lukev

27 ms·

451.
by lukev 2y ago
Sure. Your claim has some truth, but is far too strong. The articles you cited upthread do not support the notion that models consistently activate differently when generating true facts vs false facts. It is true that models can capture so
452.
by lukev 2y ago
Upon a second reading, this is an excellent point. For the sake of clarity, let's remove LLMs from the equation and posit the existence of Encyclopedia Eric. Ask Eric any question, and he will happily research it and come back to you
453.
by lukev 2y ago
This is confidently stated and incorrect.
454.
by lukev 2y ago
I do think trusting an LLM is less firm ground for knowledge than other ways of learning. Say I have a model that I know is 98% accurate. And it tells me a fact. I am now justified in adjusting my priors and weighting the fact quite heavily
455.
by lukev 2y ago
You misunderstand. I mean that the model itself is doing exactly the same thing whether the output is a “hallucination “ or happens to be fact. There isn’t even a theoretical way to distinguish between the two cases based only on the inform
456.
by lukev 2y ago
They don’t just seem it. They are by design. We talk about models “hallucinating” but that’s us bringing an external value judgement after the fact. The actual process of token generation works precisely the same. It’d be more accurate to s
457.
by lukev 2y ago
I feel like there's some semantic slippage around the meaning of the word "accuracy" here. I grant you, my print Encyclopedia Britannica is not 100% accurate. But the difference between it and a LLM is not just a matter of de
458.
by lukev 2y ago
It also lets a language model answer questions while citing a source, something it fundamentally cannot do on its own. Everyone talks about "reducing hallucinations" but from a system perspective, everything a LLM emits is equally
459.
by lukev 2y ago
What are you objecting to? The way the journalist is using tone and centering their subjective feelings? Or the account of how some government and industry leaders regard large-scale death as a profitable enterprise and are entirely unconce
460.
by lukev 2y ago
Big if true. Do you have any kind of evidence that you can share for this assertion?
461.
by lukev 2y ago
I mean, if you think of this as an "entity dating on your behalf" it's weird and dystopian. If you consider it as a somewhat novel matchmaking algorithm, I'm not sure where the problem is and could actually be kinda cool
462.
by lukev 3y ago
In the words of Charles Babbage, "I cannot rightly apprehend what confusion of ideas would lead to such a question." LLMs (by themselves) cannot reliably count. If you expect them to, then you're falling into the common trap
463.
by lukev 3y ago
What is the basis for your assertion on how the court is going to rule?
464.
by lukev 3y ago
I'm a Kagi user. None of this is relevant to me except the allegation that privacy isn't taken seriously. And yet Kagi's terms of service clearly state that they do not even log user searches. So unless the assertion is that
465.
by lukev 3y ago
Yeah, when AIs can comprehensively cite their sources I might change my opinion on that. Though note that there still isn't any need to publish static content. The power of LLMs is that they can be dynamic and responsive! Even if we
466.
by lukev 3y ago
I can see this being useful iif the content is generated on demand and then discarded. Publishing AI generated material is generally speaking a horrible idea and does nobody any good (at least until accuracy levels get much much better.)
467.
by lukev 3y ago
I have, on multiple occasions. Unfortunately mostly on contracting gigs so I can't cite specifics but I've worked on a major ecommerce platform's api, and two separate databases using this approach. For public examples, I&#x2
468.
by lukev 3y ago
This is why I'm neutral on the value of unit tests. Sometimes they are helpful, other times less so. I am always in favor of tests around the boundaries between system components (regardless of what you call them; integration tests, AP
469.
by lukev 3y ago
Given a typical professor's salary, it is... difficult to imagine that this is true, at least not without fantastically bloated administrative and overhead costs.
470.
by lukev 3y ago
Opsec is hard. Also. No matter how bullish you are on the potential of AI, it's horrifying and categorically unethical to use them to decide who to kill. It creates the possibility of atrocity with no accountability. Although the fact
471.
by lukev 3y ago
Separately: we should be careful of carrying the right wing's water and adopting terms like "DEI" as negatives without thinking critically about it. This is straight out of the Chris Rufo playbook; identify a well-intentioned
472.
by lukev 3y ago
I do not believe anyone with real power at Google values liberal values above revenue. The RLHF and fine-tuning is solely to prevent news articles of the form "Google unveils new shockingly racist AI!!", which is absolutely what w
473.
by lukev 3y ago
The quality of these results is in my experience quite poor, so this is worrying to me. (The other comment about "distilled blogspam" hits the nail on the head I think.) It highlights a fundamental tension around Google's cor
474.
by lukev 3y ago
This doesn't pass the sniff test for me. Not sure if these models are memorizing the answers or something else, but it's simply not the case that they're as capable as a domain expert (yet.) I do not have a PhD, but in areas
475.
by lukev 3y ago
LLMs are language models. Not knowledge models. That's a tremendous breakthrough! Language is really hard and we've basically "solved" it computationally. Incredible! But whether via retrieval or some other form of datab
476.
by lukev 3y ago
Often (but not always) they fall less victim to the “loudness wars” compared to the same track mixed for stereo.
477.
by lukev 3y ago
Because the site of Worldcon is selected by members (i.e, sci-fi fans) of which China has _many_. In fact, this whole debacle has hurt them the most since it seems many/most Chinese authors were excluded by default since the amateur se
478.
by lukev 3y ago
It's hard to talk about something so based in individual perception: it's true, my eyes could literally be different. However, my strong suspicion is that the visionOS native apps have a larger font size and bigger, smoother UX el
479.
by lukev 3y ago
I for one don't notice any quality/resolution difference between native apps and a mirrored Mac screen. I agree it feels "worse" than a physical screen - it is certainly less crisp. But I ran an experiment, and created a
480.
by lukev 3y ago
Eh... it's extremely good and more than I expected, but it's not quite 4k quality IMO. Not because I can see the pixels (I can't), but just because of very slight inconsistency across the field of vision, and motion blur when

More ›