DannyBee

So the agent is going to put less into context when it uses sed to see 15 lines of a file than using read and putting the entire file into context.

That is my best charitable guess at what they are hoping to achieve.

Of course, there is an obvious set of solutions for this problem that don't involve pushing the agent to use bash.

Just a few days ago they turned on an experiment the forces claude code to use bash over standard tools in auto mode.

From the system prompt directly, new as of August 18th:

  Do your work through the Bash tool wherever it can accomplish the job: read files with cat, head, or sed -n, search with grep and find, and make file changes with sed, heredocs, or short scripts, rather than using the dedicated Read, Edit, or Write tools. Fall back to a dedicated tool only when Bash genuinely cannot do the job.

I'm sure they'll argue they are trying to make it use less context tokens to do things, but if this is the best they could think of, ....

This is of course, also not documented anywhere, as is typical for anthropic, you just have to guess whether you are going crazy or if they changed stuff seriously on you under the covers.

This was the last straw for me. Their harness (models are fine) was already falling well behind the other one i use (OMP) in the past 6 months in usability/etc, and they are the only ones who don't allow me to use other harnesses with my subscription.

So I've now stopped using claude code entirely. Unless something changes, i'll drop my max plan when it expires next month.

See page 29

I'm not sure how to explain all this without writing a 70 page dissertation on HN, and it's probably not worth it :)

Overall - this is a wildly complicated area. To give you an idea how complicated: Ignoring state law, transmitting gambling information for sports events over the wire is a federal crime. See 18 U.S.C. § 1084(a), which makes it a crime for a person “engaged in the business of betting or wagering” knowingly to use an interstate or foreign wire facility to transmit bets/wagers or information assisting bets/wagers “on any sporting event or contest.”

(It's legal if you are transmitting from a jurisdiction where it's legal to a jurisdiction where it's legal).

This has been upheld repeatedly for sporting events.

New york can, and did, include a claim to enjoing them from violating this act, which has absolutely no pre-emption issue because it's not a state law.

The case you cited is going to end up in the supreme court, where it will be a toss up. (in previous supreme courts, it would be a non-starter and the third circuit would have been summarily reversed)

The "emergency powers" they speak of are 7 U.S.C. § 12a(9), and they are quite specific.

It gives them the authority to direct a registered entity to do a few specific things. None of those things are relevant to here. It's stuff like emergency margin requirements, position limits, etc. Not "violate state law". It gives them no power to enable a registered entity to violate a TRO, or anything like that. Such a power would have to come through pre-emption.

The CEA gives them zero authority to preempt state law directly, and any pre-emption would have to be argued to already have occurred under the Commodity Exchange Act. They'll argue it occurs because of their order, but it actually doesn't meet the requirements to do that, so then they'll argue the CEA preempts state law.

As you may imagine, this has been argued about before, for a very very very long time.

Gambling is core state police power, and has been found so many times. As such, presumptions against pre-emption would apply, etc. Even in the current court that ignores precedent, using an esoteric made-for-specific-situations emergency power statute like this one would to preempt new york/etc (this is not the only case) law would run clearly afoul of the so-called major questions doctrine.

Lastly, the current CEA regulations actually ban event contracts that are unlawful under state law (17 CFR 40.11):

https://www.law.cornell.edu/cfr/text/17/40.11

Prohibition. A registered entity shall not list for trading or accept for clearing on or through the registered entity any of the following: ...

1. (1) An agreement, contract, transaction, or swap ... that involves, relates to, or references terrorism, assassination, war, gaming, or an activity that is unlawful under any State or Federal law;

So trying to pre-empt state law when the existing regulations clearly don't allow event contracts that are disallowed under state law is ... not likely to succeed.

Also note that New York has claimed a violation of the wire act in there, and in particular 18 U.S.C. § 1084(a). This is a federal statute that makes it illegal to transmit sports betting information over the wire (it's okay if it's from a jurisdiction where it's legal to a jurisdiction where it's legal). They have asked the court to enjoin them from violating this. This claim is here because it avoids all the pre-emption issues - it's a federal statute. So New York is also hedging their bets on the state preemption issue.

All that said, there is also a CFTC-designated contract market that Kalshi operates, and that they could likely exercise significantly more power over, and New York can order them around less on. But that is likely to end up in the supreme court, and harder to predict. Any other court the answer would be clear - congress doesn't have the authority to regulate purely intra-state gambling, etc.

If you read the complaint, the prayer for relief is quite clear that they only are trying to stop them from operating in new york, deliberately offering gambling to new yorkers, etc.

There is no relief requested nationwide.

https://ag.ny.gov/sites/default/files/court-filings/new-york...

