←
phyrooo6,723 posts
phyrooo avatar

phyrooo

@phyrooo

Minimalism maximalist. $BTC $GRIN secp256k1: b0b00009763375d6800c7bc477431cfaaf032a065d301b22dac492728ea3df29

◷ Joined October 2009▧ 252 Photos and videos
phyrooo avatar
phyrooo@phyrooo·Nov 11

3/ Readme contains examples of some other flags/options. The thing is meant as a simple decision-model search over videos. It won't find all the things, but it's good enough for many simple cases. Tested only on linux.

phyrooo avatar
phyrooo@phyrooo·Nov 11

2/ The image below explains roughly how it works. Readme shows how search over Youtube videos or live Twitch streams (via yt-dlp) as well as your webcam video.

Post attachment
phyrooo avatar
phyrooo@phyrooo·Nov 11

1/ dvid is a simple grep for videos: ask a question and get the timestamps where it's true. dvid --every 5s "the audience is shown" talk.mp4 A bash script around ffmpeg + a hosted or local decision model with deem. Works on files, streams and webcams. github.com/phyro/dvid

phyrooo avatar
phyrooo@phyrooo·Nov 6

7/ Completely vibed and tested on a linux machine. A friend was able to make it run on Windows and I suspect it would work on a Mac too.

phyrooo avatar
phyrooo@phyrooo·Nov 6

6/ Design decisions were discussed with a good friend of mine and he told me (and I quote) "You are completely right." so we settled on this collectively. Readme.md and other docs files cover a lot more. If you have any questions, the best thing is to ask your LLM about deem.

phyrooo avatar
phyrooo@phyrooo·Nov 6

5/ Here's a decision tree as a deem program: it sorts log lines, and the trace shows how its questions are batched.

phyrooo avatar
phyrooo@phyrooo·Nov 6

4/ Every run can be traced and compared across models (deem trace compare), so you can check how close a local model gets to the hosted one. A run collects questions, gets answers and restarts the computation with answers as additional inputs until we have a pure jq computation.

phyrooo avatar
phyrooo@phyrooo·Nov 6

3/ A deem program can't read files or env vars, write files, reach the network or run commands. It sees its record, the clock and the model's answers; deem sends every question, so it can cap them. Computation is not bounded: run untrusted ones with a timeout and memory cap.

phyrooo avatar
phyrooo@phyrooo·Nov 6

2/ Beyond one-liners, you can write programs in jq (jaq) with one extra function: judge which calls the model. A prelude has wrappers around it. Why jq and not Python calling the API? I expect models to write these programs, and I don't want to read every one before running it.

phyrooo avatar
phyrooo@phyrooo·Nov 6

1/ Deem is a CLI tool that judges logs, tickets and screenshots, from the shell. Use hosted Jev or a local model (Kev, d3-mini) to keep your data local. Use local when possible. Installation instructions explain how to run both variants. Experimental. github.com/phyro/deem

phyrooo avatar
phyrooo@phyrooo·Nov 3

The idea is to make adding honest code easy and "dishonest" code harder. Much like the asymmetry in cryptography where it's (mathematically) hard for the attacker to do things. A simple digital signature from Anthropic over (prompt, response, model) would work too though.

Can we now make git commits come with a set of verifiable statements: [(model, prompt, response, proof),...] where prompt is a security audit for the code diff and the response being a "safety score/threshold" from that model? Might make libs/packages injecting exploits harder.

phyrooo avatar
phyrooo@phyrooo·Nov 3

If have you 1gb to spare consider cloning the repo (git lfs!) so this timestamp proof collection gets more redundant. git lfs install git clone github.com/phyro/ohara.git Make sure to check that the zip file in data/inputs/internet_archive/metadata is 800mb and not just a pointer.

phyrooo@phyrooo29 Dec 2025

Ohara is an open archive of verifiably timestamped video hashes. It includes an Internet Archive download wrapper that downloads IA identifier data and indicates which video files have timestamp proofs in its archive. github.com/phyro/ohara

phyrooo avatar
phyrooo@phyrooo·Oct 9

Either the ratio of bad people around Sam is unusually high and he's been unlucky (for years) to have been surrounded by them or... Sameone else is bad. The reason I never bought his product is bc it's a donation (a vote) towards these people handling the most powerful weapon.

Post attachment
phyrooo@phyrooo24 Jul 2023

Altman won't even open source things, why would he suddenly start giving free coins to everyone? A change of heart? Unlikely. Worldcoin is an attempt to buy your biometric data. It's just marketed as "prove you're a human and you'll get free coins because we care about you".

Jasmine Wang avatar
⟳ phyrooo retweeted
Jasmine Wang@j_asminewang·Oct 8

OpenAI fired me last week, along with two of my safety colleagues. I was given one reason: that I accessed an executive's email. I want to say this plainly, because too many of OpenAI’s history is smoke and mirrors when people disappear:

Two other safety researchers and I were fired from OpenAI last week. We wrote this letter to leadership. I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation.

Post attachment
Parimal avatar
⟳ phyrooo retweeted
Parimal@Fintech03·Oct 7

Brilliant take by Terence Tao on Open AI's recent math release. In 1993, Andrew Wiles announced a proof of Fermat's Last Theorem. Months later, experts reading it found an error. It took him another year, with Richard Taylor, to fix it. The proof became a proof because people had to read it line by line. The talks and workshops were part of how it got to be true. A proof now costs hours. Understanding it still costs years, and the experts who pay that cost never asked for the result.

Post attachment