounces.
Before you release:
→ Point your automated tests at Resend's test addresses, such as [email protected].
→ When Resend's webhook reports a bounce from a user's address, ask that user to correct their address the next time they open the app.
→ Sign up with [email protected] on a preview build, then check that the app shows the prompt to correct the address.
1
14
Resend advises you to test with its resend.dev addresses, because emails to fake addresses can damage your domain's reputation: resend.com/docs/dashboard/em…
4
Once business customers ask for invoices with their company's name and tax ID on them, which of these will you send those invoices from?
→ Stripe Invoicing
→ QuickBooks
→ Xero
→ Something else
1
10
Stripe prints each tax ID that you add to a customer in the header of that customer's invoice PDFs: docs.stripe.com/billing/cust…
2
Customers with Outlook and Hotmail addresses may not receive your app's sign-in emails when your domain has no DMARC record, however few emails you send.
For example, the email expert Al Iverson shows that Outlook can reject mail from his low-volume test domains, which have no DMARC record. Microsoft, however, publishes a DMARC requirement only for senders of more than 5,000 emails a day to Outlook and Hotmail addresses.
Before your next customer signs up with a Hotmail address:
→ Where your sending domain has no TXT record at _dmarc, add one with the value v=DMARC1; p=none.
→ Sign up for your app with an Outlook address that you own, then check that the app's sign-in email is in that inbox.
1
25
Bento reports that Outlook rejects these emails only some of the time, so one sign-in email that reaches the inbox doesn't prove that the next one will: bentonow.com/posts/outlook-d…
5
YOU CAN PAY MORE FOR EACH TASK WHEN YOU PICK THE AI MODEL WITH HALF THE TOKEN PRICE.
Anthropic charges twice as much for each input and output token of Claude Fable 5.2 as for Opus 5.5, yet it writes in its launch post about a benchmark of large data-collection tasks: 'At Medium effort, Fable 5.2 matches Opus 5.5's top score for about 40% of the cost per task.'
Before you choose the model for your app's AI feature, send each model that you're considering the same set of test requests that match what your customers ask, then compare what you pay for each answer that you accept.
Will you choose your app's model by its price per token, or by what you pay for each answer that you accept?
23
Delali retweeted
What do Rerankers and Chess have in common? ♟️
I am SUPER EXCITED to publish the 150th episode of the Weaviate Podcast featuring Ghita Houir Alami (@ghita__ha), Co-Founder of ZeroEntropy and now leading Model Research at Notion! 🔥
This is one of my favorite deep dive podcasts into Search! 🔎
We explored how pairwise preferences can be turned into useful pointwise scores with ZeroEntropy's zELO method! Like chess Elo, a sparse sample of pairwise battles between documents is enough to fit an absolute score through a Thurstone model, and that signal is then distilled into a pointwise reranker. The result... 🥁 ... the state-of-the-art zerank-2 cross encoder! 🔥
The podcast covers a lot of interesting topics in Search! I loved learning about Reranking with Jev / Decision Models. 🛠️
Ghita explained how many teams want cross encoders to prune and classify rather than just order results. Calibrated relevance scores can support searching with a Dynamic K, thresholding, and classification. 🧠
We also discussed the tradeoffs between pointwise, pairwise, and listwise ranking. ZeroEntropy's pairwise reranker was slightly more accurate, but distilling it into a pointwise model came close to matching it at much lower inference cost. ⚡️
We also covered infrastructure for rerankers, agentic search and reranking, instruction-following rerankers, metadata in training rerankers, and the relationship between RSI and SLMs! 📜
This was a super fun conversation, and I hope you find it useful! 💚
YouTube: youtube.com/watch?v=mQgYPtct…
Spotify: spotifycreators-web.app.link…
Fable 5.2 is even cheaper. Wow. Anthropic is cooking!
Claude Fable 5.2 is available today on all paid plans, on the Claude Platform, and on all major cloud platforms.
It’s $8 per million input tokens and $40 per million output tokens, down from $10 and $50.
Read more: anthropic.com/claude-fable-a…
29
Do you build your app in an editor with models from several AI companies, such as Cursor, or in an agent from one AI company, such as Claude Code or Codex?
ALT A dark square card shows the sentence: If you write your code with Claude or GPT, build in Claude Code or Codex, because Cursor charges those models' full API price from the first request. The words from 'Cursor charges' to 'the first request' are in blue.
5
2
64
When customers connect their AI agents to your app through an MCP server, which of these will you use to sign those agents in?
→ Clerk
→ Stytch
→ Auth0
→ Something else
27
Someone else's AI agent can sign in to your live app with a password that you set for testing.
WorkOS says that an AI agent signed in to one of its tools with a credential for local development. So before you release, check that the live app refuses every test password and key.
2
3
64
That tool's API had no rate limit, so the agent pulled records from the tool in bulk before a WorkOS employee spotted the unusually high usage: workos.com/security/bulletin…
6
If your app's image settings allow every domain, anyone can make it resize images from other websites at your expense.
For example, Vercel takes each resize in a Next.js app out of the Pro plan's $20 monthly credit and then charges from $0.05 for every 1,000 more. On the free Hobby plan, the app's new images stop loading once the month's 5,000 resizes run out.
Before you deploy again:
→ In the image settings, list only the addresses that the app's images come from, such as your storage bucket.
→ Open yourapp dot com /_next/image?w=640&q=75&url= followed by the address of a picture on another website, and check that the page shows an error.
1
30
The Next.js documentation warns that a broad address in those settings can let attackers resize images that you never meant to serve: nextjs.org/docs/app/api-refe…
7
Which of these will you ask your coding agent to use for the phone version of your app, once customers want to install it on their phones?
→ Expo
→ Flutter
→ A progressive web app
→ Something else
1
1
65
A progressive web app is a website that customers can install on their phones and open like any other app: developer.mozilla.org/en-US/…
5
Your email or marketing tool holds every detail that you send it about your customers, so an attacker who breaks into it can take the details that none of your emails use.
For example, the toilet paper company Who Gives A Crap told customers on 7 November that someone had reached their names, email addresses, phone numbers, mailing addresses, and the total that each of them had spent, through one of its marketing software providers.
To leave an attacker less to take:
→ Send the tool only the details that your emails use, such as each customer's email address and first name.
→ In the tool's contact list, delete every field that no email uses.
1
24
Who Gives A Crap says that it can't yet tell whether the attacker copied that information: abc.net.au/news/2026-11-08/w…
11