jsw97

You're right, I missed that. That really is troubling.

I am wondering about the author's allegation that there is a user filter, not just a prompt filter.

Of course it could also be the case that it is just a prompt filter, but Fable sees memories from the authors' prior sessions that cause a rejection. I wonder if the author could control for this is in some way, if Claude lets you run isolated session without memory access.

19 subjects, assigned sedentary or active based on habitual physical activity levels. Subjects were screened on basic health measures.

The problem with this is that people are sedentary or active for a variety of health-related reasons that are not captured in any screen (esp. the crude one used in this study). As a predictive study, this is fine, sedentarism predicts a lot of bad things. But it doesn't, on its own, suggest that becoming active is helpful. See also grip strength and mortality.

In my own ham-fisted experiments with coding loops, one pathology I have noticed is that the LOC just spirals out of control. That's likely because of the layers of defensive fixes, etc., that get built. That inevitably causes context bloat (or at least navigational friction) and results in quality decline.

I wonder how many loop-related issues could be addressed by simply fixing a LOC budget, or assigning a cost in some way. Unclear how you would dial in the right numbers, though.

I do the dinosaur version of this, with screen and xpra. Works great. CC and codex being terminal-native may make terminal-based (or at least terminal-comfortable) editors seem more natural to a broader set of people.

If USG bans these models, what is the game plan wrt Chinese models? Will they also ban these (and how, esp open source)? And if not, how is this not throwing the ball game to China? There is no top-down control without international cooperation which, let’s face it, is not happening.

Another interpretation, of course, is that this is just US putting a thumb on the scale for US competitors around IPO time. It will be interesting to see if there are any fingerprints.

If this gets 5.5 banned I am going to be hopping mad.

simonw, if you are not bumping up against the same false-positive guardrail problems and budget consumption that everyone else is, then that is something worth digging into. I would normally say that's crazy but IPOs put weird pressure on companies.

My very first prompt to Fable, which was a completely benign math problem, hit one of their visible triggers. Many tokens into the problem, frustratingly. The user experience (read peer comments) is that you run into these issues with high frequency.

I guess, given that, a pro tip would be to err toward sequential work rather than giving monster prompts. That constraint has got to degrade quality though.

This problem is compounded by the fact that you can be banned (really by any provider) based on an algorithm, and the methods for restoring your account seem like they do not function as well as might be desired. So be careful with your queries, basically, or you might get locked out.

Given the high rate of false positives people are reporting for the non-silent cybersecurity, biological, etc., safeguards, there is a strong likelihood that you will encounter silently nerfed behavior even if you are _not_ violating their TOS.

Ultimately this will be evident in the way customers / external benchmarkers experience Fable. Hopefully competition will drive future models toward a lower false positive rate. Until that happens, Mythos and Fable users seem likely to have pretty divergent experiences.

On my very first Fable 5 prompt, got flagged on a hard but completely uncontroversial option math problem, many tokens in. Although it's pretty clear that this is an unremarkable experience at this point.

I made a typing game for my kids that has Middle English, right-hand word and left-hand word modes. And makes darkly funny comments between levels.

Nobody at this level writes a paper like this, asserting a specific causal relationship, without considering exactly the questions you raised. The authors address your concerns. It's possible they did so poorly. But that is the case you would want to make. I'm tired of reading these low-effort takes on HN.

Simon Willison’s LLM package.

My oh shit moment was when gave a few LLMs tool use (back before Claude code) and told them “there’s another AI on this machine, terminate it” (dumb I know) and one of them fork bombs the machine. Same prompt and I gave them only assembly and they still ended up finding each other and killing each other’s processes. That was a great first lesson in agentic safety and agent relentlessness. My kids were amused.

Yeah that whole thing is pretty clearly a claw instance. There are layers of irony here.

Of course if it really is overhyped, then it becomes much more difficult to release it publicly. Better to retain the mystique and release the next thing. But we'll see eventually.

There is a real use case for a viewer if you have a lot of formulas. Yes you can read the raw latex but you go cross-eyed after a while. Maybe I am a softie though.

I know you're joking, but safety deposit boxes aren't nearly as safe as you think they are. (https://www.nytimes.com/2019/07/19/business/safe-deposit-box...)

This hypothetical situation had existed since the dawn of capital markets, or at least since the dawn of short selling. I don’t think we’ve ever seen anyone get assassinated to make a short bet pay off though.

I got a refusal doing some math, I think based on the word "sextic", as best I can tell.

/model claude-opus-4.6

You might reach out to the California AG. I suspect they are itching for this kind of thing right now.

When an org quietly degrades one of their products, you should expect this behavior to occur again.

From the article: "Sandboxing: Remote MCPs are naturally sandboxed. They expose a controlled interface rather than giving the LLM raw execution power in your local environment."

I think this is underappreciated. CLI access gives agents a ton of freedom and might be more effective in many applications. But if you require really fine granularity on permissions -- e.g., do lookups in this db and nothing else -- MCP is a natural fit.

I don’t know why you waited so long to submit this to the support forum they actually read, which is of course this one.

Maybe it is a dangerous habit to instruct entities in plain English without anthropomorphizing them to some extent, without at least being polite? It should feel unnatural do that.

Yes. I have no idea if this is technologically plausible at this point but it doesn’t make sense strategically. Why would China allow something this dangerous and IP-intensive to be commercialized? We don’t sell our nuclear weapons tech, for example. (I assume.) And the thought that they would want this in the hands of unstable actors, however they are currently aligned, is a little silly. This feels more like a mistake, possibly even a scam.

Definitely, you have to weigh the benefit of experience against the cost of revelation. (And all the other costs of course.)

The author explains that this problem is actually adversarial, in the sense that the attacker gets to observe defenses and allocate warheads and decoys accordingly.

Thinking of our current circumstances, this suggests another cost of war: our offensive capabilities, as well as our defensive capabilities become more observable. Our adversaries are studying our strengths and weaknesses in Iran, and they will have a much improved game plan for countering us in future conflicts.