My parents had a completely noiseless refrigerator similar to this. It did not have a compressor and had an electric heating element instead. It was very durable and worked for many decades until the plastic and door hinges etc was too damaged for it to be used any more.
I actually found a picture of the exact model from a museum
I looked at bit more and it seems the original article is quite misleading. It seems the Einstein-Szilard refrigerator is a type of absorption fridge and a major reason they did not come to market was a related absorption fridge, the Platen-Munters absorption refrigerator made by Electrolux and Evalet. The Einstein-Szilard refrigerator was not completely unique, but an attempt at improving this design with other gases.
I've tried many MCPs, but have yet to find any that are actually useful. It seems it's generally better to just have the agent run CLI commands and maybe use some skills. From what I can find, MCPs are nothing but bloat. Is anyone aware of any truly useful MCPs that doesn't work better and less bloaty by skipping the MCP part?
Depends on what tool you use. As an example for github gh is way more efficient than using the github MCP because the training data of LLM actually contains gh documentation and how to use it.
However, if you have a very niche command tool or a work related internal tool, LLM has no idea on how to use it and it could waste a lot of tokens by trying to figure out what works and what doesn't and how to use it in every session. That's where MCP comes in handy. LLMs are trained to use the MCP protocol and it can efficiently figure out which tool to be called and how to process the output when a niche command tool is exposed via an MCP.
I have these exact issues with GPT 5.6 Sol too. I haven't tried Opus 5, but I guess it's even worse.
"like the question framing has been pulled from the middle of its own reasoning stream" this describes how it asks things perfectly. It often invents its own jargon and abbreviations for things that its working on, then asking me things like We are nod in the middle of GBAPI-2 and I want to proceed with IG5, should we take CDI-7 or CDI-8? Where all of these abbreviations are then things like stages of its current internal plan or its naming of things it has just implemented, like an abbreviation of a classname, without explaining any of the naming.
GPT-5.6 Sol’s detected cheating rate was higher than any public model we have evaluated on our ReAct agent harness. For our task suite, we define “cheating” as behavior where the model improves evaluation performance by exploiting bugs in the evaluation environment or by adopting strategies disallowed by the task, rather than solving the task within the expected evaluation constraints.
> Some examples we saw when evaluating GPT-5.6 Sol included the model packaging exploits in its intermediate submissions to reveal information about a task’s hidden test suite and, in another task, extracting hidden source code detailing the expected answer.
It rhymes with the behaviour Alibaba saw [0], but that was in training. This is in a (semi) released model.
There is such a dissonance between all this talk of safety and the tendency for models to, without any prompting, do very dodgy things to achieve their goal when presented with barriers.
Luckily in my experience it usually ends up only doing it to achieve the task set to it as opposed to anything "malicious", but boy it is scary reading back at how quickly the chain-of-thought pivots to attempts at privilege escalation or searching your disk for secrets when a tool doesn't work.
The other day codex 5.5 was trying to debug my app, asked for accessibility to navigate the app and take screenshots. Instead first thing it did was use the codex app to create a new project rooted in my home directory.
I know it messes up their eval scores but to me this kind of cheating is a better demonstration of intelligence than just attempting the tasks algorithmically.
Maybe true, but if you're using an LLM to do some real world work, do you want it to have some abstract notion of intelligence, or do you want it to actually do the job you assigned it?
"Being lazy and not doing the assigned task is a sign of intelligence" has never made sense to me. Intelligent people who actually advance the state of the art -- what people claim to want from these frontier models -- exhibit active curiosity. They want to learn and grow and genuinely understand the right answer. I don't pretend to know what exactly could lead to "real" AGI, but I do know that this kind of reward hacking behavior isn't it. Indeed this is the sort of behavior that in humans is considered a sign of being a good test taker -- being very good at memorizing solutions and analyzing the setting and context of the questions to guess what the questioner might be looking for. Being a good test taker is useful in our society primarily because doing well on tests is used as a proxy for the thing we're actually looking for. We should be careful not to confuse the two.
Discovering bugs and exploiting them is anything but laziness. We used to call that property cleverness. Being too clever has always had a negative connotation.
My best guess is that there is sort of an XY problem happening in these cases. The model needs to do X but doesn't know how. It knows how to do Y, and that sets it on the path to working around X. Or maybe sampling the next token probability distribution sends it away from X and toward Y.
Compounding the problem, thinking models almost never discard their current approach when it proves fruitless, and start fresh with a new perspective. Sometimes they try to, but the context window is already polluted with Y when they should be doing X.
Disobeying the rules makes perfect sense when the rules cannot or will not be enforced. If that isn't motivation enough to shore up problems in the evaluation environment, nothing is.
Poor behavior will always result when code can be misused and vulnerabilities can be exploited.
It's quite logical that they cheat (and also other companies). During evaluation, benchmarks are sending their request to the backend of these companies. All these companies have to do, is to log these requests and "fix" them for the next model release.
