Bryan Sanchez

AI researcher building tools for self-learners

Building

[Bryan Sanchez @bs_c0de]

@firecrawl Crazy good numbers. Congrats on the launch!

[Bryan Sanchez @bs_c0de]

Building Bleppi has taught me that AI tools have made things faster to build, but distribution is still difficult.

I wrote about my attempts at distributing Bleppi and what I learned from this experience: https://t.co/aifkCb9QJr

[DAN KOE @thedankoe]

https://t.co/OPGS7HUKaG

[James Bermudo @jcebermudo]

Introducing Kapi UI, a visual commenting tool for your UI that automatically connects to your coding agent.

Click, comment, and send to your agent.

Works with Claude Code and Codex.

Free and open source.

[Bryan Sanchez @bs_c0de]

Massive fumble from Anthropic.

[Om Patel @om_patel5]

CLAUDE HAS A SERIOUS PRIVACY PROBLEM RIGHT NOW, A HUGE NUMBER OF SHARED CONVERSATIONS ARE PUBLICLY INDEXED ON GOOGLE FOR ANYONE TO FIND

when you use claude's share feature it makes a public link. it turns out those links got indexed by search engines, so "share with anyone who has the link" actually became "anyone can find this by searching"

and people are pulling up genuinely alarming stuff:

> api keys, credentials and crypto wallets
> personal resumes with real names, addresses and phone numbers
> a lawyer working through a potential ethics violation
> an engineers internal company project details
> what appear to be peoples social security numbers
> and a crazy number of deeply personal chats people never imagined another human would read

anthropic never added a noindex tag to those shared pages, so search engines were free to crawl and list them

one line of code would have prevented the whole thing

this already happened to chatgpt about a year ago, same exact issue, but openai patched it fast

if you have EVER hit share on a claude chat, assume it could be public

go to settings > privacy > your data > shared chats > manage

delete anything you dont want the whole internet to see, especially anything personal or financial

[Bryan Sanchez @bs_c0de]

Spending the past few weeks building Bleppi has taught me a lot:
- Keep tight feedback loops to improve your app
- The hard part is no longer building, but reaching the people you're building for:

More on my blog: https://t.co/XKv2O611m8

[dax @thdxr]

used a trick @jlongster came up with

agents can control browsers but you can also ask it to record network requests into a HAR file

then it can derive a client for any website which is more efficient than browser controlling it every time

made it build a quick uber eats cli

[Viv @Vtrivedy10]

Sound on 🔊
Chess Tutor -> Eval Tutor! shoutout Fable, now my eval review agent can interactively speak & step through how agents succeed/fail during Eval rollouts

ran this on a few Terminal Bench tasks using @harborframework

processed each rollout with a judge that had privileged access to the verifier and scoring and prompted it to identify key moments, mistakes, inefficiencies and organize a timeline to step through

design flow inspired by @chesscom -> their game review bot is great

fun one, voice was generated locally using chatterbox with @lmstudio

at LangChain we're thinking a bunch about mining traces at scale while making it easy for humans to understand what happened so we can transparently & continuously improve agents over time

come work with us if this interests you :)

[Bryan Sanchez @bs_c0de]

The app is completely free for the time being! It's still under active development, so run into any issues or have any features you'd like to add, don't hesitate to reach out.

Hope this helps and good luck with your exams!

[Bryan Sanchez @bs_c0de]

[repost]

#rpmtwt hi!

My girlfriend and her friends are reviewing for the boards and I noticed that encoding review drills in Google Forms and retaking them there was a popular method of review. However, this takes a lot of time. So, I built Bleppi: https://t.co/PDKc2SU2i9

[Bryan Sanchez @bs_c0de]

Instead of having to encode via a lot of pointing, clicking, and typing, Bleppi gives you a textbox where you can directly type in your questions in a simple format. Once you're done, the app will parse these into quizzes which you can take after.

[Justin Skycak @justinskycak]

I wrote the above entirely manually. Go ahead and @pangram my post above.

I have a gigantic corpus of previous writing to draw from in my blog, books, and pre-2026 tweets, all manually written.

Here's the full blog post that the long tweet above came from: https://t.co/UyMpPrmgQE

(This year I have started using AI to extract themes / takes from my corpus of previous writing for short tweets here.)

[Catalin Voss @CatalinVoss]

We got a lot of people asking about how we built Ello’s tutor to achieve sub second latency. The architecture behind this was a real technical challenge.

Our tutor steers the UX in real time and makes complex decisions on the fly, so it operates more like a coding agent. However, the standard tool-use loop doesn’t work here.

"Generate a call, execute it, feed the result back, generate again" stacks up to 3-4 seconds of dead air between turns. Fine for a coding agent. Fatal for a conversation with a first grader.

Here's a deeper look at how we actually built it.

[Bryan Sanchez @bs_c0de]

@Bagz_Tech I think one thing we need to be cautious about is making sure the AI doesn't hallucinate information. If it ends up teaching the wrong info, then the student has to unlearn it and learn the accurate info, which could interfere with each other.

[Bryan Sanchez @bs_c0de]

@justinskycak Would love to see some concrete examples of applying these to different fields.

[Justin Skycak @justinskycak]

The physics of learning, and why almost no one uses it. 👇

[Bryan Sanchez @bs_c0de]

@lionel_mora This looks awesome!