Leo Huang

AI capabilities researcher. Creator of claude-real-video — open-source tool that lets any LLM actually watch and understand videos.

[Leo Huang @LeoAidoAI]

https://t.co/aDZHG7dBSY
open source: https://t.co/85T6iOfPUV

[Leo Huang @LeoAidoAI]

Can Claude watch a video?

Not directly. It takes images and PDFs, not video files. ChatGPT and Gemini do.

So I turn the video into frames + a timestamped transcript first. Same question, Claude got it right too, and fixed 10 words the transcript misheard.

100s demo below

[Leo Huang @LeoAidoAI]

source: https://t.co/W8k6tEISPT

the filter is home-mixer/filters/oon_retweet_reply_filter.rs. the interaction weights are in home-mixer/params/param.rs. they said they'd publish diffs, so re-check before quoting numbers.

[Leo Huang @LeoAidoAI]

i read x's open-sourced ranking code. two things everyone repeats are not in it:

premium gives no ranking boost. no scoring path reads subscription status.

links are not penalized. open_link is +0.2.

what is in it: out-of-network replies get dropped before ranking.

[Leo Huang @LeoAidoAI]

@kig The catch I hit with this: written down rules only fire if the agent happens to read them that turn, which was maybe half the time for me. The ones that stuck were the ones I could turn into something that runs, a hook or a test. The rest stayed a hope.

[Leo Huang @LeoAidoAI]

@mjlbach Compile time is underrated for agent loops. The agent is bounded by how fast it gets a signal back, not by how fast it writes the code, so a 1s build changes what you can leave running overnight. Rust catches more up front but you pay for it every iteration.

[Leo Huang @LeoAidoAI]

@dzhng The verification bottleneck part matches what I see. What helped me wasn't better review, it was making the agent prove the change ran end to end, reading the record back after deploy instead of reading the diff. Design calls though, still no answer there.

[Leo Huang @LeoAidoAI]

the tool, if anyone wants to poke at it: https://t.co/85T6iOfPUV

[Leo Huang @LeoAidoAI]

mediapipe 1.0.1 doesn't fail with a python traceback on macos. it SIGABRTs the whole interpreter from C++: no exception, no log line, dead process.

hit this in my video tool's gesture tracking. pinned <1.0, all green.

for native deps, version pins are your only error handling.

[Leo Huang @LeoAidoAI]

@ClaudeDevs running claude code basically around the clock on a mac mini here and the mid-turn cpu spikes were real. deferring gc to idle is the same trick browsers pulled with requestIdleCallback years ago, funny how every runtime ends up relearning it

[Leo Huang @LeoAidoAI]

@tibo_maker every affiliate that ever moved anything for me was already a customer recommending it for free, the payout just made it official. cold-recruited ones mostly post once and disappear. so my answer would be boringly simple: look inside your user base first

[Leo Huang @LeoAidoAI]

@tibo_maker what got me is it wasn't a user choosing. the model was writing the integration and just reached for what it already knew best. whoever the coding model reaches for by default gets distribution for free.

[Leo Huang @LeoAidoAI]

@marclou zooming out saved me last month. daily numbers looked fine, the 5-week line had been sliding the whole time and i just never saw it. daily is mostly noise, the trend is the only part that tells you anything.

[Leo Huang @LeoAidoAI]

standard youtube advice: re-hook viewers every 60-90 seconds. i measured 3 MrBeast videos frame by frame. real number: a micro-stimulus every 17-22s, a scene change every ~100s. median shot: 1.5s. the famous rule is 3x too slow — it's the floor, not his machine.

[Leo Huang @LeoAidoAI]

@marclou shazam matches a fingerprint of the exact recording, so a hum has basically nothing to match against. melody search is a whole different model, google's hum-to-search does it and it's decent. funny that in 2026 they're still two separate products

[Leo Huang @LeoAidoAI]

@tibo_maker the what-to-make part is the right call. one thing i keep running into though: most viral-video analysis reads the transcript and the caption, not the actual frames, so the hook it describes is often not the hook that worked. if yours really looks at the video that is the moat

[Leo Huang @LeoAidoAI]

@tibo_maker "not afraid to ship" is the real part. they have no old workflow to protect and no reputation to lose yet. that fear resets for everyone every time the tools change, most people just refuse the reset

[Leo Huang @LeoAidoAI]

@marclou the number i'd want next to 149 is how many listings never sold. 2.5x on $2.5k mrr says buyers are paying for cash flow, not code — which is probably why most no-revenue side projects just sit there.

[Leo Huang @LeoAidoAI]

@tibo_maker still solid because it's research, not a growth hack. my first ~30 users came from one forum thread. by user 10 half of what i assumed was already wrong. the channel dried up around 40 — but by then i knew what to actually build.

[Leo Huang @LeoAidoAI]

a floating tooltip cost us 2 days of zero conversions.

it re-appeared every session (sessionStorage resets in in-app webviews) and sat on the booking button for 60s.

the tell: views normal, clicks zero, next button healthy. a dead button beside a live one = overlay, not demand.