Marty H

Dad life + Data life | AI/ML | Teaching machines to learn while kids teach me patience

[Marty H @oldhero51]

Hand-labeling video for object detection is miserable.

Waldo automates the whole pipeline: SAM 3 labels your footage from a text prompt, YOLO26 trains on it, and you get a serving endpoint at the end.

Self-hosted. Mac, Linux, Windows.

https://t.co/c4rYpdbVWq

[Marty H @oldhero51]

@Houseofyogi @grok verify stats please

[Marty H @oldhero51]

@elonmusk The fact that this is so easily debunked is exactly the problem. Iran Air 655 is well-documented history, not opinion, not interpretation. A 10-second search confirms it.
This has @grok sitting right there. Use it. Flag obvious historical misinformation before the algorithm rewards it with a million impressions. The correction in the replies will always reach fewer people than the original post. That’s by design, and it needs to change.

[Andrej Karpathy @karpathy]

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

[Marty H @oldhero51]

@tadgh_dc Great joke.

[AI Breakfast @AiBreakfast]

So Claude got Maduro and ChatGPT got the Ayatollah

Sounds like a new benchmark

[Marty H @oldhero51]

@NYCMayor Haha…this will age poorly

[Marty H @oldhero51]

This resonates deeply. I've had the same phase shift experience. But I keep circling back to a question you didn't quite address: what does the learning path look like now?

You mention discrimination (reading) and generation (writing) are different capabilities, and that you're noticing atrophy in the latter. But you built your discrimination ability through years of generation. You can spot the 1000-line bloat and suggest the 100-line solution because you've written both versions manually, felt the pain of maintaining the bloated one, developed the taste.

When I watch newer engineers use Claude Code, they often can't see the slop. Not because they're less capable, but because they haven't built the pattern library yet. They accept the overcomplicated abstraction because they don't know the simpler one exists. They can't push back on wrong assumptions because they don't recognize them as wrong.

So I'm genuinely curious about your take on the pedagogical paradox here:

1. If coding drudgery is where fundamentals get internalized (not just understood intellectually), how do we create that same depth when the drudgery is automated away?

2. Your "generalists vs specialists" question is fascinating, but does this assume a baseline competence that might not form naturally anymore?

3. Is there a version of this where LLM-native developers actually develop better intuition because they can explore vastly more design spaces faster? Or is that optimistic cope?

I don't think this is a "kids these days" concern. It's a genuine structural question about how expertise forms when the struggle that builds it gets smoothed over.

What's your mental model for the next generation of engineers emerging from this environment?

[Marty H @oldhero51]

@svpino Do you run any heavy processing on it? Or are you remoting to a server? Wondering if you ever notice throttling from heat dissipation.

[Marty H @oldhero51]

@CodeByPoonam Whelp looks like I need to start practicing supervising firefighters. How should I get started? @grok

[Marty H @oldhero51]

@elonmusk Why haven’t you launched an autonomous EV racing league yet?

Imagine racing but with NO DRIVERS. Just pure AI pushing the limits of what’s possible. Teams competing on algorithms, sensor tech, and split-second decision making at 200+ mph.

This would:

- Accelerate FSD development faster than any road testing
- Create insane entertainment value
- Push battery/motor tech to the extreme
- Generate massive data for Tesla’s neural nets
- Attract top AI talent globally

Racing has always driven automotive innovation. Time to do it again, but this time let the machines compete.

Who wouldn’t watch AI-powered Teslas battling it out on track? Make it happen! 🏁🤖

#AutonomousRacing #FSD #TeslaRacing

[Marty H @oldhero51]

13/15
THE TRUMP FACTOR
Our model identified "Presidential Authorization" as critical first step.
Trump's statement: Need to see if "people come to their senses" Earlier: "Iran cannot have a nuclear weapon"
Timing of ceasefire announcement matched our "immediate action required" prediction perfectly.

[Marty H @oldhero51]

12/15
LESSONS FOR CRISIS MANAGEMENT:

Time pressure is decisive (nuclear timeline)
Signals matter (advance warnings = diplomatic openness)
Regional mediators crucial (Gulf states)
Credible threats + off-ramps = success
Quantitative models can predict complex outcomes!