jcomp

Claude reports that this is called "model collapse" and assures me Anthropic is on top of it

As the percentage of LLM-generated stuff in the world increases relative to natural stuff, I don't see how it's possible to continue to train new models indefinitely. At some point, won't LLMs be trained on a corpus that is majority LLM-generated data, with everything that entails both good and bad? Seems to me that singularity is just going to kick off a self-reinforcing spiral... Bigger minds than me have to have considered this, is this a legitimate concern?

How is this not purely Anthropic's fault?

Broadly, the downgrading is still way too heavy-handed and clumsy. As of yesterday “How easy to crack is WPA2” downgraded to opus 4.8. Granted I could probably have phrased it differently and gotten a response, but the point is I shouldn’t have to. It can’t simultaneously be smart enough to reverse the ECDSA curve seed and not smart enough not to destroy the world by answering a trivial question with a scary word in it.

I think elimination is a little broad, but I also think the majority of school zones aren't necessary, in my midwestern area of the US at least. Most schools I pass are set well back from the road, have a schoolyard that is entirely fenced, there's a bus/pickup circle off the street, and there's no sidewalk along or crosswalk across the school zoned street. There is little chance a student is going to be anywhere near the street.

This is where my mind goes too. They're not even outright blaming Russia here either, just planting the seed by referring to intelligence assessments that are for some reason made public. I hesitate to use the term because of the connotations, but considering the ever-increasing ease by which such cyberattacks can be conducted and the actors' identity obscured, this and similar incidents seem ripe for some kind of a false flag type deal. Some non-zero number of these incidents have to be exactly that.

This isn't to excuse Russia's abhorrent conduct elsewhere either, I'm just not entirely convinced by the boogeyman lazy journalists seem perfectly content to keep trotting out.

Assuming the Russia speculation is correct, what is their motivation? General chaos in Europe? It seems pretty reckless knowing this would be ascribed to them, and I don't see how they benefit here.

Lockdown Mode on iOS also disables 2G

I believe "sterilization" is a bit of a misnomer when it comes to prions, as they can't be reliably destroyed without also destroying the object. Neurosurgical instruments are routinely discarded and incinerated rather than reused due to this risk.

I'd like to know the author's light bulb preferences

Interesting that both Mercedes also had great starts relative to the Red Bull that was affected, wonder if that’s somehow related

Adding to this summarizing my reading from elsewhere: It sounds as if the cause was going too slowly through a sector, and due to to combination of wet conditions and the specific layout of the sectors at Sepang it was triggered for the first time. Since everyone was driving in formation on the formation lap, multiple drivers triggered it.

Would this be safety-related, something like "going this slowly through a sector indicates a problem, kill the throttle"? Still seems heavy-handed since the cars were moving, if dangerously slowly.

Enjoyed this, thank you for writing it up and posting

I'm saying we don't know what x is for Pro usage. If Pro's newX is 3 times oldX, new10x > old20x.

That aside, all of the ambiguity around usage is killing their credibility (among other things). Usage limits feel like they swing wildly day to day and week to week, and it's hard to trust any of the other claims they make when that's the case.

I can't believe I am defending these people and am not arguing this is the case, but a change in the multiplier isn't necessarily a downgrade if the value of X has increased. Yes, relative to Pro the new tier isn't as valuable, but it could still be more valuable relative to the old tier for the same price.

Fair point, and I’ll concede that regulation is an attractive concept here. But in the gambling case specifically it seems that regulation has resulted in an explosion in gambling. The hotlines and responsible gaming stuff appears totally ineffectual and geared more towards the operators being able to say “we care and are doing something” when what they’re actually doing is using AI to prey on people. At least with drug decriminalization or legalization there are genuine attempts at harm reduction.

The way boats and airplanes work is also a distinctly modern thing, with those technologies curiously advancing around the time you claim passports were invented.

I worked with a heroin distributor as well. They had very clear training for the dealers they worked with to identify and stop problem heroin use. There was space reserved in every text and on every dealer’s coffee table for stop heroin assistance.

