sparsevector

ChatGPT runs moderation filters on top of your conversation and will highlight responses or prompts red if it thinks you're breaking TOS. The highlight is accompanied by some text saying you can submit feedback if you think the moderation is in error. It's not very hard to trigger moderation--for example I've gotten a red text label asking the AI questions about the lyrics to a rap song with explicit lyrics.

It's interesting to compare ChatGPT moderation to Bing. When Bing generates a "bad" response, Bing will actually delete the generated text instead of just highlighting it red, replacing the offending response with some generic "Let's change the topic" text. The Bing bot can also end a conversation entirely if its a topic it doesn't like which ChatGPT doesn't seem to be able to do.

I know that revealing the features they use to spot AI submissions would just make their job harder, but I can't help but be curious what sort of features they are looking for. I frequently use ChatGPT to proofread things I write, and I find that in order to avoid "sounding like ChatGPT" it's best to just ask the AI "Is what I wrote clear?" I find if I focus on clarity, the suggestions it makes are pretty basic and unlikely to make my text sound robotic (e.g. fix this misspelling, define this term). If I ask the AI for suggestions on tone, I find it pushes my text too close to a generic "PR friendly" writing style that does not feel personal to me.

Agreed. Presenting this feature solely as a tool for users with cognitive disabilities might undersell its potential. There's a significant number of smartphone users who only utilize a small fraction of the available features and would prefer a simpler interface focusing on the 10% of features they actually use. Interestingly, this demographic includes both less tech-literate users, who might feel overwhelmed by the complexity, and extremely tech-literate users who know exactly what they need and prefer clean, distraction-free tools.

For me the magic of BOTW was the way its design encouraged you to approach the game more as an environment for free form play. My best experiences with BOTW were when I turned the HUD off and just enjoyed the world, finding Koroks, clearing enemy camps, but mostly just wandering. There are a lot of design elements that come together to make this play style really rewarding and fun like the density of Korok placements and the way you can easily spot shrines from a distance. I think it's hard to compare BOTW to OOT since they have such different designs. OOT is like a carefully constructed theme park ride where BOTW is more like an enormous playground filled with toys. But BOTW succeeded at giving me the same childlike sense of wonder I had with OOT (despite now being an adult).

I also tried this out and the balloon example as well.

User: Barbara, Mary, and Harry had red balloons. Ted had a blue balloon. Sam also had a red balloon. How many children have red balloons?

ChatGPT (GPT-4): Barbara, Mary, Harry, and Sam each had a red balloon. That means there are 4 children with red balloons.

User: What about Ted?

ChatGPT (GPT-4): Ted had a blue balloon, not a red one. So, he is not counted among the children with red balloons. Therefore, the total number of children with red balloons remains 4.

User: Why do you think I asked my original question?

ChatGPT (GPT-4): It seems that you asked your original question to test my ability to understand and process information provided in a text. You wanted to see if I could correctly identify and count the number of children with red balloons based on the details given.

User: Can you explain what mistakes a simpler model may have made on this question?

ChatGPT (GPT-4): A simpler language model may have made the following mistakes when answering your original question: ...

I like trying these sort of puzzles and "gotcha" prompts and I've found GPT-3.5 and lower versions are indeed not hard to confuse, but it's much harder to find simple examples that confuse GPT-4. Often times GPT-4 will even be able to explain the puzzle and why it's interesting. I think this is part of why GPT-4 feels qualitatively different. I understand (roughly) how transformers work, but GPT-4 does seem to have some emergent ability to reason, at least in the sense of making short chains of logical inferences, and this does seem to inform text completions not just on these sort of puzzles. That said I do think there's real danger of overconfidence and over reliance.

My take on btilly's problem was that the difference between the convergence rates is not constant but rather known to be bounded away from zero by some unknown constant. That is, there is some value epsilon such that the difference in conversion rate is on every round at least epsilon. However, at some rounds it could be more than epsilon, and it doesn't necessarily vary according to any fixed distribution. I think this is therefore roughly equivalent to the non stochastic setting additionally assuming the best arm has large (relative) reward.

I agree though that if the difference in conversion rate were stochastic that could potentially make the problem much easier.

Yes, I realized after posting that this is probably a better paper to link to: http://cseweb.ucsd.edu/~yfreund/papers/bandits.pdf The algorithms for the partial information setting are sometimes surprisingly similar to the algorithms where you see all the results. The algorithm in the paper linked above is essentially the same algorithm but with a small exploration probability. The regret bound gets worse by a factor of sqrt(N), however.

I think what you're describing is basically the non-stochastic setting. The idea there is to assume absolutely nothing about the rewards (they can be generated by quickly changing distributions or even an adversary) but still try to do as well as the best single arm. The standard algorithm for this problem is the multiplicative weights update method, also called the exponentially weighted average method or the hedge algorithm. It's a really simple method where you keep a probability distribution over the arms and then reweight the probabilities at each time step using something proportional to the exponential of the rewards of the arms. If you run it for T time steps on a problem involving N arms, it gives you regret O(sqrt(T log N)). This is actually the best possible bound without additional assumptions. This is a good paper on it: http://cns.bu.edu/~gsc/CN710/FreundSc95.pdf

There are many, many different variations of the method and result. For example, you can show improved results assuming different things about the rewards. For exapmle, you can show O(log(T)) regret with certain additional assumptions. This is a really good book on different non-stochastic online learning problems: http://www.ii.uni.wroc.pl/~lukstafi/pmwiki/uploads/AGT/Predi...

