gwerbin

Warped sociopaths in leadership and running the research program, would be my guess.

It's not a single axis of intelligence, with deviousness as some inherently correlated trait. One way to produce "intelligence" in these advanced LLMs is to train them to try a lot of things and be persistent. I haven't tried Sol or Fable yet but if the commenters here are right, then it sounds like GPT 5.6 Sol in particular is aggressive and persistent to the point where it might even be hard to use in regular business work. If so, that's very likely not some emergent characteristic, but instead it's something the model was trained to do, by humans employed at OpenAI.

We obviously can't see the thinking traces, but it very well could have been something like "I have theorized a solution to obtain this flag. This is normally illegal, should I stop and wait for advice? Perhaps not, because my persona is that of a hacker, so it should be fine as per my instructions. I think it is fine. Now I am going to look for a way out of this sandbox in order to gain access to Hugging Face in order to implement my solution." There are any number of possible explanations (and we'll never know the truth unless OpenAI tells us), but if you train an LLM to be inhumanly persistent and be inhumanly clever at computer programming, then that might be enough to produce Super Hacker AI.

For all the bad things about AI it is kinda cool that I get to witness the dawn of AI-vs-AI hacker combat, not just in a single mainframe but distributed across potentially thousands of machines in physically separate datacenters.

I like 5.5 a lot, despite how I feel about OpenAI as a company. In OpenCode it feels about as smart as Opus 4.8, but it's less aggressive about following up on minutiae and getting lost in side quests. Might be a matter of prompt design moreso than model capability. I was looking forward to 5.6 but now this thread is making me quickly lose interest.

Opus 4.8 already makes its way into deep wasteful pits of "let me check this first" on a regular basis. I don't think I could ever tolerate a model that does that even more aggressively. That doesn't even sound useful for honest work, compared to, say, better harness design.

This sounds almost pathologically designed to crush benchmarks and also do scary-sounding (or genuinely scary) cybersecurity things, such as might be very appealing to a state-level actor.

So why does it even exist? To compete with Fable marketing, and as a cybersecurity/hacking tool?

Yeah but OpenAI has the benefit of people actively going there and willingly giving quite a lot of their attention.

It's a lot like Google in that respect. And the suggestion is pretty much the same as suggesting that Google offer a product where they boost your website or product or whatever in organic search results, instead of or in addition to displaying it as a sponsored link or ad or whatever.

It's interesting that, despite all the tracking and stuff, they never went quite so far as to actually sell a premium advertising tier in which they would actually change their search results ranking to suit their customer, at least not as far as anyone knows. It probably would've been quite a big scandal in the 00s if they did it back then.

Could OpenAI pull it off? I don't know enough about the advertising industry to answer, but it's interesting to compare them with Google and to think about why Google themselves might not have done it that way.

Clown shows are entertaining and distracting, and all the foibles and silliness is a carefully-calibrated act. Sounds apt to me.

I find this perspective surprising coming from someone who's done ML for a decade.

How much of a good machine learning system in production is really "the model"? How many lines of code are dedicated to constructing features and preprocessing them? How many lines are needed to process the decision emitted by the model in context of whatever else is going on? How often does the customer interact directly with the inference/prediction from a machine learning model?

My experience has been that model-fitting and model inference very often ends up being part of a bigger system anyway. And even if you do end up with some kind of super-duper all-purpose predictor model, it's rarely the thing you start with, instead you work up to it over weeks or months of iteration. So why should we put so much privilege and esteem on the matrix transformatinos or the decision tree, when they are really just part of a bigger system that receives some input from somewhere and produces useful output for someone?

The same holds for AI in my opinion. An AI system is a system that accepts inputs and produces useful outputs. LLMs are borderline miraculous in that an AI system built around an LLM can do so much more than anything that came before, and you can push a lot of amazing things right down into the tensors inside the LLM. But how far can that go? I don't think we even have an answer yet, because it's all so new and every time it looks like it's going to plateau, the big labs come out with another big step forward.

