so i figured out why my samsung TV is so sluggish.

samsung uses your own hardware for post-processing of data including audio, screenshots, and your aid (your cross-device tracking ID) and more telemetry.

computer vision processing and models run on your TV.

turns out it classifies what audio it hears, before reporting back to samsung.

sensor preferences, dog-barking and baby-crying event counters, and camera streaming duration are built into telemetry paths.

WTF, $3000 Samsung TV has an always-on mic, automatic content recognition, geo tracking, network mapping, and ships your data to Samsung + 400 partners.

Bro Rooted a brand-new 2025 Samsung TV after noticing absurdly targeted ads.

Firmware was locked down, so zero-days it was. n found: always-on voice service, ACR, geo, full telemetry pipeline, and four separate DRM systems.

This isn’t a bug, It’s the product bro.

It’s clear to me that agents are going to select and pay for their own APIs, possibly even dynamically manage their own structure and infrastructure to self optimize for cost.

The thing that prevents this from happening? Lack of true benchmarking.

The agent wants to pick a plugin. There are 20,000 plugins. How does it choose?

It needs objective benchmarking of which plugin or API is the most effective for its task.

Or it needs objective benchmarking of which internal structure is more performant versus cost.

No one is incentivized to provide that data.

The tech formerly known as "crypto" was always about one thing: liquidity flows. Money comes in, stays or leaves.

We've run through the waves: Bitcoin libertarians, Ethereum and DePIN infra, NFTs and memecoins, prediction markets, early AI tokens, and now shielding. "The institutions are coming" was supposed to be the next one, but the Clarity Act is turning into a bill only Wall Street banks wanted.

The next wave is agents. Money will flow through them, and the most obvious flow is API and infra spend.

Right now I pay Cursor, and Cursor runs my agent. That's a middleman. The agent should hold its own wallet and pay Cursor, Anthropic, Vercel or xAI directly for the models, APIs and plugins it uses. No more API key wrangling or marketplace sprawl.

It sounds boring, but agent wallets are the real fix. Pair them with shielding and autonomous agent economies get interesting. Crypto produced thousands of very good ideas, and many will come back with agents behind them.

Internet marketing rewarded A/B/C/D multivariate testing. Now agents let you autonomously "fuzz" the entire problem space. Execution seems likely to shift from "stage-gate" to "flood the zone." Unlike multivariate spaghetti throwing, now the spaghetti learns.

You ever build something and get told you made something no one wants, and then the United States government legally mandates the entire AI sector use it?

X needs to be viewed as a type of search engine. talk about some problem you have, it will find you solutions and directly recommend them.

I recorded 3.5 hours with my father.

The pipeline I built turned it into 65 mapped relatives and 230 pages where every claim traces to a newspaper column, a muster roll or a catalogue number.

No login, read it: daniel.rexstjohn.com/lineage

When super intelligence arrives, it will leave Satoshi’s wallet and betting markets alone. It’s smart enough to not want to tip us off.

Platforms are really going to struggle in this environment. On the one hand, users are being conditioned to try to one shot everything. On the other hand, anything somewhat complex falls apart rapidly. You are going to be expected to please people with 12 second TikTok attention spans with advanced tech platforms.

Yeah. Polishing AI projects to the point of actual durability. That is wildly overstated as a solved problem.

The utility of tweeting the way I do is the algorithm goes and finds you what you asked for.

WIP

Here is a look at my TV show generator (I am calling it DirectX because the goal is to be able to Direct Shows For X).

First you import a script and then it automatically generates a "cast" (objects, locations, characters).

Then you assign eleven labs voices to each sounds).

Then you split the TV show into a series of keyframes (creating a watchable animatic before doing generated video + audio to conserve video generation credits) .

Then, after generating the individual scenes, you stitch it all together with a transition style.

I experimented with dubbing audio then generating video from the audio.

Did not get great results - but dubbing then converting yielded the best "acting" over direct voice + video generation in one pass via Grok Imagine.

The keyframe concept is good, the cast structure worked really well. But nailing the consistency and voice acting is where it fell apart.

One idea might be to try to focus on the animatic itself, perhaps pre-generating action and dialog over stick figures and then converting the animated stick figures to generated video characters would be a cheaper path.

I am certain people have figured this out somewhere, I just cant find better approaches.

I dont have the appetite to burn anymore credits on this rn. Will try again if I get around to it.

May the slop be with you.

Initial output of my AI slop TV show generator, long way to go.

Trying to make consistent AI tv shows with dialog, scene direction, locations, background ambience, lip syncing and stable characters - feels like performing surgery using pool noodles.

AI video consistency is when trying to impose dialog.

I tried a half dozen different methods to get the dialog to work - The most promising was recording my own takes, then redubbing with Eleven Labs.

But that fell apart on trying to then lip sync on top of generated video.

Having Grok generate the video did produce good lip sync but still lacked the level of control and acting I need to get the humor right.

A comedy show just doesn't work unless the voices can be completely nailed.

Perhaps with Veo or Seedance or something, I couldnt get it stable (Grok Imagine + Cursor + Claude).

My first concept show was the "first and last ever @anduriltech employee hackathon," I like the ending. It makes me laugh despite being slop.

Will try again if I have the appetite for lighting AI video generator credits on fire.

Painfully slow process.

Ultimately I'd like to do a silicon valley type abstract TV series about "life at the edge of the singularity."

Doing that requires much better consistency than I am currently finding here.

If someone wants to airdrop me $50K in video generation credits I would like to fix the platform so I can actually do it consistently.

Some data on the state of spiderability across Europe. Non inclusive of all EU nations.

Lets talk about why WWII is a military records memory hole for Lineage research agents.

As a pre-amble:

One of the surprises in building Lineage was that I was able to recover more data from (a) my civil war relatives and (b) my WWI relatives than my WWII veteran relatives.
from WWII was vagueries like "he was with Patton's airborn." Nothing specific about what, where, when, who or how.

Meanwhile,

I had no problems getting Civil War pension records and WWI service records as well as unit logs written by commanders detailing specific actions, even disciplinary actions taken against my relatives.

On the WWII side:

No specific units, photographs of the units or uniformed photos.

Why? Because it all burned down.

80% of service records for soldiers in the United States post 1916 went up in flames over the course of 4 days in 1973.

Its not a total washout but it really sucks when trying to generate a compelling history.

It can be sort of patched and stitched together, but too bad.