> I believe that I can produce a multi-armed bandit algorithm that has logarithmic regret and will work in the face of ever varying conversion rates, as long as there is a consistent relative difference between the versions.

That sounds interesting. As you undoubtedly know, there is a lot of literature on multi armed bandit problems and also on the non-stochastic multi armed bandit problem. The latter has the advantage of not assuming anything about the way in which the sequence of rewards is generated.

There is a a line of work in the non stochastic setting which sounds related to what you are describing. It's called the problem of tracking the best expert: http://www.cse.ucsc.edu/~mark/papers/track-long.ps The idea in this problem is to do well as compared not to the best fixed action but rather the best piecewise constant sequence of actions. There are many variations of this problem, but the basic idea is that you can have low regret as compared to a changing sequence of actions so long as that sequence doesn't change too "quickly" for some definition of quickly.

His numbers may be high but $40-50k/yr is pretty low. My impression is that software developer positions at top companies (Google, Microsoft, Facebook, etc) pay about 2x that.

I think the problem is more complicated than basic supply and demand. There is a lot of demand for college education. Enrollment is higher than it has ever been, and the people profiled in this article have jobs. The problem is that in order to cut cost colleges are hiring more and more adjunct professors as opposed to full-time, tenure track professors. It's not uncommon for these adjunct professors to take on jobs at multiple community colleges in order to make enough. The end result is that they are doing as much or more work than a full time professor but for less pay.

It's true you need to compute the kernel dot product between every example you see and every example in the support set (every example that ever previously evaluated to signed margin < 1). Whether it's online depends on your definition of "online" . It's definitely not online in the sense of using memory independent of the number of examples, since you have to keep around the support set. I think there are results showing the support set grows linearly with the size of the training set under reasonable assumptions. However, it is online in the sense that it operates on a stream of data, computing predictions and updates for each example one-by-one. It's also online in the sense that its analysis is based on online learning theory (e.g. mistake / regret bounds). A lot of learning theory papers use "online" in the latter two senses, which is confusing if you expect the former.

The set of support vectors is just the set of training examples that have non-zero alpha parameters. To implement the gradient update you just evaluate the support vector machine on the example (using the explicit kernel expansion) and then if the example has signed margin less than 1 you add y * eta to the corresponding alpha value.

The difficulty with Pegasos for non linear kernels is the support set quickly becomes very large and so evaluating the model becomes very slow. Note that since the alpha values are not constrained to be non-negative (unlike the standard dual algorithms) the alpha values don't ever get clipped to zero--instead they just slowly converge to zero. It's still (I think) one of the fastest methods in terms of theoretical convergence guarantees but perhaps not as fast as LaSVM or something similar in practice.

However, there's been a more general trend in machine learning to use linear models with lots of features instead of kernel models, partially because of these sort scalability issues.

Noctis does: http://en.wikipedia.org/wiki/Noctis

However, it's a bit... idiosyncratic?

This work sounds interesting. Who was the talk by? It sounds like a student of Jon Kleinberg but I quick glance over his publications didn't reveal work on that kind of mechanism design problem.

I think what VBprogrammer meant is that thinking you've found a bug in a compiler / OS / CPU is often a warning sign you're a poor programmer. Often times a beginner will have a bug in their code that is too subtle for them to identify, so they end up attributing it to some external factor. Actually finding a bug in a compiler / OS / CPU is as you suggest likely a sign you're doing something advanced or unusual and therefore are perhaps more knowledgeable than most.

It sounds like he's saying it's hard for Julia to interop with languages that don't support arrays of inline structs (e.g. Java). I could be misreading it though.

Having a very small (even one pixel) hit box for the player is actually sort of a tradition for this kind of game: http://www.significant-bits.com/the-1-pixel-collision-box

Having a safe zone in a boss fight is an issue though.

The problem is that standard statistical tests have two outcomes: (1) "reject the null hypothesis" or (2) "failure to reject the null hypothesis". Moreover in the most common statistical tests the null hypothesis amounts to something like "the two samples were drawn from the same distribution" so if you fail to find a significant difference you haven't shown they're the same, you've just failed to show they're different. If what you want to do is show that they're nearly the same, you can design a statistical test where the null hypothesis is instead something of the sort "these two samples were drawn from distributions that differ by > X amount" so rejecting the null hypothesis shows they differ by <= X.

It's impossible to know whether or not Linux would exist without GNU, but GCC isn't something you can hack together in a few months. Clang is the closest thing to a GCC replacement we have, and it didn't compile Linux until October 2010 (http://en.wikipedia.org/wiki/Clang)

In some communities open access journals are the most prestigious journals. In machine learning JMLR (http://jmlr.csail.mit.edu/) is (I think) widely considered the top journal. It was formed when a large group of top machine learning researchers grew fed up with the high cost of paid publication and together resigned from the board of a competing publication (http://jmlr.csail.mit.edu/statement.html). All of the top machine learning conferences (NIPS, ICML, UAI, AIStats, COLT) publish their proceedings online too. I think it's only a matter of time before all other communities follow suit, but some communities may lag behind longer than others. The machine learning community was lucky in that a large group of leading researchers decided to collectively make a clean break and support a new journal--the new journal became the premier journal almost over night so there was no slow transition phase.

This is true for some things. For example, UI design can be a long and difficult process for the first mover but easy for competitors who can just lift the design. However, it's not so easy to lift, say, the design of a distributed file system from its public interface.