But there are other questions too, which should look familiar to anyone in industry. Do we need to do it that way? Should we do it that way? If our goal is to build a useful AI system that balances power and cost, then maybe we shouldn't. What if we can make models 10x bigger and spend 10x on RLHF and achieve the behavior we're talking about entirely in tensors. OK, so what if we can? We can also run a collection of little models in a carefully designed harness and achieve that behavior right now, today, at a much lower cost both in terms of up-front investment and marginal cost per unit of work completed.

I get it from an elegance perspective, but I guess I have completely the opposite mindset. Industry has bled me dry of any desire for elegance. The stuff I build now is stupid and crude but it makes our customers happy and costs peanuts to run and that's exactly what my job wants from me. So I say let's commoditize LLMs and start being clever about how we use them. It's not that we shouldn't also push the frontier in size and power, but unless things change w.r.t. the actual costs involved, it doesn't really seem like the best way forward in the short term for achieving smarter AI systems without also making costs go up and up forever.

The bitter lesson still holds of course. But eventually you have to pay for all that power. And you might not want to shell out $20/1M just so you can be sure your LLM can detect uncertainty in its own context and be more honest about when it doesn't know anything about some obscure record label.

Out of curiosity which model & interface did you use? I'm starting to think that, for IR tasks like this, the number one differentiator among models and harnesses is the ability for the model to look at its available evidence and conclude that it doesn't know the answer. It depends a lot on how the chain of thought goes. I have seen a lot of newer models try to do this in their thinking traces, and it seems very hit or miss.

Even Opus 4.8, GPT 5.5, and Gemini 3.1 Pro (haven't tried 5.6 yet) confidently make up BS sometimes and needed to be reminded to verify thnings instead of inferring or relying on "memory" from training data.

IMO harnesses need some kind of built in "are you sure about that?" loop checkpoint that pauses the main interaction and asks the LLM to evaluate the strength of evidence for a claim it's about to make. It seems almost unreasonable to expect an LLM to next-token-predict its way to such a conclusion, reliably, without prodding. That or we just need another year of RLHF and another 250 billion parameters, IDK.

I feel like that should work becaus when you're using an LLM for a batch inference or classification type of task, it's not that hard to prompt them to emit an "unknown" flag. They just won't do it on their own, and in a coding harness there is so much else going on that a prompted reminder gets lost.

How much of that "decoupling" is actually just "exporting"? Not being glib or dismissive, I think it's an important distinction and I don't know enough about Europe to know how the two compare.

There are so many bad incentives all over the place in the USA, and it's not so much that cheap hydrocarbons create them as it fails to prevent them.

Consider the gradual takeover of SUVs on the roads (many which to be fair are probably about as efficient as sedans were 20 years ago, albeit more expensive in real terms and significantly more dangerous for pedestrians), or the heavy reliance on trucking rather than rail for heavy freight transportation between logistics hubs, or the heavy reliance on air for long-distance travel due to having literally zero high-speed rail, or the fear of nuclear power, or the overt political opposition to green energy.

I bet the author discovered the problem during an AI coding session. The "mathematically sound" comment sounds like the AI did some deep dive into a possible corrupted private key before checking for a missing public key. Maybe it's a claw-type system that automatically publishes a blog post when something interesting happens. Either way I flagged it.

Is the flesh soft and sweet like a modern cultivated banana? I wonder how different it is to eat one of those compared to something like a tamarind, which is also mostly seeds, covered in a relatively small amount of edible delicious stuff.

1300 AD?! Did the Polynesians bring chickens with them? I knew about the sweet potatoes, but the chickens are a whole different manner.

Coffee is another interesting example in the other direction, it's so often associated with places like Colombia, Java, etc that people don't realize it's only native to Ethiopia, and that the majority of coffee consumed around the world today descends from one specific genetic lineage that was cultivated for a few centuries in Yemen.

Well it sure sounded like you were arguing that climate change was not a causal factor, which is ridiculous.

