jerkstate

3.7 flash was by far the best model for image recognition tasks according to my benchmarks. 3.8 flash didn't regress any candidates and improved some specificity (positive ID of common name vs species name of exotic fruit, correct identification of cast/replica of artifact and statue) but is still relatively weaker (26/30) on esoteric public figures (Korean beatboxers). I'm going to have to make my benchmark harder.

My benchmark is tiny compared to WorldVQA or FG-BMK, which are available, so I'd point you in that direction if you're interested in a VLM benchmark. My use-case isn't exactly captioning as in "what is in this image?" -> caption, I am using the VLM to validate captions, as in "is this an image of [supposed subject]?" - my ranking of models I've benchmarked is gemini-3.7-flash > seed-2.1-turbo > gpt-5.6-luna > qwen-3.7-plus > qwen-3.7-flash. Gemini is almost perfect on my test dataset, only failing on some esoteric pop-culture minor celebrities and being over-specific in some cases (i.e. Q: is this [common name of fruit]? A: that's a [latin species name of fruit], not a [common name of fruit]; false). However, gemini-3.7-flash is only in my test list because openrouter has it on 75% introductory discount; otherwise it would be about 4x more expensive than seed.

> The deepseek-v4-flash-vision-exp model accepts images alongside text, so you can ask the model to describe pictures

it doesn't specify what type of images it can and can't describe, I'm pointing out what type it isn't good at compared to other models.

I just ran my image recognition benchmark on it ("is this XXX public landmark"?) and it misses a lot that bytedance seed 2.1 turbo gets right; for example: Asked "Is this Salisbury Cathedral" and supplied a picture of Wells Cathedral, it answers "Yes, the west facade of Salisbury Cathedral". Bytedance seed 2.1 turbo correctly says no. Similar results for a picture of Manhattan Bridge sent as Brooklyn Bridge, Chartres Cathedral sent as Notre Dame, etc. I have a benchmark of 12 such images and seed gets 11/12 and deepseek only gets 6/12.

the only thing that still kind of annoys me is constantly being told what something is not, but even that statement is load-bearing (see what I did there) because it records how it ended up with this decision, because it's not that other choice that it mentions.

FWIW, I also think the constant chorus about how new models are worse than old models is a human hallucination. They're certainly not perfect but every one becomes more steerable in terms of actually completing more and more complex work.

I don't know that much about math but I've read that one possibility is he had found a valid proof for n=4 and assumed that it generalized. Hopefully someone else who knows more about the subject chimes in!

There’s no particular reason it’s not OSS, but my main interest is collecting a lot of data on different athletes and publishing original research. Most weightlifting studies are small n and over a short amount of time. My particular interest is how volume, load, and fatigue are related to strength, endurance, and compliance over time. My intention is to run it for a while, look at the data to generate some hypotheses, pre-register them, then run some experiments (and by that I mean just keep collecting data). If someone else was particularly interested in this goal, I would definitely invite them to the project. That’s why it was important for me to design it to be hosted for just the cost of the domain name, because I don’t really intend to make money from it, I’m just interested in the data.

If this is something you are interested in, I did make a mobile friendly SPA similar to this: https://curvefit.app

> high reps are better for hypertrophy work

Some nuance here: the latest research shows that proximity to failure is the main hypertrophy driver regardless of load and rep count; high rep count makes proximity to failure harder to gauge; so high load/low reps close to failure is probably better for hypertrophy (there are other good reasons to do higher reps/lower load work though)

I wrote this app as a SPA! It uses a curve formulation similar to Brzycki, except I added a “shape” parameter (an exponent gamma between 0 and 1) that slopes the 1rm downwards at the right side.

My main finding for “pick whatever weight you want today” was that picking a lot of different weights made the curve less identifiable, so my latest iteration encourages you to pick a ladder for a few sentinel exercises per mesocycle in order to improve the statistical power. In addition, strength improves more quickly at >80% of 1RM, and hypertrophy depends on proximity to failure, so if you pick a lower weight, you really need to go to failure, which burns you out for the rest of your session, where leaving 1-2 reps in reserve is probably sufficient for hypertrophy and leaves a lot more gas in the tank for the rest of the session. Definitely open to suggestion/discussion here.

https://curvefit.app (it runs on Cloudflare free tier, so I won’t have to start running ads or charging until I hit a couple thousand users)

not everyone who can think and solve interesting problems can write prose well, but there's still a lot of complaining about AI-generated text.

Don't pay too much attention to the naysayers if they don't have constructive advice to offer (but do try to collect what signal you can from the noise). There are a lot of people who viscerally hate AI for a ton of different reasons, or maybe are perfectionists about the subject matter. Even if these AI technologies are currently imperfect, they solve Bloom's Two Sigma problem - students tutored to mastery one-on-one perform two standard deviations better than students taught in traditional classroom settings. Imagine a world where everyone is 30 IQ points smarter just from better education technology!

