> They don’t just lean into epsilons, the session context tolerance is used
for almost every single point classification operation in geometric kernels
and many primitives carry their own accumulating error component for downstream math.
The GP wasn't wrong. To "lean in" means to fully commit to, go all in on, (or, equivalently, go all out on).
I think his point is: rather than "leaning into" it as in, masking through epsilons, he argues that tolerance is fundamental to the problem space, not a way to resolve edge cases.
To me it seems it's used with the intent "They don’t just <do X>, they <do Y>," implying that Y is a proper superset of X. My point is that X is in fact a superset of Y, making the most charitable reading "They don’t just <do X>, they <do X in more words>."
Is there another potential reading of "just" that I'm missing?
I'm still waiting for the update that explains why, after going from clock speed to FLOPS, etc. (which I understood) we suddenly started measuring processing power in GW. It makes as much sense as measuring automotive motors in decibels.
Speaking in terms of GW makes sense when we are discussing performance of an individual datacenter since power is, kind of, the limiting factor.
Also, FLOPs per watt hasn't changed as much lately, so thinking in terms of watts over a few-year time horizon does at least give you a ballpark of how many FLOPs.
If you're talking about a chip, you obviously want to know about FLOPs moreso, but, even down to the level of individual rack-units, wattage is a serious concern. Not every facility is built for these crazy 200kW racks.
Yeah, I get that. But using something that should be on the "costs" side of the ledger as your metric still seems fraught. It feels like another LOC trap.
Evolution is survival of the fittest. That's not a tautology, it actually says something, namely that the traits which survive and thus propagate tend to be the ones that enable some form of adaptation to its living conditions to the individual. The paper lists a bunch of examples:
- lactose tolerance
- immunity and disease resistance
- lighter skin at northern latitudes
- metabolism and vitamin D processing changes in response to changes in diet after the rise of agriculture
All these traits go beyond just increasing the odds of survival, they improve the life of the individual directly. I.e. they confer fitness. Individuals carrying those traits will, on average, in that ecosystem they are inhabiting, be more healthy than those who don't.
It's more reproducing than surviving. If the population of some species increases and the number of copies of some allele remains constant we could consider than gene less fit than the other alleles, in the population genetics sense. So it's frequency rather than survival that geneticists look at. But that proves that there are indeed other ways that they could have defined fitness if they wanted to.
Well that's a given mate, nobody's talking about populations that dont reproduce are they? And even then you're still making the same tautological mistake.... If they survived they must have been fit
The flip side is everything is being degraded by random mutation.
It's like holding a large ball in place on a hill that sees frequent tremors. If the ball is still halfway up the hill it's being held in place, if it's being held in place it's still halfway up the hill. It might be considered a tautology if you're only working with symbols and ignore all the mechanistics.
Remember, all improvements are changes, but most changes aren't improvements. The trick that makes evolution works is this: out of lots of random changes, most of which are harmful, the harmful ones tend to be weeded out and the useful ones tend to spread.
Whatever does not survive stops registering in later times; most of the time, what helps survival is retained, and what helps survival is what increases fitness.
As stated, this feels wrong. Specifically, it does not account for traits being appropriate for environment. I like to say it as what was needed for one stage could be the problem for the next stage.
That is, traits that stop registering may no longer be something that helps survival. But that does not mean they were not necessary for survival at an earlier point.
How exactly does that contradict the concept of fitness?
Several examples from the paper are exactly that. E.g dark skin was better for survival in Africa, but as populations moved north light skin was strongly selected for. Given the levels of sunlight in Europe, lighter skin increased fitness.
It is against the idea that the beneficial traits will survive to the present. It could be that there was some trait/gene that was absolutely needed for survival in the past, that flat out became irrelevant and dropped off before the present.
That is, it is not an argument against any of the traits that are present. Is why I said the problem is with how it was stated. But you do not have everything with you to provide evidence for all of the things necessary for you to have gotten here. At best, you have evidence that nothing you have with you prevented you from getting here.
That make sense? I grant that pulling it back up, I see the comment I was responding to was hedged. My concern is largely against the idea that things that "were selected for" in the past can be determined by evidence. I'm not convinced it can't be. But I find this presentation of it to be somewhat weak.
More to the point, TFA is specifically addressing the issue (which is part of what makes it a big deal).
They aren't saying "we see these things now, so they must be good" but rather things like "we see these selected for from 9kya to 3kya, but from then to the present they were selected against"; they are specifically looking at how apparent selective pressures changed over time.
> the idea that things that "were selected for" in the past can be determined by evidence
When the evidence is a copious selection of ancient genomes, distributed over both space and time, they certainly can be.
Apologies, I only meant my gripe with the comment I was responding to. Is why I put "as stated." I meant that to be that I was not arguing what I think they were messaging towards.