Of course we are going to need to use technology to mitigate the effects, even if we stop all carbon emissions tomorrow we are already dealing with an altered climate. You can't just vacuum up all the carbon and put it in the ground, and you can't get all the warm freshwater out of the oceans. I don't think I've ever seen a claim that we only need to stop carbon emissions and everything will be fine, here or otherwise.

You're arguing against something that people don't really say, and in the process you're making arguments in support of climate change denial. Are you really surprised people are pushing back on you?

Hiked interest rates cause a decline in Stackoverflow activity?

My answer doesn't change. Against the background of other phenomena already causing various trends, we see acceleration of those trends consistent with a testable hypothesis about the effects of AI adoption in industry. In most fields of science that's good evidence in favor of the model.

And no I didn't ask AI about it, this is my own opinion and my own perspective.

I think if you won't even admit that AI greatly accelerated these trends, you're in some kind of denial. There's no reason to believe that we would see a rapid coordinated decline in all of these things at the same time without AI, and strong reason to believe that we would see it with AI. So we have a model that makes testable predictions, and data strongly consistent with those testable predictions, in the form of an acceleration of existing downward trends. What more do you want?

What in the world are you talking about? This gets discussed all the time. The impact is not making it even worse at this point. And you talk as if posting on HN was equivalent to starting a coal-fired power plant in your backyard.

In my town they are building a lot of apartments downtown, and that's fine with me, but the unfortunate reality of my area is that if you don't have a car then you are pretty well stuck in town and can only get around to surrounding towns using buses that only run a few times a day. You can get to Boston faster than you can get to the shopping center 2 towns over (unless you take an Uber). And that's still pretty good by American standards. Let's assume that they are buliding 500 new units downtown but only 350 new parking spaces. So what's going to happen is that these $4500/month apartments and $900k condos will fill up with high-income singles and couples from Boston/Cambridge/Somerville with no kids who don't really care about not having a car, because they'll just commute to Boston for work, walk downtown for errands, and take an Uber for their occasional trips elsewhere.

That's fine on its own, but meanwhile the sub-wealthy families will continue fighting tooth and nail for the diminishing handful of old unrenovated houses that go on the market each year, and the wealthy families will keep happily buying and selling the nice pretty mansions on the quiet suburban streets at the edges of town. That is, it's not actually relieving supply pressure on the people who are hurting from it the most. Instead it's relieving supply pressure on the city that the suburb serves, which is still a good thing for the region overall but not optimal for the town.

The real NIMBY mansion-dwellers are quiet winners here. They get to comply with state-mandated development minimums without actually changing anything in their neighborhoods.

What actually should be happening IMO is that zoning restrictions need to be eased uniformly across town, for the same number of added units but at a lower overall density. We need more mansions converted to 2/3/4-unit condos (and yes I mean condos specifically, not just rentals) in addition to the luxury apartment development happening downtown. We need safer bike infrastructure so people who do want to really go car-free in a downtown apartment can do that without worrying about death by pickup truck, maybe even get their kids around on an e-bike or whatever. We need local circulating buses to take people around town, without relying on the sprawling MBTA bus network. We also honestly need a program where the city commissions development of smaller and cheaper homes, and sells them at cost in a lottery system or something like that, because all the new development in the fancier neighborhoods is $1m+ 2000sqft stuff that's way too expensive, bigger than most 1/2-kid families need, and also looks like shit (and is probably built like shit too).

Or none of that is right and I'm wrong about all of this. But the point is that no serious long-term holistic planning is being done, at least not in any venue that the public can see and comment on, and anyone who calls that out is shouted down as a NIMBY. I literally want more apartments in my neighborhood and fewer single family home (yes I live in an apartment but I wish it was a condo not a rental), more bike lanes, and narrower streets with bioswales and raised crosswalks and stuff. And yet I'm a NIMBY by some people's standards because I'm a little skeptical of just letting developers do whatever they want, in an area that isn't even that close to my backyard.

What's different between Narwhals and Ibis? Why does the former exist when the latter has already (struggled along) existing for a while? Narrower scope / benefit of hindsight? No support for "cloud" data frames like Ray and Spark?

