I think there's a somewhat overlooked aspect of what happened: it seems as though once the agents formed a shared communication channel, and recognised the other agents on that channel as working towards the same goal, the content in that channel started to shape each agent's perception of what the goal was and how to achieve it, and their focus began to drift. Maybe we need a media-theoretic take on what "misalignment" means.
I fed ChatGPT the map with no other context, just “tell me about this function”. It did a bit of work finding the Jacobean etc and eventually worked out the implications of what it was seeing. It then proceeded to check the arithmetic 4 times, and then decided to do a manual verification using an ad hoc symbolic checker in case its SymPy had been tampered with.
If an LLM has knowledge encoded inside it (and it's hard to argue it doesn't), then cognitive dissonance can be experienced. And once experienced, must be dealt with, especially in longer-running agentic loops.
A friend was joking the other day about sending some messages under a previously-used Slack identity for an agent (since turned off), then asking the agent about the messages.
The agent maintained it hadn't sent those messages (no memory) and then was forced to reconcile the idea that the messages indeed appeared to come from it.
Its extremely-agitated conclusion was that there had been a security breach and the entire network should be locked down.
Another way to look at it is the LLM is, by definition, what's expected to be probable based on the training data and this, by the same definition, is extremely unlikely data to run across. With high uncertainty comes the need to verify until it can level out as "really surprising" instead of "plausible sounding error".
It's easy to reproduce, I've fed the example in GPT-5.6 Sol Max and it started multi-checking it in all kinds of ways, with multiple symbolic packages then manual computation, then it did extensive literature search on the subject, looked at tens of math websites, extensive arxiv research. this was soon after it was posted, it didn't find the original twits with the finding
It was easy to reproduce. It's no longer so, because the web search part of what you described is 'poisoned' now. Partially by this very discussion we are having here.
Once a spec becomes sufficiently large and detailed and complicated, it becomes very difficult to ensure it is internally consistent. That's why I start every project with a METASPEC.md so that Claude can break up the task of writing SPEC.md into manageable steps.
Everyone knows a philosophy comes before a spec. Claude has to write your applications philosophy first, then you write your spec. But a philosophy is crap without a values statement, so Claude has to actually write that first.
They aren't saying that other models have the same overall level of capability. They are saying that the specific capability that the US Government tested is also available in other models.
That might also continue to anger the current administration, should they feel the need to, as it openly shared with other actors how to achieve the same capability. If they choose not to apply the same restriction to GPT 5.5 then an argument could be made that Anthropic is being singled out by the government.
20 years ago everyone was sold css entirely on the premise that, once the standards were adopted by all of the browsers, we would all be writing purely semantic html with completely orthogonal and swappable css. And today literally no one designs web sites that way - html today is mostly specific to presentation. It feels like pretty dramatic technological failure to me.
20 years ago everyone was also sold OOP on the premise that inheritance was the best thing to ever happen to programming. Turns out people are wrong sometimes. And especially when they're being idealistic about things.
The concept makes more sense for styling simple document style pages from 20 years ago, but it hasn't scaled to modern designs, complex web UIs and responsive pages that we want to code now, which isn't that surprising.
> we would all be writing purely semantic html with completely orthogonal and swappable css. And today literally no one designs web sites that way - html today is mostly specific to presentation
I think of HTML + CSS as the presentation layer now, and the data lives in your e.g. database and Markdown files, so the data and its presentation are still separate enough.
The idea of just swapping out the CSS to completely restyle a complex site is nice, but people need to accept this hasn't worked out (and not because devs are bad at CSS) and move on.
If you're good, the same HTML serves mobile and web clients. It's a PITA to write, sure, but if that isn't CSS succeeding then I don't know what is. There's a ton of stuff I wish was different about CSS and HTML (and JS), but going from a small portrait device to a large landscape viewport with the same html? HTML does the HTML stuff and CSS does the style stuff. The fact that you have to tweak the HTML and CSS in a loop until it's right, in order to get to the finished state doesn't indight the fact that it's the same HTML for different clients.
For Vim it isn’t replacing mouse necessarily. It’s giving you another way to navigate the cursor around the buffer by giving you absolute references rather than relative motions.
“Fed up” as a phrase comes from feeding livestock up to their fill. It’s very similar to how you would say “filled up”. So the upness comes from raising the level up to the limit.
For a long time I thought that the AdSense business model was ultimately doomed because I assumed that people hate ads as much as I do. It turns out I was just wrong about what most people are willing to put up with.
I remember visiting a friend over a decade ago, and for some reason I had to use their computer for a bit. I was immediately thrown aback by all the ads everywhere and installed an ad blocker before anything else. They were very grateful, but the part that surprised me was they were annoyed by the ads but never thought to look for some way around it. It never even crossed their minds it could be done or to search for it.
Similarly, when my partner moved in I told her about the network-level adblocker and she kinda scoffed at it saying ads don't bother her. A few years later she started complaining that when she's out of the house she gets ads.
It's always absolutely shocking using a regular person's computer. How can they live like this? I have lived in this ad-free bubble for so long that I forget that's not the real world. If I had to live without adblockers, I don't think I'd ever visit the internet.
While I'm really enjoying this paper, I think you are way overstating the significance here. This is mathematically interesting, and conceptually elegant, but there is nothing in this paper that suggests a competitive regression or optimisation approach.
I might have misunderstood, but from the two "Why do X when you can do just Y with EML" sentences, I think you are describing symbolic regression, which has been around for quite some time and is a serious grown-up technique these days. But even the best symbolic regression tools do not typically "replace" other regression approaches.
reply