This is the complaint.

If you go to page 29 you'll see what they requested.

The claim they are trying to prevent them from offering all event contracts nationwide is simply false. The closest anywhere is a claim to enjoin them from violating some federal criminal statutes that they would not be violating if they were not operating in new york illegally (IE do not stop them from operating nationwide).

You can also see their is no specific number on the damages. In fact, the only specific number is the request for Kalshi to pay $2000 in costs to the state of NY.

The CFTC is, understandably, relying on people not bothering to read it and so has put out an "alternate set of facts".

Business & Professions Code § 5412 is the current controlling statute, and that is a codification of what the courts have required.

That requirement is not just a california constitutional problem, but a federal one.

So even if you modified the california constitution, you'd still have the federal problem.

The federal problem is much less than the state one (SCOTUS has found the floor here to be quite low), but it's still bad enough that feds require compensation as well. Just less compensation.

For example, 23 U.S.C. § 131(g) conditions federal highway funds on paying just compensation for removals along Interstate and federal-aid primary highways — 10% of apportionment forfeited otherwise.

I looked into this quite heavily, and had long discussions with legislators, because there was a huge ugly billboard 150 feet from my house in San Jose :)

They have not gone extinct.

It's not a government problem - it's a constitutional one - Banning existing billboards is considered both a first amendment problem, and a takings problem.

Doing it without compensation is a constitutional violation, and even if california modified it's constitution, it would be a federal one. Less of one for sure, but still an issue.

The current law (Business & Professions Code § 5412) is a codification of what the court requires.

FWIW - Lots of people seem to believe the "in any other business", but it turns out there are tons and tons of places it is equally untrue, even where there is near zero time or cost barrier to entry.

So it's more "what's special about corn". It is also fairly hilarious to claim the parent is denying basic economics and then bring up corn as an example of having successfully managed economics. If the scales were not being thumbed, and "basic economics" were in play, corn would be in very very bad shape.

In the case of DRAM, there is an incredibly long history of these gloom/glut cycles, and they have stayed roughly the same timeframes (~3 years) since the 1990's.

Almost all the ones who have survived this long are either in the same kind of boat as corn - protected in various forms from the downside - or don't increase production and get caught out until they are absoultely forced.

The very temporarily increased profit is not worth going bankrupt for - they make more money long term by being very cautious and know this.

There are a near infinite number of economic studies you could look at (and several sibling comments cite some) - DRAM manufactuers don't chase the price and probably couldn't anymore if they want to.

None of this denies basic economic theory, of course, since economic theory is not exactly "rigorous", even to the degree it could be (IE even the parts that are pure analysis of data rarely reproduce!).

While you can generally restrict new ones, you can't eliminate existing ones without basically paying out the lost revenue+etc to whoever is running it.

They only seem to care now because it affects them.

But Jeff was responsible for a lot more of what is actually used today than Demis.

Demis is responsible for a lot more of the hype though ;)

I think saying Jeff's contributions were a long time ago must represent some kind of lack of understanding of Jeff's recent contributions.

Jeff has still been focused more on infrastructure, and that is just more hidden most of the time.

I would say,if i was forced to pick someone whose vision to follow, it would definitely be Jeff and not Demis, even today.

I will state a fairly blunt position: Unless literally nobody thought or tried at all here, i would give it a 0% chance of meeting the bar of criminal negligence.

The rest is a distinction without a difference.

As for what you want them to do - i don't agree the investigation should cause them to feel pain - that's not a good goal for investigations, and definitely not one we should want, because it essentially presumes they did somethign wrong in the first place. A bad outcome does not mean a broken process. All processes have error bars. You can desire the error bars to be smaller, and try to back that up with criminal penalties, but an expectation that error bars will be 0 makes no sense.

You can do absolutely everything right and still have people die - star trek was not wrong in that regard. Punishing that will not fix this inconvenient reality, which is why we generally don't punish it. This is also why we distinguish between inherently dangerous activities and not, for example.

As for the laws, sure, i think it's totally reasonable to explore whether you want the law to be different, but again, i totally disagree with your second part.

A successful attack does not imply anything is actually wrong with criminal law, or should be changed. The question is more of what error bars you want on the activity and where what they did falls - inside or outside those error bars.

Making mistakes of reasoned judgement are basically never criminal negligence.

In every state i'm aware of, it would require total disregard for safety in the case of a huge and obvious danger.

It would also have to cause injury or death.

The bar for criminal negligence is pretty high.

As i mentioned elsewhere, the standard is basically "total disregard for safety in the face of an obvious and huge risk that resulted in injury or death". I don't think anything we are talking about here comes close to these criteria.

Negligence is not a crime, it's civil liability.

