Hegseth Announces Grok Access to Classified Pentagon Networks

7thKeeper said:

Anthropic is out from what it seems. They refuse to remove guardrails from their model that Hegseth is insisting on.

And no matter what version (free/professional) you are using, if it's an LLM, you will never be able to get rid of hallucinations. It's unfortunately baked into how the models function and you can only try to mitigate it, but it will always be in here in LLMs.

Click to expand...

Seems like Anthropic is changing their tune pretty quickly


“Rather than being hard commitments, these are public goals that we will openly grade our progress towards,” the company said in its blog post.

The change comes a day after Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei a Friday deadline to roll back the company’s AI safeguards, or risk losing a $200 million Pentagon contract and being put on what is effectively a government blacklist.


With regards to the other aspect you mentioned, while the hallucinations you speak of are certainly pronounced in the free tiered versions, the higher tier versions and enterprise solutions are a different user experience. Where the free tiered ones on the web are basically just a glorified LLM/google search hybrid made for speed, the more robust offerings are more tailored to accuracy and depth.

So while one may be used to hopping on ChatGPT, Perplexity, or Claude and typing a question, and getting an answer back in 5 seconds. The more robust solutions are asking follow-up questions, and giving users the option to prioritize different aspects of what they're trying to do. For example, when utilizing the top tier version of Claude Code, it's not unusual to basically have an 1 hour+ discussion with it, and have it "think" for 5-10 minutes between answers and build a complex solution that can pass SOC2, HIPAA, and PCI audits, where as if you hop on the free version of Claude that's online, it'll spit out a quick and dirty html page with some vanilla JavaScript.


To describe the difference in the past, I've used the steak dinner analogy.

You can order a steak dinner at Cracker Barrel or Applebee's.
vs
You can order a steak dinner from Ruth's Chris or Fleming's

Both will involve sitting a table and having someone bringing you a steak to eat that's edible, but you'll be getting a very different product offering and experience at the latter.

That's not to say that you can't end up with a poorly cooked steak at Ruth's Chris, it obviously happens once in a while. But your chances of getting an overcooked piece of shoe leather, or one that's still cold in the middle (because they didn't thaw properly) when you ordered medium-rare is much less likely at the latter.

MarcusGregor said:

Related:

AIs can’t stop recommending nuclear strikes in war game simulations

Leading AIs from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95 per cent of cases

www.newscientist.com

"Would you like to play a game?"

Click to expand...

Some of the article appears to be behind the paywall...

I was able to see as far as this part:
Kenneth Payne at King’s College London set three leading large language models – GPT-5.2, Claude Sonnet 4 and Gemini 3 Flash – against each other in simulated war games. The scenarios involved intense international standoffs, including border disputes, competition for scarce resources and existential threats to regime survival.


Not to keep harping on the "version is everything here folks" talking points but.

Using several-versions old free tier models doesn't necessarily prove what would happen in a realistic application.

For example, they mentioned using the 1 year old free-tier Claude implementation (they're currently on Sonnet 4.6 for the free tier, and Opus 4.6 for premium use)

And there's a big difference between the two

GPQA Diamond is where the gap becomes dramatic. This benchmark measures PhD-level questions across physics, chemistry, and biology. Opus 4.6's 91.3% vs Sonnet 4.6's 74.1% represents a 17-point chasm — the single largest performance difference between the two models. If your work involves expert-level reasoning, Opus is in a different league.

Similar story with OpenAI. Their "deeper reasoning" is determined based on mode. Meaning, the answers/approaches you'd get from Sonnet 4.6/GPT5.2 are very different than what you'd get with Opus4.6/o1 Pro-Thinking Mode.


Basically, this guy's experiment is tantamount to someone reproducing an already known flaw from a several-versions old, un-patched release of Windows 10, and using that as an argument for why it's dangerous for a government entity to be using the latest & greatest version of Windows Server that's fully patched and up-to-date.

mindlight said:

Musk is a Trump supporter while much of the rest of Silicon Valley are just doing what they need to survive him. It seems that military decisions as to which platform is reliable and which not are being made on the basis of personal loyalty to Emperor Trump.

Personally I found Grok to be infected with Musk's atheism when I asked it questions about Bible dating for example.

The biggest question with this AI as with all of them in fact is to do with the tendency to hallucinate based on statistical models that do not map to reality. I would be interested in what legal constraints and what correspondence to truth tests will be in place for this system.

Also how will command and control hierarchies integrate with the workflows of decision making. AIs can identify targets but it will still need some kind of empirical test and a human decision to fire.

Click to expand...

And that leads us to the current conflict between Hegseth and Anthropic, the corporation operating the AI system called "Claude."

Hegseth is threatening to destroy Anthropic because they refuse to develop fully autonomous weapons that take humans out of the loop entirely and automate selecting and engaging targets. I suspect it's because he can't find enough soldiers who will reliably obey illegal orders.

Here is the full statement of Anthropic's CEO on the dispute:

Here is a detailed news story on the dispute.

ThatRobGuy said:

Seems like Anthropic is changing their tune pretty quickly

https://edition.cnn.com/2026/02/25/tech/anthropic-safety-policy-change

“Rather than being hard commitments, these are public goals that we will openly grade our progress towards,” the company said in its blog post.

The change comes a day after Defense Secretary Pete Hegseth gave Anthropic CEO Dario Amodei a Friday deadline to roll back the company’s AI safeguards, or risk losing a $200 million Pentagon contract and being put on what is effectively a government blacklist.

Click to expand...

No, that new policy does not indicate a reversal from Anthropic's resistance to creating an LLM that can target humans without human control.

RDKirk said:

And that leads us to the current conflict between Hegseth and Anthropic, the corporation operating the AI system called "Claude."

Hegseth is threatening to destroy Anthropic because they refuse to develop fully autonomous weapons that take humans out of the loop entirely and automate selecting and engaging targets.

Click to expand...

I normally don't say this, but destroy Anthropic? Don't threaten me with a good time, Sec. Pete.

RDKirk said:

I suspect it's because he can't find enough soldiers who will reliably obey illegal orders.

Click to expand...

This is entirely too plausible.

MarcusGregor said:

Related:

AIs can’t stop recommending nuclear strikes in war game simulations

Leading AIs from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95 per cent of cases

www.newscientist.com

"Would you like to play a game?"

Click to expand...

"A strange game. The only winning move is to wipe out mankind."

RDKirk said:

No, that new policy does not indicate a reversal from Anthropic's resistance to creating an LLM that can target humans without human control.

Click to expand...

If you read the fine print of the policy change, it does significantly weaken their prior promises.

Or, at the very least, it's a reversal on what their perceived "commitment" level was. They've long tried to brand themselves as the "ethical AI" company.




However, when you look at the details of this policy change, it's basically tantamount to saying "Yeah, as long as we're out in front, we'll stick with the commitment to safety, however, if other competitors start to surpass us and start making more money, we'll just match whatever their level of guardrails are"

Nithavela said:

"A strange game. The only winning move is to wipe out mankind."

Click to expand...

Where is HAL when you need it.