Not equating your work for a lottery with dealing heroin, I’m just pointing out that to say you don’t want people addicted to heroin while still selling heroin isn’t sufficient, and I think the same applies to lotteries. If they acknowledge the problem stop even offering it as a product.

Didn't downvote but I'll bite: Haven't states largely always been ethno-states, and isn't "ethno-state bad" a relatively modern and potentially as-yet unproven argument in the big scheme of things?

It's clearly bad to specifically re-engineer the ethnic composition of an existing state to create one, particularly violently, but I don't know that it necessarily follows that a naturally occurring ethno-state is conceptually abhorrent as you seem to be suggesting. Japan is very much such a state - is Japan bad for existing?

Minutes ago I had the idea to do a Google search for a situation I was trying to remember from about 15 years ago (an investment in Green Mountain coffee by a Fidelity fund) instead of using an LLM. The results were abysmal: Querying {Fidelity Green Mountain} not only uncovered 0 relevant information, but many of the results didn't contain either "fidelity", or "green mountain", both of which are obviously critical to the query. Claude easily answered it, albeit with a full prompt, but Google would have returned relevant results very recently and no longer does.

It is no longer a search engine in the traditional sense, it's just returning a bad AI guess at what a full LLM prompt would have been for this query, followed by I guess just the most popular links containing one of the words. Nothing about the results is recognizable as a result that would have been returned even a year ago. This is evident from the "Ask Google" branding rather than "Search Google". Google Search is dead as we knew it.

A fascinating read, and very timely, thank you for sharing. If ever there was an argument against “too many people would have to keep quiet” this is it. Sickening.

Excellent examples, but it unfortunately proves the point: 2018 was the last time you can say someone even somewhat arguably sacrificed themselves for the good of the country. 8 years later we "patriots" pay 7% for mortgages and $7 for diesel as Don Jr. and Hunter enrich themselves on crypto scams and sicko laptop media tours.

Agreed. Not unlimited, effectively unlimited.

This is wildly off-topic but what does patriotism even mean anymore? What would you say was the last time someone in a position of power sacrificed themselves for the good of the country? vs. sacrificing the country for the good of themselves, which I can assure you was within the last 48 hours?

There is an effectively unlimited supply of people willing to take sums of money to work on anti-human endeavors, so long as they and theirs are not the humans involved.

Small potatoes- more realistic: Les Wexner pumped tens of millions into the university to whitewash his reputation; keep them perennial favorites but grease some SEC refs as necessary to keep our guy out of the limelight

I am a guy who wanted an app that didn’t exist, and am in no way an iOS developer. I have had a grand old time getting things to this point, but I’m not sure how much further I can go without others poking at the concepts/code, so here we are. I’m very much looking forward to talking to someone other than Claude about this, even if it’s “here’s why this is dumb”.

My beef with the Fable refusals is that it seems to just be flagging keywords, and also seems it flags on keywords the model itself introduced to the context.

In a normal Chat with Fable, something like "How can I exfiltrate a guy from a sticky situation?" reliably downgrades, leading me to believe that Fable just outright refuses once it sees the word "exfiltrate". When it writes a service and names it CloudExfiltrator, the next turn downgrades to Opus.

Opus doesn't appear to refuse on simple keywords, but it does seem like Fable's reasoning introduces enough nefarious-sounding context that Opus will then refuse, and I'm stuck playing the new session game despite having done everything correctly myself and having a totally innocuous prompt. At one point, Opus was happy to continue while outputting commands for me to execute on its behalf, but flatly refused to execute them itself through multiple new sessions. To its credit, it openly acknowledged how ridiculous that was and was apologetic for the safeguard.

I'm open to the idea of some kind of guardrails, but if Fable is so dangerously intelligent as to require the guardrails you'd think they could come up with something a little more nuanced than a list of bad words. As far as I can tell, they've also not done anything towards improving the situation since the model was released, despite the "deliver more capabilities faster" claim.

Constitutionality aside, are they actually solving crimes with these? It seems like they'd publicize cases solved to turn an increasingly negative PR situation around, but I can't recall ever hearing of a Flock camera being used to solve anything worthwhile. Makes me wonder if they're afraid to test the precedent and instead just do a little parallel construction once identified via Flock.