Gross negligence (reckless disregard for human life) is often a crime, and often there are crimes related to it (reckless driving, etc). It also does not require intent to injure, so it could be committed by, say, an operator by operating an autonomous vehicle knowing it was unsafe and could harm people. So it usually requires knowledge but not specific intent. Again, crimes like this are state specific, and sometimes even municipality specific so it's tricky to give an exact result without pinning it to a state.

However, for example, all states where autonomous vehicles are operating have statutes explicitly defining civil and criminal liability right now, so it doesn't get into the more general legal question of AI.

The same thing is already starting to happen with AI agents in general, it's just not there yet.

As i mentioned elsewhere, criminal law is often reactive, not proactive. We usually do not make things crimes until after someone has already been hurt, and society gets really upset about it. As i also mentioned elsewhere, attempts to be proactive have also rarely worked out effectively, so it's sort of lose-lose in that sense. But it usually does not take anywhere near as long as people think for law to catch up.

Let's assume the case none of them do, since it sounds like you are asking about "what would happen in the case this isn't specifically answered by a statute".

In that case, the short answer is:

Criminal liability - you could only really charge crimes that don't require specific intent. Reckless driving is an example. You could charge the company since they are the operator and the car is simply an instrumentality. In the end though, there just isn't a lot of people here with legally culpable conduct.

Civil liability - the company pretty clearly because civil liability often does not require the same kind of intent crimes do.

This is, of course, why states where autonomous vehicles operate have autonomous vehicle statutes :)

As a general rule, criminal law mirrors what society overall wants to decide is culpable/not, and the lag time isn't as bad as most people often think. That doesn't mean nobody ever gets hurt or dead without someone being as culpable as society wants, they do, and it often leads to a law with a name - megan's law, etc.

Criminal law is mostly reactive though, not proactive, and to be honest, proactive attempts don't have a high hit rate.

Even though agents are not agents in the principal/agent legal sense (because agents have to be human), for the purposes of criminality, it does not matter.

Agents do not act autonomously (and every court to ever consider it has agreed), and therefore they would simply be considered an instrumentality of the crime.

So that part does not need a new law.

The real blocker is often that a lot of the crimes you could charge here require specific intent. Because the agent is just an instrumentality, it does not have separate intent (and can't be part of a conspiracy), so it's the intent of hte person using the agent that would matter. Without whatever intent the crime requires, they haven't committed a crime.

There are not a lot of non-intent crimes in this area, and this is on purpose. Otherwise you could get charged with a crime for say, running nmap and having it accidentally shutdown something important or killed a person or whatever because someone hooked it up to a TCP port.

Not quite.

Agents in the principal/agent sense have to be human.

However, every court to have ever considered it have held the human/company driving the agent responsible under vicarious liability/negligence/etc principles.

The only real defense that folks have tried is to claim the agent acted "autonomously", which no court has bought so far.

It is totally true that they don't think like humans, but this is mostly irrelevant.

The token outputs will change as a result of this particular input, and will be closer to the tokens in training data where people felt hurried or rushed or like their job was on the line.

That doesn't mean the LLM feels at all, but it's definitely going to push the output towards output that came from/was trained on people who were in that state, because the input will push it much closer to that latent space as it starts predicting.

As such, what you are saying is one of those rejoinders that is basically pedantic and wrong.

It is true they do not think, act, or feel like humans. But that doesn't mean it won't output text that looks like hurried or scared humans. It definitely will, because, again, the training data these inputs will be closer to is the training data that came from scared or hurried humans, and thus the predictions will be closer.

So either you don't think this will happen, which would mean you don't understand how the models work (or at least, you aren't giving any sense you do), or you do think this will happen but want to pointlessly argue that this isn't "human feeling", which is true but totally irrelevant to what words it will predict and therefore the actions it will perform.

Either way, i'd downvote you.

It would be more accurate to say the word predictions the model makes based on the input text will likely be closer to the ones that were made from the training data where people felt like their job was on the line than the ones that were made from the training data where people felt otherwise.

So while the model does not feel, it's predictions are definitely going to change as a result of this input.

I was literally told in one of the comments i made about how difficult a lot of the changes were, for very little practical gain:

"the new API WILL require you to rethink your code

the new API WILL be a bit harder to use for Rust newcomers, as they really need to think about lifetimes, mutable aliasing and so on"

etc

Their argument, of course, was that it was "technically better" anyway, as if that is the bar by which you should make things super painful for your customers and newcomers.

ESP non-rust works pretty darn well.

But ESP-rust feels like it's just a mess run by folks who simply don't value their customers time.

I either use STM32 or NRF exclusively now, and could not be happier.