I had the exact same experience. Claude is not my validation engine, I do not tell it when it has completed. I verify that in the code/project and end the session.
I also got this,
> A useful next habit is to end each correction with a concrete acceptance test, owner artifact, or stop condition: “write it into PROGRESS.md,” “make nix run .#bench fail until this is real,” “rerun this exact command,” or “do not proceed until these two choices are explicit.”
> You already do this well in the biggest penance sessions. Apply it to the smaller ones too.
Which I have found to be counterproductive in my personal work. Current models can generally infer acceptance tests of this level of granularity (not true for larger project-level prompts, but those don't produce good enough code for me yet -- even with specific acceptance criteria).
I also got penalized for using claude in read-only mode for the same validation reason?
> For read-only work, end with one of:
“turn the top finding into a PR-sized plan”
“mark these as accepted/rejected/deferred”
“write a cleanup checklist”
“give me the exact command I should run safely”
“stop, no action recommended”
No thanks, I'm literally just exploring the codebase. I don't want any of these.
It's a little sad to be honest, I would actually enjoy a product that helped me improve prompting + ai usage.
It's notable that they're so valuable because they feature 800Gbps of memory bandwidth. About twice what's available on the top end of M5, and exactly what makes llm inference fast.
But if SpaceX does anything _other than become the most valuable company in history by delivering at least two technologies predicated on the largest stock rise in history based on a single technology (LLMs)_, the gains from zero to now will have been privatized and the losses will be born by the public. Which is what everyone is thinking about.
see also: we're all developing on min spec Macbook Pros while being encouraged to burn > max spec Macbook Pros cost in tokens/month while still waiting for builds to compile.
Yeah, if you're arguing that "this, according to anthropic, existentially dangerous model has only had its safeguards partially circumvented so we shouldn't step in" ... it's hard for me to take you seriously?
Put another way, the thing we are all concerned with is the complete circumvention of safeguards that is normally possible with llms. If you _aren't_ arguing that this isn't possible, you're not engaging in discussing the the thing that is concerning to regulators or those discussing the regulation.
The existence of a jailbreak free llm in 2026 is extremely contentious to me. You can argue about the specifics of this exact jailbreak, but generally pliny and amazon both reported mythos jailbreaks in <7 days. It seems very reasonable to expect that a well funded state actor could achieve better results given significantly more funding, determination and most importantly unfettered access.
This is the rub, Bob would not be promoted if he consistently provided unreliable LLM output. In order to get promoted, Bob needs to learn the skills that get reliable output out of an LLM. These may not be the same skills that Alice learns, but if the argument is that Schwartz's LLM output is valuable -- why are we to assume Bob's path isn't towards Schwartz?
No he has clearly said there are differences. He has said that the points around what it may be used for are the same. HOWEVER, he has also stated that the definitions and enforcement are left to US law in the OAI contract. These were left to Anthropic in theirs.
One positive thing I will say about this administration is that they have really drawn into focus the difference between de jure and de facto law.
My hope is that this gets us some real concern for things that have been defended with de facto arguments (i.e. privacy) going forward.
edit: Anthropic argues that your Crayola analogy is fundamentally incorrect.
> Legally, a supply chain risk designation under 10 USC 3252 can only extend to the use of Claude as part of Department of War contracts—it cannot affect how contractors use Claude to serve other customers.
Isn't the point that they aren't entering into a contract with them, they are just ensuring that none of their still trusted suppliers repackage Anthropic without their knowledge?
I'm in a weird spot where I do agree with your assessment of the core claim. But putting that aside, in the world where the DoW's claim _is_ correct -- I think you don't have any choice other than to designate them a supply chain risk.
Disregarding who is right or wrong for a moment, if the DoW are right (which I'm not personally inclined to believe, but we're ignoring that for the moment) -- how else can they avoid secondhand Claude poisoning?
Supposing they really want to use their software for things disallowed by Claude's (now or future) ToS, it seems like designating it a supply chain risk is the only way they can ensure that their contractors don't include Claude (either indirectly as a wrapper or tertially through use of generated code etc)
I also got this,
> A useful next habit is to end each correction with a concrete acceptance test, owner artifact, or stop condition: “write it into PROGRESS.md,” “make nix run .#bench fail until this is real,” “rerun this exact command,” or “do not proceed until these two choices are explicit.”
> You already do this well in the biggest penance sessions. Apply it to the smaller ones too.
Which I have found to be counterproductive in my personal work. Current models can generally infer acceptance tests of this level of granularity (not true for larger project-level prompts, but those don't produce good enough code for me yet -- even with specific acceptance criteria).
I also got penalized for using claude in read-only mode for the same validation reason?
> For read-only work, end with one of:
No thanks, I'm literally just exploring the codebase. I don't want any of these.
It's a little sad to be honest, I would actually enjoy a product that helped me improve prompting + ai usage.