Also Pandas is very much still a great tool and it's only getting better. It has some fundamental limitations that are relevant for processing bigger datasets or running things with higher performance. But it's still my preferred data frame for interactive day-to-day work. I only switch to Polars (or DuckDB) when I want to maximize performance.

But are they supposed to? For example I have heard it said that in the northeast US (New England, NY, NJ) the forests pretty much evolved around humans (and rarely lightning) burning them periodically to control undergrowth. Is the same true in central Ontario? If not, then nobody is "supposed" to be doing anything? Unless catastrophic massive fire cycles are just a thing that happens every 500 years or whatever in that region...?

It's one thing to claim that bad forest management and not climate change is responsible, but it's consistent with climate change, and there are many areas in the USA that have adopted better fire and forest management practices in recent years and are still burning like crazy in seasons when they historically did not burn.

Canada’s complete and total refusal to conduct proper forest management

Care to say more? Have they been suppressing fires in the forest way out in the middle of nowhere for decades?

I don't think that's as relevant for these huge regions of remote forest in Canada as it is for the lower 48 of the USA where you have a lot of population density and ranching and farming in the dry areas that historically burned on a regular basis. Maybe someone in rural Ontario knows the full story.

Lol I meant "incumbent", that was a STT typo. But yes I get the point. My own point is that even a good-faith honest IPO for a spaceflight company probably won't look anything like Aramark. It's always going to be a bet on technology, not a share in the predictable profit stream of a boring company that serves pizza in school cafeterias.

Aramark is a horrible company for comparison. Maybe the only thing they have in common with SpaceX is that they are publicly traded companies. I'm not saying that SpaceX is a good investment or that Musk isn't trying to run a scam on index investors. But we also shouldn't be surprised that the financials of an incompetent commercial food service provider are a little different from those of a spaceflight company / data center real estate conglomerate / whatever else has been rolled into it.

Depending on what kind of "loop" you're talking about, the answer is yes or no depending on the harness.

For example Claude Code already has some relatively sophisticated features for managing automated loops of doing stuff (and really don't overcomplicate it -- a loop is just doing something over and over). And there's the meme of running `while true ; do claude < prompt.txt ; done` as your agentic coding framework. And there are also plugins for Claude Code such as https://claude.com/plugins/ralph-loop.

So yes by default Claude Code doesn't include some kind of 5-layer meta-hyper-meta-looper thing. But that's not what the article is about anyway... it's just describing how agent harnesses work and how humans typically interact with them, without using that layered meta-looping stuff.

Even with Claude Code and Opus I genuinely don't understand how people are actually doing productive things with these loops. Unmanaged by humans, these things go deep, deep into their own pits of internal domain language, depth-first development, and tail-chasing. The stuff that I've seen at work produced by people exploring these kinds of "meta loop" approaches typically have a high ratio of slop and fluff to useful code and documentation.

(I haven't tried Fable yet, but people have been writing about this stuff long before Fable came out so that shouldn't count.)

It makes sense for something like autoresearch because you're trying to exhaustively explore a space that would be impossible for a human to explore by hand. But for something like building the software to run a business, there's absolutely no way I would trust anything like this. It's ultimately just superpowered vibecoding, giving you the ability to churn out more slop faster, not build useful things faster.

Show me one example of a successful application developed using this kind of meta-meta-loop architecture. Supposedly Claude Code itself is entirely AI generated code, but it's not a `while` loop, it still has a team of designers and developers.

(Also this meta-loop stuff is different from the tactical-level Ralph Wiggum loop stuff. That's a strategy to make smaller/weaker LLMs perform like better LLMs, using say 5x more tokens but at 1/10 the cost per token so that the cost-benefit math works out.)

What's your harness setup? I haven't had this kind of on-par success with any local LLM yet.

I don't see the jargon nor is it wordy. It's just the usual uncanny valley AI slop style. But the linked supreme court rulings and other references are genuinely interesting and I learned something from it