Thomas Dietterich

The frustrating thing about this is, agencies have used automated targeting filters and machine learning for MASINT and scenario planning for many years. LLMs aren’t necessarily a huge analysis leap except now they are making command decisions. That is INSANE and Anthropic is right to balk at it.

Alphabet/Waymo vs Amazon/Zoox roboclot:

20th St by Lexington, Mission, San Francisco

If their respective remote human assistants could communicate directly this would be faster and surer/safer.

OP: tiktok.justjimmynajera

Done!

Gather 'round Bluesky, while I tell the hoary tale of epidemic vs. endemic.

How can it be that pandemic interventions that were vitally important in 2020 are marginally effective in 2025?

Science will give us the answers!

Follow me...

1/

Indirectly, it mostly goes to support graduate students

Yes. National Science Foundation (US). Funds primarily basic research in math, physics, engineering, social sciences (at least pre-Trump), education, and biology.

Is your scenario the ICU (or similar)?

Cancer screening from medical imaging is similar to target recognition. It would seem to be straightforward to give a “probability of tumor” for a detected region, but I’m sure the defn of “region” is itself uncertain, and other contextual information is important.

My only experience with uncertainty in medical settings is for lab tests and imaging. I find it very frustrating that my blood tests are returned as point estimates without error bars. At least the “normal” value is given as a range.

Yes, without question

I’m curious to learn more about your experiences. I’ve mostly thought about this in self-driving vehicles, where caution in the face of uncertainty is easily implemented.

Now we know this was all just an Altman-orchestrated move to steal the federal business from Anthropic. It had no technical or strategic basis.

The lesson of Ukraine is that we need cheap, disposable UAVs, not $30 million units. And we aren’t going to put expensive GPUs on such platforms

These systems will need to operate in comms-denied environments.

The fact that Secretary Hegseth made this an issue shows either that he doesn't understand the technology or that this is just a power play not relevant to national defense.

There are many scientists and engineers within the Army, Air Force, and Naval research labs who understand all of this very well. Likewise in the relevant commercial companies.

These systems also do not provide quantified uncertainty (particularly for novel situations), which is essential for making high-stakes decisions.

LLM-based technology is good for many things, but it is not a reliable tool for autonomous weaponry. LLMs require large GPU computers to operate; you can't put them on standard military platforms, and certainly not on drones.

Will he be getting a salary? Anyone can achieve salary-free freedom...

I'm confused by this thread. If library 1 allows Google to train on book A, what claim does library 2 (which also owns A) have?

Automated cars must make decisions under uncertainty all the time: fog, occlusion, blind corners, etc. When uncertainty is high, they behave more conservatively. So, *if* they can recognize the sign, they should respond appropriately. I don't think theory of mind is needed for these cases.

Not exactly a good way to win friends!

Re: mapping errors, I would love to know how map information is fused with perceptual signals to detect errors. Under high uncertainty, the car should plan an alternate route rather than trusting either the computer vision or the map.

Of course, we have all seen the examples of adversarially modified road signs. I don’t know the current sota in defending against such attacks.

Reading road signs should be easy—provided that uncertainty is properly quantified. I’ve driven cars that read speed limit signs, but they made many mistakes. Ex: In Italy, there are speed limits that only apply when it is raining. My rental car missed that detail.

This conversation has convinced me that theory of mind is important for autonomous driving. I clearly underestimated its value, particularly in novel situations. Thank you to the participants, esp @sifu.tweety.fish @sven.blacksky.app @aniccia.bsky.social and @caravellin.bsky.social

I totally agree

Sure, but the situation is fundamentally uncertain. What is the best action to take under uncertainty?

There are plenty of rude drivers who don’t negotiate. Waymo could learn “winning” tactics rather than doing full negotiations. And then, of course, everyone will say Waymo’s are rude and Waymo could just answer with safety statistics

Or you will need billions of hours of non-simulated data.