nighthawk454

Usually, we destructively compress (mean-pooling) both the query and the document, and then compare the two compressed forms.

With ColBERT, we compare first - at the full token level - for more detailed comparison. Then reduce the full set of comparisons to a single vector. Naturally this takes more memory and compute to do the more comparisons. The idea is it’s worth it because the more detailed comparisons lead to better results

tokens —> reduced vector —> comparison

Vs

tokens —> comparisons —> reduced vector

Contrast ratios don’t go away in light mode. The contrast ratio of black-on-white is the same as white-on-black.

Curious then what is the real driver of these issues? If accessibility teaches us anything, it’s that there’s certainly no one size fits all. But it’s gonna be hard to update the standards or make plugins/tweaks successfully unless we get at the root of it.

No, no, the arsenic was in the wallpaper. Plenty of other things in the food like alum (which causes constipation actually)

Presumably it's a spin off of Google's 'Model Card' from a few years back https://modelcards.withgoogle.com/about

Easy to try here: https://huggingface.co/spaces/NexaAIDev/omnivlm-dpo-demo

https://i.imgur.com/44XYyXU.png

> In the fall of 1981, three of the four children followed their mother to the grave. According to Peskov, their deaths were not, as some have speculated, the result of exposure to diseases to which they had no immunity. Both Savin and Natalia suffered from kidney failure, most likely a result of their harsh diet. But Dmitry died of pneumonia, which might have begun as an infection he acquired from his new friends.

“That’s the name of the movie!” - most Pitch Meeting videos

Hey, just wanted to say nice app :)

Lotta people talking about broader app/subscription ecosystem issues, which hopefully you can take as off-topic to your specific project. Making a cool useful thing and putting it on the internet is great, cheers.

That would be kind of neat. Bc realistically the marginal costs on most digital things is negligible. But if it were practical to charge people the 1e-11 dollars per page view or whatever maybe could do some interesting things

Repo: https://github.com/lmcinnes/datamapplot_examples

Also, lmcinnes is the author of UMAP and HDBSCAN!

That's... not the user's problem. This is a fine and cool project don't get me wrong. But the overall 'subscription everything' model is not really justified by costs. The subscriptions are usually orders of magnitude more than the true operating cost.

It's not the customer's job to pay you forever bc Apple wants a developer license. It's the business's job to make sure it's sustainable with the costs that it has / has chosen to bear.

That's the backpressure on business models - they're not all viable. Just because you _could_ add in a bunch of servers and cloud costs and whatever, doesn't mean it's inherently justified.

The problem is more that it's gotten _so_ cheap to run, that charging each user a seemingly-nominal 5c/day fee doesn't feel bad to an average person for a chance at value. And at scale you get enough people who figure "ah it's not that much", and end up with massive profit margins. Profiting off the disparity between the individual choice and the aggregate.

The error lies in thinking they're 'real' systems to be taken for granted and blindly reasoned forward from, instead of sometimes-helpful academic categorizations.

You can always factor things into groups. e.g. 'Thoughts about now vs thoughts about the future'. Extending that to say there are therefore two modes of thinking and that the brain must handle your two groups differently, at some fundamental or physiological level, and there are only these two modes and all things are either one or the other ... is perhaps quite misguided without more support.

"It's only a model"

Oh thanks, I didn’t realize that. Having recently evaluated Vispy that really helps me place what this tool might be for

I mean that seems a bit uncharitable. I don’t think the author communicating their intent behind the name is aggravating or manipulative. It’s simply an explanation. Shouldn’t be worth more than an ‘ah, I see what they’re going for’ and move on.

It’s a sign, not a cop. Doesn’t seem like any undue pressure to control people, or prevent the reader from doing what they will with.

One wonders if the IPA pronunciations on Wikipedia is similarly bristling

title is missing a " Presents " in the middle

Not if you’re throughout limited with $1 games

A few links through, this Ship Investor simulator is a great super-simple game to test some intuition: https://xkqr.org/ship-investor/ship-investor.html

Seems like the terminology of calling it a ‘true’ belief led to some confusion. Of course there is an huge difference between evidence and proof. Correlation is not causation, Godel’s incompleteness theorem, all abstractions are leaky, etc.

Desperation to ‘know’ something for certain can be misleading when coincidence is a lot more common than proof.

Worse yet is extending the feeling of ‘justified’ to somehow ‘lessen’ any wrongness, perhaps instead of a more informative takeaway.

