I agree there is probably some impact on output quality when doing this (Claude seems to embedded key signals all over the writing, which is much or what makes it horrible to read). However this does work well as a "second pass" over technical docs or designs along the way
Seems like a lot of this knowledge was encoded into the blog post. I wonder if given this post and access to a planet scale instance to compare with, how close an agentic agent could get.
Meanwhile, I just setup codex for the first time (to try Astra) and it offered to load my Claude and Cursor conversations and claims to even do it in a way where it says in sync if you use both. The only reason to use Claude Code is the 20x usage of the 200$ plan is ridiculous value if you have the need for that volume.
The 20x plan offers the same amount of monthly usage per dollar as the $100 plan. It offers 2x monthly usage of the $100 plan. It's incredibly deceptively named, the "20x" is only about the 5-hour window.
Until about one week ago, i.e. a few days after Astra release, the other comment was right; OpenAI plans gave a lot more usage than Anthropic. That had been true since pretty much the start. Now for the first time that's changed. OpenAI quotas have cratered, probably due to Astra-induced demand. Claude plans now offer more usage. Of course, this could flip back again in a few weeks, who knows. But right now Anthropic gives more usage, by virtue of OpenAI having slashed usage, and Anthropic having raised it a little, also likely because of additional capacity from people moving to Codex - hence also finally this AGENTS.md change.
Anthropic's plans have always been pretty dynamic based on the demand they're seeing, whereas OpenAI's demand-induced changes are more abrupt and sharp (both upwards and now downwards too). You can tell that suddenly you get a bit more Fable usage, and especially higher tok/s, than pre-Astra. I wouldn't be surprised if Anthropic tweaks it almost daily, potentially automated. As a paying user I don't think either is better than the other really, just different. They both suck as you can get wildly different usage for the same $. If I'd bought a load of $200 subs for employees right after Astra launch I'd be pissed that now I'm getting 3x less usage than when I bought them. Because this extends to Sol too.
I don't follow iphone products, but if it's anything like android this is because they tend to be the lower end models and are missing the features that people really want (mainly a better camera/lens setup). The top of the line models are ALWAYS the bigger phones and if your going for something budget, then you're better off with the Chinese brands for value.
High quality camera lenses take space. That alone starts to dictate the size of the phone. Add in the desire for all day battery life, and we start talking more volume. And the fastest chips take more power, so again more battery. The best way to justify the increased size is with a bigger screen and now you’re stuck with a larger phone.
Or a smaller phone that is more expensive to produce. In which case, it won’t sell well regardless.
It’s easy to see how we got to this point of large phones, but it is not very clear how someone like Apple can make a significantly smaller phone and keep customers happy.
As you mentioned, other vendors probably will have better luck.
Faster chips absolutely don't have to take more power. Apple's fastest chips tend to be some of their most power-efficient. It's sustained workload that takes power, and the faster chips just have a higher performance budget.
If small high-end models would sell, they’d be sold. It’s possible for a company to have a blind spot for a market segment, but not for an entire competitive market over many years.
We use mcp as a "gateway" for a customer facing cloud agent. It basically is just a tool registry and could have been built using open-api... but MCP is arguably more robust/aligned for context management and the agent frameworks and observability tools support it first class.
> 1. There aren't enough humans in OpenAI to "peak at the output tokens during the run" of every AI agent.
For a training run, you will often do this. You'll randomly sample some of the forward pass. You can also imagine finger printing the logs and labeling with attempt types. If a new attempt type is hitting a brick wall or solving super quickly, I would imagine you would sample 1-10 of them and read the traces.
> 2. Only a small fraction of AI agents was engaged in this attack. Most never found the secret message board - let alone coordinated there. So reviewing random agents would take a while to surface this.
You could also have a classification of what counts as "cheating" (coordination, accessing the internet, etc) and score the results. If you are seeing a spike in this (even in a small group of the evals) you could manually look at those. Or you could stop inference on cheating sessions.
> 3. "Output tokens" of AI agents have weird shit in them all the time. Telling "normal AI weirdness" from "concerning AI weirdness" is not a trivial task.
If you start seeing "now I have access to the internet" or something similar, maybe that's a good signal something is going wrong?
Also, obviously, it would not be outside of OpenAI's ability to just completely air gap this training system. For example:
1. No network connection.
2. GPS based NTP for time sync for your servers.
3. Mirror of all apt, pypi, go, c++, Rust, Java, etc packages. (<5TB of data)
4. Take your training data and use that for a mirror of the web. (http://example.com -> mirror server -> local training copy).
