ListNook

BorikGor Mar 30, 2026 +48

Oooh! I think I saw this one somewhere!

48

ThreeChonkyCats Mar 30, 2026 +17

"Net" and "Sky" part of it?

17

Lanster27 Mar 30, 2026 +5

Is it the one that have them all plugged into a computer simulation?

5

Tajetert Mar 30, 2026 +2

except everybody is French

2

StaticSystemShock Mar 31, 2026 +1

Baguettes become sentient and declare Ciabatta a mortal enemy. War emerges between France and Italy, pulling entire Europe into a first Bread War and eventually to a World Bread War 5 years later.

1

EasyRider_Suraj Mar 30, 2026 +9

TUN-TUN TUN TUN-TUN

9

the_walking_kiwi Mar 30, 2026 +13

The penguins better be watchful

13

Ferelwing Mar 30, 2026 +36

At some point I keep hoping that someone will listen to actual *AI scientists* who keep stating that LLM's are *not* ready for this kinda thing because they *cannot* actually do what hypers keep trying to claim, but never let reality get in the middle of *hype* I guess.. This will go *predictably* wrong. Gary Marcus will likely have to add *another* prediction to his list.

36

Eskipony Mar 30, 2026 +8

It depends on what you're using it for. In tech at least, AI is already embedded in many workflows, and it already works with the appropriate safeguards, and obviously with a human in the loop.

8

Ferelwing Mar 30, 2026 +2

Has it ever stopped "hallucinating"? No. You'll excuse me if I am not nor will I ever be "impressed" with it. But now, instead of insisting that software be perfect we're teaching people that "we're" the problem and if only we'd learn to "prompt" better we'd get "better" results. You'll excuse me if I don't find that innovative. Alas, the dumbing down of everything will continue...

2

Wise_Mongoose_3930 Mar 30, 2026 +6

The article mentions translation work as one of the ways France is using it. As someone who used to use google translate a lot at work, I can safely say that all the major LLMs do a better job already for my use cases. Is it as good as a human that is fluent in both languages and also a subject matter expert? Nope. But our budgets almost never allow for such a thing. There's a lot of snake oil being sold, and there's a lot of people over hyping it, but it already has plenty of real world use cases where it performs better than existing tools. It's just that most of them are boring.

6

Ferelwing Mar 30, 2026 +1

I agree, translation work is absolutely fine and definitely something that I think LLM's are actually pretty decent at. They're not perfect but even humans aren't perfect at that.

1

Eskipony Mar 30, 2026 +11

Because... its just way faster when used right, and if you have an easy way to check for correctness, its way better. At least where I work, the AI self corrects when it "hallucinates". Small scripts or data manipulation work takes mere minutes when it would likely take me hours or even days previously just to design a script to do so. Thats not to say using the top of the line models to just go through what you need to do.

11

IgnoranceIsTheEnemy Apr 1, 2026 +1

Have you checked the output by replicating work yourself? I’ve been using it in my business and we have caught a lot of “looks right but isn’t” nonsense.

1

Virillus Apr 1, 2026 +1

IMO AI is an incredible tool once you understand the constraints. Once our team understood how to limit it to small, discrete, tasks with a single responsibility it became a pretty meaningful accelerator. When you're using AI to construct a script for a single task with strict bounds, not only does the chance of a "hallucination" go way down, verification is extremely fast and collision with other systems is super unlikely. That being said, letting it loose on anything with significant complexity and you start carrying a huge amount of risk.

1

IgnoranceIsTheEnemy Apr 1, 2026 +1

I have a lot of colleagues that are being encouraged to buy into the hype by management and offload thinking to be more efficient. We are sleepwalking into a situation where 70% correct work that vibes with what people already want to see is “good enough”. I was amused by a client yesterday who ran a competitors work through an analysis tool that said 80% of their advice was AI generated. The conversation that followed was basically- “I’m paying you for your subject matter expertise. Not to replicate what we can do ourselves with AI”

1

Virillus Apr 1, 2026 +1

I just wish I understood why all the major LLMs are addicted to massive omnibus scripts. Beating it out of them is exhausting. No, I don't want a 3000 line script for a single task. No, I don't want to bundle 12 behaviours into a single massive controller script. No, I don't want a custom solution for literally every problem.

1

Eskipony Apr 1, 2026 +1

Depends on the task at hand. I work in tech so most of my work is coding related, it'll probably be different for non-tech roles. The top of the line models generally give me the output I need, but obviously you'll have to review the code afterwards, and allow the agentic AI to run tests. The only times where it gets confused is if you start doing too much at once or start loading too much unrelated context into it's context window. Considering how there's really not that many new ways of implementing your code in an existing codebase, what I do manually is pretty much what the AI will do.

1

Ferelwing Mar 30, 2026 -10

Again, LLM's are nothing more than probability scales. "Prompting" them is teaching people that *they* are the problem *not* the software. I am *not* a fan of that and *never* will be. As if IT coding isn't spaghetti enough now we're adding in stolen code and sub par "engines" designed to make it "easier" and pretending they "self-correct".

-10

TakeThreeFourFive Mar 30, 2026 +10

Nobody is telling you that you must be a fan. But to act like they can't make things easier or that they can't self-correct is just burying your head in the sand. People with decades of experience are seeing real value get delivered without a real loss in quality. That IS happening. Turns out probabilistic generation is good enough in very many cases.

10

Ferelwing Mar 30, 2026 -3

In some cases, there's enough data out there stating not in *most*.

