Pinned Tweet
Ever read an explanation three times and still couldn’t picture it?
I added diagram generation to canvas. Turn a note into a diagram, then ask follow-up questions about it.
here’s how it works.
what I found working is posting similar content in more languages - there is good correlation between them.
advantage is that you test out different hooks and content 3x faster.
if something works on the us account, post on pl etc etc.
10
X has made some nice changes to how the algo treats long form videos
We run a few theme pages and post our YouTube videos on here too
And they no longer die out after ~36 hours. Some videos are still getting views after 10+ days
Great reach if you post good content, I guess what’s left is X making it worth your while financially
3
48
2,786
I have 3 accounts that post slideshows everyday.
1000 views every time.
How did you guys scale past that?
16
btw, i have been doing a launch video for my app. Used this and also connected elevenlabs - for sfx generation.
i was quite surprised with the sound effects it generated. It was just back and forth with claude
Today I'm launching earshot v1
Your coding agent can now make its own sound effects! You can just prompt and earshot will render the sound in code on your laptop
This works really well for games, animations, videos and a lot more!
npx skills add Leonxlnx/earshot
sound on :)
1
30
need a comparison with @reasonmachines
what is the pitch in few bullet points?
why i joined dusk bot, and what we're building...
there are dozens of Grok Bot alternatives today, and almost all of them run on the same recipe:
> plug into the subs and harnesses you already pay for
> wrap it in a simple UI
> add always-on features and more control
i've tested pretty much all of them with +800hrs in Hermes, Grok Bot, OpenBot and the rest
and they're all missing the same thing: real, integrated business applications
my job here is to bring that to dusk bot, so your bots run whole departments:
> SEO and AEO on autopilot
> content pipelines that grow every day
> automations for X, Insta, Tiktok, Youtube...
> lead gen for your agency
> ads generated and deployed autonomously
> the day-to-day of running your business
and much more...
we're building this layer with partners to always bring the cheapest options to grow an agent-first business with dusk
because right now most personal agents end up as an email filter, a calendar manager or a morning brief
we're going to fix that
now, here's exactly how we're building it:
dusk started with OpenMausBot, an Apache 2.0 licensed open-source foundation that we're building extensively on top of
we've been transparent about that from day one, and you can explore our code here: github.com/DuskBotAI/duskbot
our code is source-available and free for personal use, hope that clears up any confusion
it was the right call and i'd have done the exact same... no point spending months rebuilding what dozens of projects already built
what we're adding on top:
> a fast, reliable, beautiful mobile app to reach your bots from anywhere
> better performance, connectivity and proactivity
> a mode for heavy coding projects
> and A LOT more
but the part i care about most...
we're pouring a decade of business knowledge, decision making and operations into your setups
to make dusk bot the best co-founder you can have, by always deploying more use cases for our users
get it for free at dusk.bot
1
34
if you're building an app that reads pdfs, this will save you 3 weeks of testing...
we ran 10+ ocr models on real student files: scans, handwriting, latex homework, a 167-page book. here's what actually matters:
> most pdfs already have text inside. read it first. on our homework file that took 1.6 s vs 13 s for the full parser
> but check it before you trust it. latex math fonts can come out as garbage: ∫ turns into "Z" & accented letters fall apart
> decide per page, not per file. one upload can be half typed, half scanned
> for clean printed scans, free old tesseract is enough. 0.6% wrong characters on printed books, ~0.5 s a page
> ai vision models are great at handwriting but silently drop things. in one run haiku returned a 50-row table with 10 rows & said nothing
> so keep tables, printed math & diagrams away from them, & count lines against the page to catch what's missing
> a small model does the job. haiku came out at ~$0.4-1.4 per 1,000 pages
> printed formulas on scans are still the hard part. tesseract got 0 of 20 fully right
> show the text you have right away & upgrade pages in the background. nobody should wait on ocr
and the dumbest one: half our upload time was the file just waiting in a queue. time every step
day 8 of building an app that helps you learn faster
beta counter: 23
my quizzes used to forget your answers. the new ones will keep track
pick a few notes, take the test, and once you master a concept its flashcards unlock
sketched with a friend
day 7 of building an app that helps you learn faster
beta counter: 23
a quiz that remembers what you know
pick a few notes and quiz yourself
a little black button keeps your progress on the canvas
master a concept and its flashcards unlock
rough wireframe, not live yet
day 6 of building an app that helps you learn faster
beta counter: 22
2
4
171
day 5 of building an app that helps you learn faster
onboarded two people
fixed the bug we found, live the same evening
one slideshow hit 1k views
beta counter: 22
Andrew Pulc retweeted
you shouldn't have to trust an ai answer blind...
so widemind now shows its work: what it searched, which pages it read, and a link to check each one yourself
still in testing, soon live
beta counter: 21
day 4 of building in public. do you ever check your ai's sources?











