@brandonlive.com
Architect for AI in Word at Microsoft. Previous: ~14 years working on Windows, with a stint as small startup CTO in the middle. I also made Tweetium (may it rest in peace). https://brandonpaddock.substack.com
Barbaric.
Why are messages limited to such a small size on Bluesky? It’s a lot more difficult to make reasonably sized posts/replies/threads here than Threads.
That said, I got your post/reply via a reply to a thread that was discussing an article written about the paper. That article (and post sharing it) were terribly misleading about what the paper says, and the replies were all echoing it, so my reply here may have been colored some by that context.
That is, one theoretical way to “solve” a lot of these problems would be to separate out all but the most fundamental knowledge from the LLM, leaving behind just the reasoning, cognition, and language understanding parts of the model - with knowledge learned, stored, and retrieved externally.
But some things can’t be generalized especially in a context-free way. For some non-generalizable facts, the training process can result in the model memorizing them, but this only gets you so far and many of us think is really an unfortunate side effect more than anything.
One example reason for this is that if you train on statements like “Today is Monday” on a Monday, or “the price of milk is $X.XX”, the accuracy of these statements isn’t true on a Tuesday a year later.
More importantly, the main goal of training is to generalize, not memorize.
More that I think you misunderstood it. The idea of “perfect” data referred to is just saying that if your training set has no false statements in it, your base (pretraining only) model will still be capable of producing false statements.
Nothing in the paper says anything about it being fundamentally impossible to solve these problems, though even defining the problem(s) and what it means to solve it/them is glossed over.
It’s unfortunate that so many people encounter misinformation like this on social media and believe it.
One nuance is that the pretraining process for *today’s* models is such that some types of “hallucinations” are expected (again, the paper pointed this out). But pretraining isn’t where we stop and base models are never used in production. Maybe the author of the article got confused on this point?
I never liked the term “hallucination” for a few reasons. People actually use it to describe a variety of behaviors with different causes. The paper the article references tries to generalize this but at least acknowledges many (though not all) of the nuances.
100% this.
This is not in any way what OpenAI’s research says. The quote here is just a lie, plain and simple.
I think there are some fundamental misunderstandings here, and the multiple variant uses of the poorly chosen “hallucination” term don’t help.
The “problem” is that LLMs are language models, and now more and more they are reasoning models, but they are not (effective) knowledge models.
Two plausible explanations:
1) Selection bias, the people who see it (and leave verified reviews which is what the score shows) are in the cult and were always going to give it 5/5 no matter what.
2) They just bought a bunch of tickets and spammed fake reviews. They did that with Don Jr’s book.
He’s desperate to be the coolest guy on the planet, and he knows there’s only one way that will ever happen.
lol.
Not really, if you know how these things work, it’s pretty easy to determine they cannot be conscious.
My bigger worry is that in trying to define consciousness, we’ll determine that it doesn’t really exist in humans either.
I don’t think anyone serious anywhere would argue that today’s machines are conscious. But it would be a good idea for us to try and figure out how we would know if they were.
I knew if I took a little time I could create a simple utility app that would do what I wanted, but I never find the time for little projects like this.
But using CC, I was able to get something really handy up and running in no time at all.
Claude Code really is bonkers. Last night I was trying to figure out a way to make it easier to play PC games on my big home theater screen. I couldn’t find a good tool or way to set up Windows to do what I wanted…
People like Musk believe that free speech is like money. He and his friends should have it, you should not.
Whatever this is, 0 stars. Do not recommend.
A fun thing about having a doctor for a father is that you can have a legitimate basis for worrying that the doctor prescribed you the wrong antibiotic (in this case, if the infection turns out to be MRSA).
I know it’s only been 8 hours but I would really really like it to start working now.
Trump and his ilk may not be smart, or honest, or even care about the particular direction they’re leading their followers.
But they do lead. And people hungry for someone to tell them what to think don’t even have any other options.
Mamdani is an all-to-rare example of a Dem acting like a leader.
I think a big part of the problem is that Democratic politicians don’t try to convince voters of anything. Their approach is to do polls to try and figure out what people already think, and then align what they say with that. And when their base is split, they equivocate.
That’s not leadership.
Another copy on YouTube until it is taken down:
youtu.be/IIm66Tsao8s?...
Here’s the 60 Minutes segment that Trump and Bari Weiss don’t want you to see.
Don’t look away.
In eastern Congo, where rape is widespread, the cancellation of USAID funding for life-changing — sometimes lifesaving — post-exposure prophylaxis (PEP) kits has left many victims vulnerable, according to nearly 50 interviews.
As is the case with Musk, and even Trump, the most grotesque part is not their existence — awful people have always been around and always will be. The most grotesque part is nominally smart or honest or decent people taken in by them. Or, at least, pretending to be taken in, for advantage.
Neither is “inferior”, just adapted to different environments. Humanity originated in Africa with early populations close to the equator, where dark skin is protective against more intense UV radiation from the sun. Far from the equator, light skin is advantageous for vitamin D synthesis.
