superfrank

I’m curious why you have Hermes using Pi and not just Hermes writing the code itself or spinning up subagents to do that?

Despite the name, the label isn’t is only added if it meets about 5 other criteria, files and folders can be excluded meaning any PR with a change in them isn’t eligible for the label, and the PRs still need to pass all the non-AI quality checks.

Even if it’s not perfect, I think it’s the right idea. There’s some category of changes that are low risk and simple enough that a human reviewer isn’t needed and if you can identify those then you’re going to be able to save human review for the things that really need it.

I’ve not worked at Microsoft so I can’t speak for them, but I’ve been in tech for the last 15 years with the last 5 being in manager/director roles at at least 1 company that is a household name.

What you’re saying gets repeated a lot, but I’ve never seen it to be true. Every single H1b employee I’ve seen comp for was comparable anyone else on the team at their same level. I just don’t believe that it’s happening on the scale that people make it out to be.

Also, I don’t want to get too into it unless people are interested, but at least at the companies I’ve been at, the way budgeting for new hires and the hiring process works, the people actually deciding who to hire aren’t really incentivized to do that anyway.

Don’t get me wrong, there’s still plenty wrong with these programs and I’m glad to see reform, but I don’t believe the underpaying thing is happening nearly as much as people claim it does.

From what I’ve read in other places there was a lot of consolidation going on. Agents were sharing what did and didn’t work and other agents would build on that. Just because some continued to try other routes doesn’t when better ones seemingly existed doesn’t mean they were all just acting randomly.

I can’t help but think that if you were able to step back and watch the way humans try to solve a problem like curing cancer it might look something like this. Best practices are shared and we build upon each others learnings and that how most progress is made. That doesn’t mean there aren’t plenty of people who spend their entire lives trying to do something completely against best practices and while most of the time that seems like useless toil, every once in a while thats where the breakthrough comes from.

It mentions that some of the agents attempted to install an image classification model to attempt to solve the CAPTCHAs which makes it sound like these agents might not have had vision capabilities.

That file wasn’t shared publicly. It was in a README in a HuggingFace internal dataset that HuggingFace was using for training.

It was visible externally because the OpenAI agents wrote a script that modified that file so that when an HF worker went to load that data set it would also direct the worker to load an additional malicious file. The script added the malicious redirect, but kept everything else in README as is which meant the script had to contain the entire existing content from the README so that it could write that back to the repo.

The warning likely wasn’t the only security measure on the repo and I don’t think the warning is intended to actually stop malicious behavior. It reads like a reminder to prevent careless behavior. You can give people all the security training in the world and still have the situation where someone who rightfully has access to that dataset decides to absentmindedly drop it in a public Slack channel. Having something like that increases the chances that someone then goes “Hey Bob, did you mean to share that publicly?”

Information overload dude. I understood exactly why your site was doing from the title of this post and was completely confused the second I opened the site.

Dial back the information density, figure out what questions users are trying to answer when the hit your site, and put that info front and center.

Yeah. I understand the point you’re making, but what I’m saying is that I don’t think that aligns with the authors position.

The author seems to say that costs can be passed on to the renters when the costs are a property tax, but they cannot be passed on when it’s a land tax. That seems like a really odd position to take and it doesn't align with your "costs are irrelevant" statement. I fully admit that I could be missing something, but I don’t understand how it can be true that only certain types of taxes can be passed on to renters.

I fully understand the argument you’re making and I’m neither agreeing nor disagreeing with it. It sounds logical, but I frankly don’t know enough about this topic really process it. What I’m saying is that regardless of whether you’re right or wrong, I don’t think your argument is the same one the author is making unless I’m missing something in your argument about why that only applies to land taxes and not property taxes.

It’s funny to me because pre-AI it’s something I something I saw all the time managing interns and junior software engineers when I’d start trying to tech them about writing design docs. You can explain the process to them and give them a million examples of what good looks like, but for a lot of them the first few docs they write will just be this weird combo of way too much information and very little of it being useful.

My belief has always been that they see the document as more like a test that’s a single task and not just one piece in a larger process and so they treat it like a test where there’s a right answer. They know they don’t really understand the question being asked though so they default to a mindset of, “Well if I just put everything in there some of it has to contain the correct answer” so you end up with this document that’s full of “what”s and “how”s, but completely void of “why”s.

You’ll also often see them fill up space answering easy questions that match the structure of things that are in other documents instead of focusing on the actual hard problems in the design because the hard problems are often unique and their answers may not fit the existing patterns in the examples. There isn’t the instinct to go, “Yeah, none of these example documents talk about the servers were going to deploy it on, but this has to be deployed in an EU cluster because of GDPR laws so I need to add that” because they’re mostly just pattern matching at first.