Iiuc I think it does work, but the only guarantee is the table will be stable. It very likely will not be _level_

Hey cool article, thanks! Might be time to finally dive in to DuckDB

^ this one was posted by OP, maybe dedupe to that

Here's another way to look at it. With a naive binary search strategy, which is an optimal algorithm for search for a sorted static array, we have:

Step # | % of reachable numbers in [1,100]

1 | 1% = 2^0

2 | 3% = 2^1 + 2^0

3 | 7% = 2^2 + 2^1 + 2^0

4 | 15% = 2^3 + ...

5 | 31% = 2^4 + ...

6 | 63% = 2^5 + ...

7 | 100% <-- 2^7 > 100

So out of all possible single-trial outcomes, only 31% of outcomes guess the number in <= 5 steps. In other words, in 1 trial you would expect a 69% chance of a non-positive outcome.

So an EV of +$0.2 or +$0.07 alone does not match the actual odds of winning in 1 trial. EV is most predictive in the infinite limit, and least predictive in a single trial - which this is. First off, $0.2 isn't even a possible outcome so there's a good reminder that the mean value does not always occur in the dataset.

This is also pretty straightforward from the classical interview-y 'Big O' perspective. If you translate the question to the natural CS-worded equivalent, something like "can you search a sorted array of length 100 to find an arbitrary value in 5 steps or less?", one readily sees that O(log2(n)) -> log2(100) = 6.6 > 5, so you definitely can't guarantee it. If we put odds on it as above, we can see that the odds are also not in your favor.

Now, if you want to look at it as saying after 5 steps you've removed ~96% of the search space, that's cool and all that you've reduced the candidates to only 3-4 remaining numbers, but those aren't the odds of winning. We know that 7 guesses is enough to guarantee finding the number, so after 5 guesses we ought to be 'close'. But the game is not horseshoes, so the fact that we're close is not relevant

hey, thanks for the blog post and your reply! I think I follow - a generalization of a coin-flip type game. I agree that if you have more starting money, you would never lose. From the binary search idea, even if chosen adversarially, the worst case is still log2(100) ~= 6.6. So if you get 1,000 guesses or just any number of guesses >= 7 you literally can't lose. Then you should definitely play.

Setting the limit at 5 brings you to the interesting point of there being a good mix of win/loss outcomes. 4 would be too few guesses and you'd very likely lose, and 7+ you'd definitely win. So the question is only interesting _because_ the limit is chosen so that the spread puts your odds on both sides of the 0 line. Otherwise it'd be clear cut.

The standard deviation being ~$1 is interesting. To me that suggests that with a mean of $0.07 and a deviation of +/- $1, it's essentially 50/50 odds. There's technically a slight edge in your favor, probably 53/47, but barely. So given a game with essentially no edge, would you play? Framing it that way - deciding to what degree the game is winnable - it's essentially not. You should not particularly expect to win, no matter your strategy.

I think part of the trick with the Ballmer question as well is the question is not necessarily about 'can you find an optimal strategy?' - it's 'do you play the game or not?'. The paths chosen within the round don't ultimately matter to that question. It's only intermediately necessary to model the intra-round decision paths in order to get to the overall win/loss distribution for a single round.

If you do end up getting the time, do make another blog post!

It's a good question, sort of related to optimal bet sizing. Check out this wiki which another commenter also mentioned: https://en.wikipedia.org/wiki/Kelly_criterion?useskin=vector

Yeah, exactly - you wouldn't stake your life on hitting the average case in an n=1 scenario. Parkour community sounds really cool!

That is an interesting question. I suppose it's equivalent to the 'worst case' performance for binary search, which would be a relevant topic. Finding the optimal strategy in the case where the opponent _may_ be adversarial could be interesting. I don't imagine the odds improve compared to the base situation, but not sure

People are really zeroing in on the word 'bankrupt' here. The point was if you used that strategy, all in 100% of the time for positive EV, you will _probably_ go bankrupt even though the n=infinity limit of that strategy is positive.

The whole poker thing was merely an analogy in the first place.

> If the question is "would you play this game" the reasonable mathematical translation is "determine if the expected value is greater than zero".

Not really! And that may be the point of the question. It's not testing if you can pattern match to plausible CS concepts.

If you get one play, and the goal is to win, do you take the chance? The whole question is about the difference in likelihood in the limit (expected value, infinite plays) and what is a likely outcome _of one round_.

You could? There's no reason you couldn't do $50,000 with bets of $10,000. The point is you get 5 guesses before you lose. The bet sizes don't really matter.