The callout on "evidence" I have there is that I meant that to only be present evidence. And again, I am not convinced it can't be done. It takes a lot of work. Which, the article is doing. But just saying that traits that helped you survive are typically retained, so by definition increase fitness, does not.
According to the Ofcom regulation checker [1] (linked to by The Register article), the Online Safety Act does not apply to this content.
Here's the most pertinent section (emphasis mine):
> Your online service will be exempt if... Users can only interact with content generated by your business/the provider of the online service. Such interactions include: comments, likes/dislikes, ratings/reviews of your content including using emojis or symbols. For example, this exemption would cover online services where the only content users can upload or share is comments on media articles you have published...
is this legal advice you are offering, as someone practicing law in the uk? because you are all over this thread stating your opinion very confidently.
(conveniently, there is no risk to yourself if you happen to be wrong or misinformed.)
No, I'm not offering legal advice, and neither am I stating an opinion. I'm simply quoting Ofcom, the regulatory body responsible for overseeing this law.
A valid point, and maybe I should have phrased it differently. I've deleted the comment which used the word "misinformed", so as not to cause any confusion.
My point is simply that the Ofcom quote clearly states that user comments on an article are not subject to the Online Safety Act. I assume this is a fact, as it's from the horse's mouth.
Some people appear to be basing their opinions on the assumption that the OSA does apply to such comments (hence my use of the offending word).
>Please note: The outcome of this checker is indicative only and does not constitute legal advice. It is for you to assess your services and/or seek independent specialist advice to determine whether your service (or the relevant parts of it) are subject to the regulations and understand how to comply with the relevant duties under the Act.
I mean even the site itself says it really shouldn't be used for legal advice...
On top of that, none of this matters until said law is settled under a case. Most often it's the first judge and the set of appeals after that point that define how the law is actually implemented. Everything before that is bluster and potential risk.
So the story is... a publication that opposes the party currently in power, quoting a few people from the side that's presently out of power, saying that their being out of power is really bad, and we may never recover?
How is this different than the whining we get when the roles are reversed?
I realize you folks hate each other, but it would be nice if either of you could talk about something without turning it into a rant about how great, noble and good your side is and how awful the other side is.
To someone neutral (yeah, humor me), the Trump administration has done far more to demolish the reputation of the US than any other administration in my lifetime (OK, maybe Nixon - I don't remember all that much about him firsthand).
But I would also say that Biden, while not as bad as Trump, was worse than anybody since Nixon.
Which of Biden's policies and actions did you find worse than any since Nixon? And where do you rank the Iraq debacle that Bush started? How about selling arms to Iran to fund the Contras in Nicaragua?
Remember what we're talking about. It's not about their policies per se, it's about what they do to the US's international reputation.
So what did Biden do? The botched withdrawal from Afghanistan was the biggest thing. But his own frailty didn't help (speech fumbling and falling on stairs). Yeah, I know, his personal frailty shouldn't affect the US's reputation. But I think it did.
I mean, yes, the fact that we were leaving at all is due to Trump. (Either credit or blame, depending on whether you think we should have stayed there.) But the absolute debacle of how we left is on Biden. And it's that debacle that tarnished the reputation of the US.
There isn't enough training data though, is there? The "secret sauce" of LLMs is the vast amount of training data available + the compute to process it all.
This is essentially a distillation on the bigger model; you'd wind up surfacing a lot of artifacts from the host model, amplifying them in the same way repeated photocopying introduces errors.
> Catching an LLM hallucinating often takes a basic understanding of what the answer should look like before asking the question.
We had the same problem in the early days of calculators. Using a slide rule, you had to track the order of magnitude in your head; this habit let you spot a large class of errors (things that weren't even close to correct).
When calculators came on the scene, people who never used a slide rule would confidently accept answers that were wildly incorrect (example: a mole of ideal gas at STP is 22.4 liters. If you typo it as 2204, you get an answer that's off by roughly two orders of magnitude, say 0.0454 when it should be 4.46. Easy to spot if you know roughly what the answer should look like, but easy to miss if you don't).
We do know. There have always been ways that people could avoid the painful process of learning, and...they don't learn.
Here's a competing thought experiment:
Jorge's Gym has a top notch body building program, which includes an extensive series of exercises that would-be body builders need to do over multiple years to complete the program. You enroll, and cleverly use a block and tackle system to complete all the exercises in weeks instead of years.
Playing devil's advocate here, but in theory, you could claim that setting up harnesses, targets, verification and incentives for different tasks might be the learning that you are doing. I think that there can be a fair argument made that we are just moving the abstraction a layer up. The learning is then not in the specifics of the field knowledge, but knowing the hacks, monkey patches, incentives and goals that the models should perform.
https://worldstats.io/clock
reply