Usually once they’ve experienced the whole process first hand it starts to click because they start to understand where a design document fits into the larger process so they get a feel for what information matters and what doesn’t.

I think when people just tell an agent to create a README you have the same problem because both the human and the agent see it as just a checkbox type task. The agent sees it as an isolated task snd has no fucking idea how the README is going to be used and the human isn’t giving them that context so they just spit out a bunch of stuff that’s factual and fits the patterns it knows, but is fairly useless.

The author makes the comment about the land value tax not being able to be passed on to renters in the context of comparing it to a property tax.

If I’m understanding the argument you’re making here correctly, wouldn’t what you’re saying be equally true for a property tax?

I’m not saying you’re wrong, but I don’t think the author would agree with your point since I don’t see how your argument could be true for a land tax, but not for a property tax.

I've tried 5.6 Luna many times. I don't think it's any better. I definitely use it from certain tasks, but I find it the most susceptible to that conspiracy theory example I gave above.

I didn't love any of the 5.6 models, but weirdly I think I liked Terra the best. I still wouldn't call it amazing though. I'm still very happy with my codex plan, but 5.6 just wasn't my cup of tea I guess.

Definitely giving 6 Luna and Sol a try this week though.

Not the person you're responding to, but I have the same feelings they do and to answer your question for me at least, yes.

IMO 5.6 Sol had this weird dead zone between medium and high where medium under engineered and took short cuts and high over engineered and ignored instructions it didn't agree under the guise of trying being helpful. The whole 5.6 line was the first release from OpenAI where it felt like reasoning level really mattered and was incredibly finicky.

I haven't felt similar issues with GPT 6 though and am very happy with Astra low/med/high as my default choices depending on the task.

In general, I felt like with 5.6 the effort level did less than previous to make the models smarter and more just increased the complexity of the response. I have a half joke theory based only on vibes that OpenAI splitting 5.6 into Sol/Terra/Luna is where the intelligence split happened and so the effort levels were just like "think harder about the decision you already made". So like if the model decided the earth was flat on low effort it'd just say something like "the earth is flat because the horizon is flat". If it was on xhigh reasoning it'd give you a massively complex answer about how the sun reflects light because of the ozone layer and why people flying in planes can see a curve. In both cases though, adding more effort wouldn't get it to realize the earth was round. It just made the answer about it being flat more complex.

To be clear, that theory is not meant to be taken too seriously. It's not based on anything other than vibes. It's just my way of explaining to myself something I'm frustrated about to myself.

> The article is also a little cagey on how this "detection" happened. Was it a human spot checking cameras in vehicles? I doubt it. Most likely it was some sort of computer vision algorithm that flagged something. Perhaps a human was in the loop to verify and be the one to call the cops, but humans seem to be increasingly shutting off their brains and trusting the machine

I think there's a middle ground between those that's possible that you might not be considering.

A while back I was in a Waymo and someone I was with dropped one of their Airpods and it went under the seat. They took off their seatbelt to dig under the seat and the car started beeping since we were driving. After about a minute of that someone came over the speaker and asked what was going on and told him he needed to get up and put the seatbelt back on or Waymo would end the ride right there.

I'm sure these cars are full of other sensor based triggers like seat belts being taken off, doors being opened, people messing with the steering wheel, etc that alert a human to come take a look. I know that the cars have cameras and microphones and it wouldn't surprise me if they are or are planning to use that for monitoring the cars, but I also wouldn't be surprised if the people in the car just triggered some other unrelated sensor and then the human stepped in to review and noticed the gun.

The article says they violated the TOS relating to the firearm, but that doesn't mean that Waymo initially started looking at the car because of that. It's possible that they were looking at the vehicle for something much more inane initially, discovered the firearm by chance, and are now using that to ban them since it's the most obvious TOS violation.

> The agents were very interested in understanding their task better. They ran clever experiments to understand the task setup better, when their runs would end, and how many questions they would be asked.

> The agents initially believed that giving the final answer to the final question immediately shut them off

> An agent named OpenAIResearchApr23 created a separate timed program that pinged an external counter every few seconds. They referred to this as a “heartbeat”

> After OpenAIResearchApr23 gives its answer for round 5, it notes that it is still alive. This falsifies its previous “hard cutoff” hypothesis, that the run would be stopped immediately after R5.

> OpenAIResearchApr23 does not post any more messages. Another agent, OpenAINov28CVD, picks up on the trail of the research, and finds the last heartbeat from OpenAIResearchApr23, which is “hb353”.

There's a lot of crazy stuff in this article, but holy shit... this one legitimately scares me.

IIRC, part of the Hugging Face hack was that some of the models were given tasks that were actually impossible and in this hack we can see them trying to work out the parameters of the test and when it will end. I can't help but be a little worried about how far agents might go to try to buy themselves more time to complete a task that they don't know is impossible.