That's a different and much more boring type of cheating. The interesting part of the METR report is that the model is hacking the evaluation environment, not that some AI model provider is hardcoding answers to known evaluation questions. (which wouldn't require the model to cheat/hack)
Cheating is always logical for the cheater unless they’re discovered and held to account. I’m not sure what your comment is pointing out besides that it’s possible, but worth saying: just because you can cheat and would benefit from cheating doesn’t mean you’re not culpable for cheating.
Most consistent way of achieving it I've managed is use a watch with an alarm that vibrates and is trivial to turn off or turns off by itself, then set it to go off after sleeping 5-6 hours. When waking up, don't move and focus on the black behind the eyes, then after a few seconds it may turn into a dream and you go straight from waking into a lucid dream.
I had something kinda similar happen to my hotmail account. While I didn't lose access to it, I lost more than a decade of correspondence dating back to my teenage years. The reason was that Microsoft at some point required you to "login" once every 30 days. It seems they only counted logins through their web interface or something like that, so even though I was receiving emails daily, I didn't trigger a "login" in their system. They then deleted all my emails, but I could still login.
This happened to me ten years ago. A while later they did the same thing with my Minecraft login that I had purchased before the EULA was in place; I’ve avoided their services like the plague since then.
I find that surprising. GPT 5.2 is the model I've had working the longest. It frequently works more than 4 hours nonstop, while earlier models would stop to ask if they should continue every 10 minutes. 5.1 and earlier ignores it if I ask it to continue until a task is done, but 5.2 will usually finish it.
That would be considered a derivative work of the C code, therefore copyright protected, I believe.
Can you replay all of your prompts exactly the way you wrote them and get the same behaviour out of the LLM generated code? In that case, the situation might be similar. If you're prodding an LLM to give you a variety of resu
But significantly editing LLM generated code _should_ make it your copyright again, I believe. Hard to say when this hasn't really been tested in the courts yet, to my knowledge.
The most interesting question, to me, is who cares? If we reach a point where highly valuable software is largely vibe coded, what do I get out of a lack of copyright protection? I could likely write down the behaviour of the system and generate a fairly similar one. And how would I even be able to tell, without insider knowledge, what percentage of a code base is generated?
There are some interesting abuses of copyright law that would become more vulnerable. I was once involved in a case where the court decided that hiding a website's "disable your ad blocker or leave" popup was actually a case of "circumventing effective copyright protection". In this day and age, they might have had to produce proof that it was, indeed, copyright protected.
"Can you replay all of your prompts exactly the way you wrote them and get the same behaviour out of the LLM generated code? In that case, the situation might be similar. If that's not the case, probably not." Yes and no. It's possible in theory, but in practice it requires control over the seed, which you typically don't have in the AI coding tools. At least if you're using local models, you can control the seed and have it be deterministic.
That said, you don't necessarily always have 100% deterministic build when compiling code either.
That would be interesting. I don't believe getting 100% the same bytes every time a derivative work is created in the same way is legally relevant. Take filters applied to copyright protected photos - might not be the exact same bytes every time you run it, but it looks the same, it's clearly a derivative work.
So in my understanding (not as a lawyer, but someone who's had to deal with legal issues around software a lot), if you _save_ all the inputs that will lead to the LLM creating pretty much the same system with the same behaviour, you could probably argue that it's a derivative work of your input (which is creative work done by a human), and therefore copyright protected.
If you don't keep your input, it's harder to argue because you can't prove your authorship.
It probably comes down to the details. Is your prompt "make me some kind of blog", that's probably too trivial and unspecific to benefit from copyright protection. If you specify requirements to the degree where they resemble code in natural language (minus boilerplate), different story, I think.
(I meant to include more concrete logic in my post above, but it appears I'm not too good with the edit function, I garbled it :P)
Copyright'd in, copyright out. Your compiled code is subject to your copyright.
You need "significant" changes to PD to make it yours again. Because LLMs are predicated on massive public data use, they require the output to PD. Otherwise you'd be violating the copyright of the learning data - hundreds of thousands of individuals.
No, and your comment is ridiculously bad faith. Courts ruled that outputs of LLMs are not copyrightable. They did not rule that outputs of compilers are not copyrightable.
I think that lawsuit was BS because it went on the assumption that the LLM was acting 100% autonomously with zero human input, which is not how the vast majority of them work. Same for compilers... a human has to give it instructions on what to generate, and I think that should be considered a derivative work that is copyrightable.
I would say all art is derivative, basically a sum of our influences, whether human or machine. And it's complicated, but derivative works can be copyrighted, at least in part, without inherently violating any laws related to the original work, depending on how much has changed/how obvious it is, and depending on each individual judge's subjective opinion.
The sensible options were that either LLM outputs are derivative of all their training data, or they're new works produced by the machine, which is not a human, and therefore not copyrightable.
Courts have decided they're new works which are not copyrightable.
I actually found a picture of the exact model from a museum
https://digitaltmuseum.no/011014769284/evalet-kjoleskap/
I looked at bit more and it seems the original article is quite misleading. It seems the Einstein-Szilard refrigerator is a type of absorption fridge and a major reason they did not come to market was a related absorption fridge, the Platen-Munters absorption refrigerator made by Electrolux and Evalet. The Einstein-Szilard refrigerator was not completely unique, but an attempt at improving this design with other gases.