fxwin

it's less about the interfaces and more that it feels disingenuous to compare token usage to a pi setup that injects a bunch of stuff into context rather than the base version. it makes me think they tested it with pi, found they couldnt meaningfully beat its token usage + pass rate, and omitted the comparison. i'd love to be proven wrong, but this seems like the easy explanation

why would you include oh-my-pi in the comparison but not vanilla pi?

for some reason this is really funny to me. it's like the "black museum" black mirror episode where a consciousness in a toy animal can only communicate using very primitive predefined responses

> not to mention the egregious target leakage

I was curious about this so I skimmed the paper [0]:

> SalesRLAgent achieved 96.7% accuracy, outperforming the best commercial alternative by 23.7 percentage points and the best LLM approach by 34.7 percentage points.

For a fuzzy natural language task like this, this magnitude of improvement should already set off alarm bells (Though i admit I'm not even sure what accuracy is even measured here, and the paper doesn't help either). Also, "best LLM" here refers to GPT-4 (at the time of upload, the public already had access to GPT-o3 and). I would have loved to contextualize the performance by looking at model size, but the paper is frustratingly devoid of detail in that regard:

> The core of SalesRLAgent is a reinforcement learning architecture consisting of: • A state encoder network that processes Azure OpenAI embeddings and features • A policy network that estimates conversion probability based on the current state • A value network that estimates the expected cumulative reward • A meta-learning module that assesses prediction confi dence

Also:

> Beyond technical metrics, we evaluated SalesRLAgent in real-world sales environments through A/B testing. [...] After 90 days across 217 representatives and 12,433 con versations, we observed: • 43.2% increase in conversion rate for the test group

This would be a pretty huge result but the fact that this is just shoved into a single paragrpah with no further discussion on methodology, baselines and setup makes me very suspicious.

[0] https://arxiv.org/abs/2503.23303

It's presumptuous for them to assume that a reader of their blog is familiar with their product?

Also I feel like the obvious way to read the very first sentence is that GLM is a language model

> As we develop GLM, the model sometimes exhibits capabilities that surprise us

That's fair, I had assumed they hadn't read the article at all ;) If they have, it would have been nice to know which parts of the problem definition the paper describes as "simple" they struggled with, otherwise I default to "I haven't read the article and would like a summary for a technically inclined layman"

I take ELI5 in hackernews comments to mean: "Explain like i'm someone with a vaguely technical background but no knowledge in this particular field", not "I'm a literal five year old". I think it's probably close to impossible to explain this adequately to an actual five year old while staying true to the essence of the paper.

I should have added "for a technical paper". They are generally not written for five year olds, and "metric space" is a term I've seen introduced anywhere between semesters 1 and 3 in most technical Bachelor's degrees.

I feel like the paper itself does a fairly good job:

> The [k-server] problem’s definition is simple: There are k servers located at points of a metric space. At each time step, a request arrives at a point of the metric space. An online algorithm must serve the request immediately by moving a server to the requested location, without knowledge of future requests. The goal is to minimize the total distance traveled by servers.

> The k-server conjecture states that a deterministic online algorithm can achieve competitive ratio k on every metric space.

I only had to look up what "competitive" means in this context, and wikipedia [0] had this to say about it:

> An algorithm is competitive if its competitive ratio—the ratio between its performance and the offline algorithm's performance—is bounded.

The ratio by which this performance is bounded for a k-competitive algorithm is k (plus some constant) [1]. We can consider the analogy of k support technicians ("servers) located in different (physical) locations ("in metric space"): The conjecture/theorem states that in any metric space (Not necessarily two- or three-dimensional), there exists an online algorithm that results in travelled distances of no more than roughly k times that of the optimal distance if all requests were known in advance.

[0] https://en.wikipedia.org/wiki/Competitive_analysis_(online_a...

[1] https://www14.in.tum.de/personen/albers/papers/brics.pdf Section 1.1

its not quite the correct method though, the commenter suggests using the first row as a page index, and the second row as the word index, but the actual solution was using both rows as word indices within the 32 paragraphs ("Proquiritations") paired to the 32 numbers in each row

I found this german blog: https://scienceblogs.de/klausis-krypto-kolumne/2014/11/17/we...

which links to: https://archive.org/details/s9notesqueries03londuoft/page/12...

which is in reference to the original proquiritations here: https://archive.org/details/worksofsirthomas00mait/page/416/...

i had also never heard of this before today and wonder if people had even seriously tried to decipher this at all?

Because in academia, your name and your body of work is a big deal that can open or close doors, and assigning credit is an important part of it. This isn't a new thing/exclusive to the AI era either.

I haven't played it myself, but surely an article written one month after the game's release doesn't do it justice, considering there have been multiple major updates attempting to address the gripes players initially had with the game in the 10 years since then (And review trends on steam indicate that these attempts were successful)

False positives (Pangram labels something slop that isn't) I'm not that worried about, since the worthwile articles generally circle back to me some way through the communities I'm involved in, so I like to err on the side of not instantly reading something anyways.

