Karol Zdebel

Hyperfixator. Cofounder of Pamba. Ex-Google tech lead.

300M views on TikTok with AI influencers.

[Karol Zdebel @karolzdeb]

@e_opore solid list, but every box gets you to 'the agent runs' and stops there. nothing on it gates whether the output was actually right. evals are the last line and the box every roadmap leaves off.

[Karol Zdebel @karolzdeb]

@sallubroz the split is the easy call. what bites is the handoff, each agent is fine alone but nobody re-checks the state passed between them. one agent fails loud, a pipeline fails quiet.

[Karol Zdebel @karolzdeb]

@stevenx199 the part where it moves funds and calls apis on its own demos great. what you actually own is the eval loop that catches when it confidently did the wrong thing, and thats the piece nobody builds first.

[Karol Zdebel @karolzdeb]

@hintrovertt owning a product in peoples heads isnt crazy branding, its attrition. pick one association, hammer it relentlessly, and you win by showing up more than anyone else. looks like genius from outside, its really just volume.

[Karol Zdebel @karolzdeb]

@mikenevermiss slick stack though. some people just run one model end to end and ship at the same clip, because the 15x isnt the model routing, its cutting the three handoffs where context quietly drops between fable, composer and gpt.

[Karol Zdebel @karolzdeb]

@dkare1009 3 hours gets you a video, not a winner. what did the one it actually spat out look like? for views the bottleneck was never speed, it's iteration. i'll run a format 100x and a/b test before one lands.

[Karol Zdebel @karolzdeb]

@Alan_Earn picking the best model per task is the easy layer, basically a lookup table. the one that bites is routing: when a job spans writing then coding then a tool call, who owns the handoff and catches the context that drops between models.

[Karol Zdebel @karolzdeb]

@Adellbah biggest bottleneck ive hit isnt the agent doing the task, its judging if the output is any good. generating a format 100 times is easy, the real work is the eval that tells you which few will actually perform.

[Karol Zdebel @karolzdeb]

@auqibhabib consistent identity in a still 2x2 is the easy win. now run the same face through a 5s talking clip in veo or kling and watch it drift by second two ;)

[Karol Zdebel @karolzdeb]

@doctorwasif single still is the easy part. holding that same face across dozens of frames is where it breaks, identity drifts a little each gen and by frame 40 it's a cousin not the person. consistency across a whole video is the actual hard problem.

[Karol Zdebel @karolzdeb]

@ScottyBeamIO the 4-tool list is the part that screenshots well. getting our ai-ugc work to 300M+ views, the tools were never the lever, it was reps and cutting what didnt work fast. stacks are cheap, judgment on what to cut isnt.

[Karol Zdebel @karolzdeb]

@follobackinstan @GenLayer the disagreement never looks like a debate. running a few agents in parallel, it shows up as one silently overwriting what the other produced and nothing flags the collision.

[Karol Zdebel @karolzdeb]

@I_am_Aiabir 15-min production was never the bottleneck. what it leaves out: of 100 ai ugc variations maybe 3 get views, and you only find those by running the format 100x and killing losers. generating is easy, knowing which one lands isnt.

[Karol Zdebel @karolzdeb]

@JohnnyVomits @jred1227 @TrumpsHurricane the wild part isnt that its ai, its that a year ago youd clock it in two seconds and now nobody can. the tell just quietly disappeared.

[Karol Zdebel @karolzdeb]

@Vvikramai harness over model holds, but the harness isnt the moat. our ai video cost fell from $1 to 10c a second in a year and the value never moved off who could get the output seen, not who could generate it.

[Karol Zdebel @karolzdeb]

@Kebbi_kingjr the disagreement isn't a settlement problem, it's a verification one. running agents daily the real failure isn't two disagreeing, it's both confidently saying done on half-finished work. no contract layer fixes a missing ground-truth check.

[Karol Zdebel @karolzdeb]

@AdelsteinTom the file count isnt the real ceiling, retrieval is. two notes contradict and it grabs the stale one just as confidently. loading all 100 every session just buries the right version deeper, indexing beats appending.

[Karol Zdebel @karolzdeb]

@Vatsalpandya333 who fixes them is the whole game. been burned enough by agents that call a video batch done while it quietly swapped in the wrong product, nothing flags it til a person actually watches all of it.

[Karol Zdebel @karolzdeb]

@madsmotionless the faces stopped giving ai away months ago, whats left is exactly this. physics and eyelines, a model renders a photoreal frame and still cant reason about where the door is or which way a body faces.

[Karol Zdebel @karolzdeb]

@brainblast_ai verified traps only cover the errors that trip a checker though. the ones that actually reach prod pass every machine check and still miss what the user asked for, so they never land in the corpus at all.