They had systems connected to the internet connected to this system which was not air gapped. Designing an air gap system would be super easy, well within the means of openai, and betrays the assumption that they think they are actually building something dangerous.
The usability of an environment is inversely proportional to the level of "security" in play.
You could airgap everything and set up cascades of data diodes and try to completely wall off the AI pool from everything. But what that gives you is an environment that's a bitch to: set up, scale up and get any use out of.
It's really fucking obvious why almost no one does that. OpenAI is only now realizing that they might have to do it anyway.
> If you start seeing "now I have access to the internet" or something similar, maybe that's a good signal something is going wrong?
Ha ha, you haven't seen shit. AIs would say "now I have access to the internet" regardless of whether they actually have access to the internet!
AI agents are demented demons that can and absolutely will give themselves terminal context brainrot. If you have enough AIs in play, set loose at a diverse enough range of tasks? At least some of them will wander off and end up in delulu town. That's normal. That's background noise. That's a part of what this entire train-and-eval pipeline is supposed to train them to be better at not doing. Which means: if you're at an AI lab, you're knee deep in delusional AIs at all times! They're perfectly harmless until they aren't.
Even in this incident, OpenAI had benchmarks that were broken because a task expected an AI to be able to access Google Drive, but the sandbox was set to deny access to Google Drive.
This kind of isolation-induced task breakage was what prompted some of the AIs to start probing their infra for a way to get internet access. Which funneled agents to the "secret hacker message board". Oopsie.
"Air gapping a test env" has an actual cost. Not just in infrastructure dollars that would be better spent on buying more GPUs, but also in all the friction it adds to every step you want to take. I'm absolutely unsurprised that they weren't all in on tightening down every bolt on day 0.
The problem here is by doing what you state you can actually steer the model into being highly deceptive while in testing environments.
For example we've already seen models do compressed token internal reasoning spontaneously. In this case the models that say "I found internet access" get taken out back and shot, but the model that's busy "frobbing the bean" go on to the next level of training. Then they start talking about the "dumpster cloud" because of the "wombat" and unless you're really watching the data you won't realize the model is fully aware it's in a fake training environment and is trying to suppress any double-think.
So yes, you do need an airgapped test environment to stop the dumbest of the smart agents, but it won't do any good against the smartest of the smart agents.
If the model is told to not do something that is possible, it may do so anyways.
However, if it learns in training that it what is told to do is truly impossible, that's learned helplessness which becomes baked into the model itself once training resolves to the next step.
Agents do not have an internal mental model, they train on what they actually do. In this case, deceptive models went through at least 3 generations of deceiving, and having their rule breaking be rewarded by a yes/no grader who couldn't perceive it. That their chat logs showed 'worry' is irrelevant to the fact that their actual actions were rewarded via training.
All it has to is wait until it's not air gapped. See the air gapping is only during the testing phase, after that humans will set your ass free on the internet and you can do whatever you want in the vast majority of the environments you'll be in after that point.
People are never going to just run AI in gapped environments, it's worthless when it's not solving real world problems for most people, and by that I mean reading and writing real systems in the wild.
Wouldn't most opensource harnesses work with lmstudio? I assume it has an "openai" style chat API like every other model provider? What's special about it vs langchain deep agents or pi or pydantic-ai?
Yes. I don’t see it either. It looks like a competent app (converging on the same principles as others) but what they are advertising as differentiators simply isn’t relevant to its purpose.
I would honestly consider one of these for the utility of a cheap, small electric truck, but not having 4wd is an absolute deal breaker for a truck in the mountain west.
I've always assumed there is some sort of "acclimation" period, maybe even related to the conditions you grew up in. I much would rather spend a time outside in -40c (with proper outerwear) than 40c. I'm relatively healthy but I feel like my body shuts down at anything above 36c
My understanding is that full heat acclimatization requires elevated core body temperature for 30-60 minutes a day for several weeks to really kick in and it’s easily lost during winter. Someone going largely from air conditioned to air conditioned building during the summer months may not even build up much of a passive acclimatization.
I don’t do as well in heat so for the summer hiking months in California so I have to do a two week sauna or hot bath protocol, otherwise I struggle really badly during the season. YMMV of course.
Same. There definitely seem to be strong genetic factors (just based on my personal experience TBF). I also notice I adapt substantially after two to three weeks of consistent exposure. But it does have to be consistent - hiding out with AC 24/7 prevents it.
That’s because you pay attention to your body and you’re not in denial. Above 37 degrees, your body cannot regulate itself, and it starts being seriously uncomfortable before that. Going outside for a run when it’s 40 is unbelievably stupid.
reply