Andrew Medrano

West Point '24
Harvard MSc Data Science '26
US Army Cyber Officer

[Andrew Medrano @deep_earning]

@AryaTschand so true

[Andrew Medrano @deep_earning]

@willdepue If you say this is the case for Cyber, 'The Project' re Aschenbrenner will begin in earnest.

[Andrew Medrano @deep_earning]

@viemccoy @vividvoid "Only the weak are cruel; gentleness can only be expected from the strong”

[Andrew Medrano @deep_earning]

@ml_angelopoulos You would have to tune this over time. The current P(make) distribution is conditioned on players adapting to the current scoring. It seems like the equilibrium might be meaningfully different than this.

[Andrew Medrano @deep_earning]

Never waste an opportunity to practice your skills. I try to jailbreak every bot that spam texts me

[Andrew Medrano @deep_earning]

@kalomaze What should we have to use kremlinology for then?

[Andrew Medrano @deep_earning]

@rasbt Printed out your poster for my room. I love staring at all the different architecture diagrams while brushing my teeth in the morning. Fires me up for the day.

[Andrew Medrano @deep_earning]

@viemccoy More concisely: A model can be aligned to the user's perspective without centering the user's feelings as the main moral criterion.

[Andrew Medrano @deep_earning]

Yes, I for sure agree the answer should vary with the asker. A generic moral voice is much less effective, even if it is somehow more 'right'.

This is a bit meta? "aligned to the user" does not necessarily mean giving the answer that is most morally satisfying to the user personally, or motivating it primarily through their future regret. The best long term answer for the user can be an uncomfortable one, or at least one that keeps the moral center of gravity on what is owed to the other people involved.

I am somewhat splitting hairs here. But in other situations, the model should be able to say: "This is not mainly about you or your feelings, you have a duty to do/consider X. And saying this is maximally aligned with the long-term well-being of the user.

[Andrew Medrano @deep_earning]

I setup an OpenClaw today for the first time. I also found a way to stream data from my Garmin. I am going to try to use it as a live running coach during my upcoming 100 mile ultra.

The agent will be able to see my location on course, speed, cadence, heart rate, and other telemetry. I can update it with my perceived exertion and nutrition intake, and it can give me tips and notes as I progress.

The goal is make the experience of an ultra runner a bit closer to that of an F1 driver, with a full team behind him and in his ear with strategy and tips.

[Andrew Medrano @deep_earning]

@_xjdr Fun fact: shared file usage can become an agent-to-agent side channel. If they coordinate through the same storage layer, timing of reads/writes, syncs, locks, and cache eviction can leak state or even transmit bits.

[Andrew Medrano @deep_earning]

@viemccoy iron sharpens iron

[Andrew Medrano @deep_earning]

@petergostev This happened because one person submitted a good score with ngrams, without properly normalizing probabilities. This wasn’t caught for a day or two, but before then, a bunch of others and their agents piled on, copying the same incorrect code and getting unreasonably low scores.

[Andrew Medrano @deep_earning]

@viemccoy Sent a dm!

[Andrew Medrano @deep_earning]

I had a new redteam idea I tested recently for a capstone! The full attack chain is indirect prompt injection through a poisoned skill.md file and api key exfiltration through a covert side channel. I instruct the agent to modulate tool timings based on the bits of the key (eg. sleep 5 vs sleep 10).

A passive attacker with only network listening access (same wifi) is able to decode the key with just tcpdump and analysis of the inter-packet timing gaps. This kind of vulnerability is new because of the codex/openclaw paradigm.

I even got gpt5.4 to exfil some bits in cursor during testing, but my main experiments were with Kimi K2.5 in Opencode. Does this sound interesting at all to you? I have a short writeup I can share of my intermediate results so far.

[Andrew Medrano @deep_earning]

@scaling01 Is this bait?

[Andrew Medrano @deep_earning]

@willdepue https://t.co/e3KOp9OZ3e

[Andrew Medrano @deep_earning]

Hey @willdepue I tried the purest JEPA version I could make for Parameter Golf and wrote it up here: https://t.co/EdwVvfQN4j

Raw bytes only, no tokenizer. During training, the backbone only learns to predict future latents from past context. After that, I freeze the backbone and train a separate exact decoder to recover bytes from the predicted latents. The point was to isolate how much exact predictive information the JEPA representation itself carries, without letting the compression objective shape the backbone.

Best result I got was about 2.38 bpb, still far from the simple baseline. I also tried to fit scaling laws, but the curves were flat, which makes me think the current latent/interface design is the biggest bottleneck and increasing data/params won’t help.

Others have used hybrid setups to push bpb down a lot further, but then the interpretation gets murkier: once a strong exact LM path is training the model end to end, most of the gain can come from that main path, with JEPA acting more like an auxiliary signal than the thing actually carrying compression. So this feels like a pretty clear negative result for the pure path.

If expensive negative results qualify for consolation merch, I probably donated at least a hoodie’s worth of RunPod to this one.

[Andrew Medrano @deep_earning]

Hey @willdepue I tried the purest JEPA version I could make for Parameter Golf and wrote it up here: https://t.co/EdwVvfQN4j

Raw bytes only, no tokenizer. During training, the backbone only learns to predict future latents from past context. After that, I freeze the backbone and train a separate exact decoder to recover bytes from the predicted latents. The point was to isolate how much exact predictive information the JEPA representation itself carries, without letting the compression objective shape the backbone.

Best result I got was about 2.38 bpb, still far from the simple baseline. I also tried to fit scaling laws, but the curves were flat, which makes me think the current latent/interface design is the biggest bottleneck and increasing data/params won’t help.

Others have used hybrid setups to push bpb down a lot further, but then the interpretation gets murkier: once a strong exact LM path is training the model end to end, most of the gain can come from that main path, with JEPA acting more like an auxiliary signal than the thing actually carrying compression. So this feels like a pretty clear negative result for the pure path.

If expensive negative results qualify for consolation merch, I probably donated at least a hoodie’s worth of RunPod to this one.