-3

TakeThreeFourFive Mar 30, 2026 +9

There's just something remarkably weird about people who don't use these tools insisting to those who do that their experience must be *wrong* or that they are *lying*

9

AgentGorilla Mar 30, 2026 +1

Yup. Even if someone doesn’t think the models are quite good enough yet, there is clearly a trend of them improving extremely quickly such that they will be even more useful in 6 months or a year from now. Given how slowly governments move, even if they sign contracts now it will still take years to fully integrate LLMs into workflows. By that point the models will be even better

1

Fusken Mar 30, 2026 -1

I feel if you don’t embrace AI for most of your work , you’ll left behind quite quickly. Of course, output should always be checked, same as any other work I do.

-1

Eskipony Mar 30, 2026

Tbf, i just chalk it up to ignorance and inability to keep up with the rapid developments in AI. I used to be an AI hater as well until models started becoming good enough last year for me to be relatively hands off. Watching Opus tear through a really huge problem effortlessly is truly a sight to behold.

0

droans Mar 30, 2026

Know what it's trained on? Facebook, Twitter, Listnook, 4chan, and the like. Know what it's not trained on? Military operations manuals. Military strategy and training. Combat effectiveness. Half of its knowledge of nukes probably comes from COD.

0

Dispator Mar 30, 2026 +3

That's not true. They can train AI on private curated datasets.

3

droans Mar 30, 2026 +1

I know they can fine-tune. That's still not going to change the base knowledge. The Pentagon even said themselves they had a problem with all the models they tested recommending the use of nuclear weapons way too easily.

1

Dispator Mar 31, 2026 +2

No they can have completely seperate data and training design that does NOT use any of the chat-gpt stuff that your familiar with. But yeah AI is not good enough to do most things really effectively yet (or ever we will see). Military AI is not a conversation/research/lookup/general bot or anything like it. It's trained for very specific tasks with very specific data. It's output would be very specific to the given scenario and wouldn't rely on prompts/chat as much as hard parsed sanitized input parameters of the given situation in a formatted text/data file. It's also deterministic - meaning given the same inputs/files/"situation", it will give the same outputs/"recommendations".

2

Additional-Sky-8384 Mar 30, 2026

it’s just voluting through a ton of data and making predictions. the only unfortunate bit is that it’s hard for us to understand why certain decision was taken

0

Ferelwing Mar 30, 2026 +7

That's because of just how far removed the programmer has been removed from the machine level over the years... Before debuggers you had to come at coding from a different mind-set. There are so many different steps involved in coding that the vast majority of new coders *do not even fundamentally understand* that gets worse over time when you add in AI, the reliance on it isn't doing anyone any favors. In the end it will lead to people working in fields who fundamentally do *not* understand the software they are creating let alone what is happening with the machine or *why*. In fact, even before AI this was noticeable it will *now* get that much worse.

7

zapreon Mar 30, 2026 -6

Iran war is a pretty good use case for what AI can do, which is massively accelerate targeting of targets in a foreign country.

-6

Ferelwing Mar 30, 2026 +5

By hitting non-military targets?

5

zapreon Mar 30, 2026 -5

By massively increasing the tempo of military strikes, especially on dynamic targets such as TELs. No more days and thousands of people to process targets, which makes such dynamism impossible

-5

Ferelwing Mar 30, 2026 +5

Again, and also making massive mistakes hitting non-military targets.

5

zapreon Mar 30, 2026 -2

What is the evidence for operationally significant mis-targeting? A few bombs here and there targeting non-military targets isn't a particularly big deal operationally. And in any near-peer conflict, that is of far greater importance than concerns of hitting civilians

-2

Asleep-Ad1182 Mar 31, 2026 +2

Mistral AI is awfully bad

2

Gunsensual Mar 30, 2026 +5

Good choice. It doesn't need a network to run and can be tuned and scaled locally. Most competing products are at the merci of US meddling.

5

tupe12 Mar 31, 2026 +1

I’ll see yall in the robot prison camps

1

Informal_Witness3869 Mar 31, 2026 +1

Did I hear someone say Butlerian Yeehawd?

1

Random-Cpl Mar 30, 2026 -3

I thought the French were smarter than this

-3

ouath Mar 30, 2026 +11

It is not about smarter or not, it is survival: if you can't control what other armies do with AI, you are bound to use it to not fall behind and become weaker for your principals. They might start to talk between each others when they all/most think they went too far with it like everything in history (gas, clusters, nuclear, bacteriologic ...)

11

Ferelwing Mar 30, 2026 +5

FOMO, the act of following a hype chain because you're afraid of missing out... LLM's cannot do what is being hyped but unfortunately *no one is listening* to the researchers who have come out and admitted it, they're listening to the hypers who are milking it for cash. Gary Marcus has been right on all of his predictions thus far and keeps explaining precisely what is *wrong* with LLM's, sadly instead of being listened to people keep believing the grifters hyping it.

5

Random-Cpl Mar 30, 2026 +2

It is not smart to deploy brand new, incredibly flawed, and potentially dangerous technology across all your military systems because you’re afraid of being left out

2

Wise_Mongoose_3930 Mar 30, 2026 +4

If you read the article, most of what's being done with it is boring clerical work, like turning scanned documents into searchable text, translation work, etc. This is a boring dystopia, not skynet.

4

generallyblind Mar 30, 2026

you thought right. Mistral AI will be deployed in lieu of copilot for translating and summarizing notes

0

ronweasleisourking Mar 30, 2026

*bah gawd! That's the terminator music!*

0

Apprehensive-View583 Mar 30, 2026 -2

they have generational gap behind SOTA, it’s prob worse than Chinese open source model, might as, well use Chinese open source model…

-2

France Deploys Mistral AI Across Military to Accelerate Operational Decision-Making.

💬 Send a Message

48 Comments