Delali @delali

Building AI-native Cloud for agents and humans. Also building: - https://t.co/vyJ60BU0wk - https://t.co/uu9UuQuu4W sondelali.com Earth Joined August 2012

Tweets

9K
Followers

533
Following

235
Likes

59K

Delali @delali

2 hours ago

How many lines of the last pull request your agent opened did you read before you approved it? One number.

0 0 0 13 0

View Details

Delali @delali

5 hours ago

Vibe coding is leverage if you already know what good looks like, because you can tell slop from a version you would deploy.

1 0 2 28 0

View Details

Delali @delali

6 hours ago

@hello_code_ makes sense

0 0 0 4 0

View Details

Delali @delali

7 hours ago

Your vibe-coded app has to send its first password reset from a brand new domain. Which one do you use? - Resend - Twilio - Something else? Tell us

1 0 1 57 0

View Details

Ask your agent about something your documents do not cover, then read what comes back. Your code calls the model whatever your search returns, so the model answers from its own training when the search finds nothing. Handle that case yourself. Where the search returns nothing, tell the user you have nothing on it and skip the model call. Then count how many questions in your logs reached the model with nothing from the search behind them. I discussed how, when the search step fails, the model appears to fabricate information and how to evaluate this step in your own documents: sondelali.com/blog/ai-halluc…

1 0 1 31 0

View Details

Delali @delali

8 hours ago

@ryanvogel Fine, but maybe we should get better at finding unnecessary tests than banning the agents from writing them outright. I personally follow the TDD principle. It ensures that the tests are made to assert the intended spec instead of making the code work against it.

0 0 0 256 0

View Details

Delali @delali

10 hours ago

@sl1ma4 Is it writing code? Otherwise, Sonnet might not be a good comparison against GPT 5.6 Sol

0 0 0 14 0

View Details

Delali @delali

10 hours ago

@nicetrygrover That's probably true. They did similar when they bumped Opus 4.7 to 4.8. But the leap is higher for Fable 5.1. Using it would tell us if the benchmark numbers are worth it.

0 0 0 26 0

View Details

Delali @delali

11 hours ago

This is what we have been waiting for. Now i am not going to trust the benchmarks. I'm going to test it out myself.

Claude @claudeai

11 hours ago

We’re introducing Claude Fable 5.1 and Claude Mythos 5.1. They're the world’s most advanced models for coding and knowledge work.

2K 6K 54K 11.4M 8K

1 0 2 62 0

View Details

Delali @delali

10 hours ago

Socket on the compromise, the packages, and the preinstall hook: socket.dev/blog/popular-n…

0 0 0 11 0

View Details

Delali @delali

11 hours ago

On 4 August somebody published poisoned versions of keyv, cacheable, flat-cache, and other packages in those namespaces, which people download tens of millions of times a week. Socket's writeup says that all of the malicious behaviour was in an added preinstall hook, so each package worked normally after installation while the machine was already compromised. It took cloud keys, Kubernetes service account tokens, and npm tokens off those machines. Turn ignore-scripts on in your npm configuration, so that npm installs the files and runs none of the scripts a package declares. Better still, use pnpm, which runs a dependency's build scripts only for the packages you have listed as reviewed.

1 0 3 30 0

View Details

Delali @delali

11 hours ago

Did @claudeai limits just reset, or are they luring us not to cancel? 🤩

0 0 0 36 0

View Details

Delali @delali

11 hours ago

If this is true, I don't see why we shouldn't be getting reduced costs. Or am I overthinking?

Claude @claudeai

11 hours ago

Cache reads with Fable 5.1 cost 75% less than Fable 5’s. This reduces the cost of the model in practice by around 25% for typical workloads, and up to 45% for highly agentic ones.

37 90 2K 523K 148

1 0 1 48 0

View Details

Delali @delali

12 hours ago

@EncryptoKnights @isamirwankhede @X 👏👏

0 0 1 7 0

View Details

Delali @delali

12 hours ago

@major101x Ah right. That makes sense.

0 0 1 8 0

View Details

Delali @delali

23 hours ago

Where do you store your users' files and media assets? - S3 - GCS - Cloudflare R2 - Something else? Tell us

3 0 3 181 0

View Details

Delali @delali

12 hours ago

@johnternus Hello John

0 0 0 16 0

View Details

Delali @delali

12 hours ago

When your agent opens a pull request, your CI runs the code inside it with the keys your workflows can read. That can be dangerous since without oversight, your AI can hallucinate anything and tamper with your workflow. For sensitive actions, move the deployment keys into a GitHub environment and name yourself as a required reviewer, so that a job reads them only after you approve the run. The catch is that no deploy goes out until you approve it, including the ones you start yourself.

0 0 1 40 0

View Details

Delali @delali

13 hours ago

@hello_code_ True. You cannot beat the speed of AI. For me the one thing I completely rely on the AI for is deep research. Doing the kind of research the model does would take me weeks.