For false negatives, that depends on the type of article at hand. If an error leads to me spending 2 minutes reading an article that ended up being not worth reading I'd be fine with an order of magnitude more than if false positives happen on articles that take 30 minutes to go through. Maybe on the order of 1-3% overall?

Unfortunately, given the sheer magnitude of text on the internet, we need heuristics to gauge whether something is worth reading or not. For authors/blogs that I don't yet know, LLM-generated text is a clear anti-signal for me, and Pangram has been very useful in that regard.

> "I have a bucket of mud, in which I stick my hands for 15 minutes every morning.

Not quite but close enough :D https://www.youtube.com/watch?v=DWzWwx8T3nE

In a rug pull, the thing you own (usually some kind of digital asset) goes down in value, leaving you with less money than you started with, whereas theft leaves you no longer possessing the asset itself.

how is it a rug pull? if it's an inside job, doesn't that just make it theft?

"censor" feels like the wrong word here

> I'm just going off what regular typing keyboards go for.

Are there any regular typing keyboards with velocity sensitive keys? I genuinely don't know, it's just the first thing I'd compare this to

> Why AA’s plot is misleading

> The first issue I have with it is that it uses a logarithmic scale on the cost axis. Using a log scale is the only way to make you spot the difference between a model that costs $0.015 per task and one that costs $0.032, while the same plot contains a model that costs $3.69 — almost 250 times as expensive. However, the net result is that the viewers can no longer appreciate the immensity of the price difference between the cheap models and the heavy ones; nor can they realize how inconsequential the price differences are between the cheap models.

This is an asinine complaint, and nobody can seriously tell me that the last plot on their page [0] is more readable than the AA one [1]. If I'm using a model at the lower range of the cost scale for whatever list of tasks, and i switch to another model at the lower end of the cost scale, my spending might double anyways! This should be reflected in the plot, and linear scale doesn't do it justice.

It's also much easier to see the mentioned pareto frontier in the log plot than in the linear one.

I can see why they disagree with the pricing determination for open/local models, but I don't think there is one clear right way to do it. So how do they do it instead?

>Hardware is priced at zero, on the basis that both an RTX 3090 PC and a 64GB Strix Halo are desirable gaming/work machines anyways.

...oh

Would have been nice to mention explicitly how the pareto frontier changes with those new calculations.

[0] https://openteams.com/wp-content/uploads/2026/09/all_models-... [1] https://artificialanalysis.ai/#intelligence-comparison-tabs

> This course teaches the craft of building practical quantum circuits from scratch.

Can't wait to have an ion trap in my living room!

I would love to appreciate the work that is going into this curriculum, but when the lesson overview is overwhelming amounts of AI copy slop I just can't bring myself to be interested.

Edit: i'm talking about the text, not the pictures. It's not overwhelming in the sense that it is a lot of text, but in the sense that the amount of Claude-isms in relation to both word count and actual content is way too high for me to care about. Take lesson 9 for example:

9. Where the quantum lives

    every page is a classical drawing

    the animation is impossible

    each step has a classical model; they refuse to glue

    one story per step, no story overall

    where the quantum actually lives

    and the gap has a price, in coherent memory

I have no idea what is going on here or what the topic of this lesson is, and the same is true for like half of the lesson overviews here. Like i said, i'd love to be able to judge the content, but there is none here.

> That’s self-aggrandizing white male behavior!

I think you might be victimizing yourself a bit much there

Hmm, maybe I'm missing something, but what's the difference between "top of canopy" and "No trees" then?

I've seen the setting, and I've seen that you suggest toggling it when tree shadows seem inaccurate, can you elaborate on what that setting does exactly?

looks like both words are valid, TIL!

I think it's still valid to see this as an example of the Jevons paradox: Your resource is money, and reduced per-unit cost means you are using your money (resource) more efficiently. If the total spend now goes up instead of stagnating or decreasing, you've got Jevons

> The set of points is defined to be not collinear, thus we cannot draw a line passing through more than two of them

This is wrong and you are misunderstanding what collinearity means. You could have a set of points where all but one are on the same line, and the set of a whole will be not collinear, while we can obviously draw a line that passes through more than two of them.

A different way of stating the theorem is that any finite set of points has either a line passing through all points (i.e. the set is collinear) or there exists a line that passes through exactly two points. This dichotomy (why two and not three? Why can't we construct a set where any line passes through at least three points?) is not immediately obvious.

> It could pass through three if those points were collinear but the statement says they're not

The theorem doesn't presume that no three points are collinear, it presumes that the set as a whole isn't collinear, which is a much weaker statement.

> hypothesis

I think you meant premise?