Personally, I just used chatgpt the other day to _finally_ understand why a dot product of vectors is equivalent to cosine similarity. Mathacademy's lesson is very good but the explanation wasn't enlightening for me, it took chatgpt breaking it down into tiny bite-sized chunks, patiently re-explaining what a cosine actually is and where it disappears to in the dot product, making little animations of vectors with sliders that I could move myself to try stuff with differently shaped and rotated triangles, for me to _actually and viscerally understand it_ and after spending probably 2 hours on this simple concept, now I _actually_ feel like I understand it rather than having just memorized it. So I feel much more confident that I can build deeper understanding on top of that.

it interleaves the end note in the text at the point where the endnote is referenced (by number), reads the endnote, and then sounds a bell to indicate that we've returned to the main text.

Sometimes endnotes reference other endnotes and you have to wait until the point in the text, e.g. endnote 304 is referenced from earlier endnotes than it occurs in the text (rather late in the book) and you just have to wait until endnote 304 is read to find out what it is. that's a bit of a disadvantage, but infinite jest is the kind of book that you really need to read multiple times in order to get it anyways.

if you're interested in completing it, check out the 30th anniversary audiobook, it just came out a few months ago, very accessible

I visited this and several other Gaudi buildings about 15 years ago in Barcelona, and many of them are truly breathtaking, or at least dramatically original and unique. I went to the Gaudi museum as well and found it fascinating that the architect himself was not a professional mathematician - he did not use hyperbolic cosine to calculate the dimensions of the catenary curves, he traced the outline of hanging chains. Really interesting to hear about how he also heavily used ratios and symmetry. I love how artistic taste can be partially derived from math (but the math itself isn't sufficient to develop artistic taste)

It’s definitely got AI tells, and I understand the objection - it’s hard to tell the difference between slop (written without review or direction, meant to fill space or present a simulacra of information) and actual informational content. With this article, it was clear that the author wanted to describe what problems they had and how they overcame those problems, and it was concise and I could tell it was reviewed so the content seemed trustworthy to me. I understand that some people have pretty high standards - if you don’t have the time to type out all of the code yourself one character at a time, then write stylish, interesting prose about it one letter at a time, if all you want to do is solve interesting problems and communicate how you solved them, you should shut up until you get good. Me, I am also imperfect and lazy, so I don’t care to demand style points or demonstrations of perspicacity from others.

Thanks for posting this, OP, I thought it was interesting.

> In some ways, the concentration of power in a dictatorship might be better, if the dictator was well morally aligned with the people.

This is pretty much the exact argument that Hayek makes - socialism leads to fascism through political gridlock.

It’s time for people who understand how to use AI to stop trying to convince people who don’t. Let them churn their own butter, weave their own clothes, and type out their programs one character at a time. It’s what they want.

How do you know those aren’t the same thing?

Their cohort studies are across three completely different populations. I would be willing to listen to an argument that for some specific population there’s a firearm ownership level above which there are diminishing safety returns, but I find the idea that this same level would hold across wildly different populations absurd

Try the audiobook, although I’m sure purists would consider it cheating. You can zone out and it keeps going and something will pull your attention and you can rewind to get the context.

How do you know what volitions and abilities it doesn’t have? My agent teams continually improve the process and tooling they use for teamwork

Get a sleep study done to find out how serious it is

Do you consider it humanitarian to further weaken a nation where the average gdp per capita is less than $3000 and who are in no way a threat to us? In what way do the immigrants arriving from there improve America? Can you give a coherent argument? I can easily argue the opposite from statistics:

From Census Bureau’s American Community Survey (ACS) and Annual Social and Economic Supplement of the Current Population Survey (ASEC CPS):

Of Haitian immigrants ages 25 to 64, 17.1 percent have not graduated from high school, 30.3 percent have only a high school degree, 30 percent have some college, and 22.6 percent have at least a college degree. This compares to 6.6 percent, 25.4 percent, 30.4 percent, and 37.6 percent for U.S.-born Americans. (2022 ACS)

Of households headed by Haitian immigrants, 52.7 percent use at least one major welfare program. For households headed by native-born Americans it is 28.4 percent.5 (2023 ASEC CPS)

So it’s clear that the previous immigration regime that allowed this was optimizing for something other than improving America or intentionally weakening rival countries.

Are you implying that we should be trying to weaken Haiti?

What about the immigrant groups that don’t dramatically out-earn American-born citizens?

So what prevents the incompetent and lazy from immigrating?

so in order to increase our median intelligence, we should make the process super easy?

It’s a shame they were so anti-nuclear. Best song on that album was Ohm Sweet Ohm.