i have very interesting empirical evidence that recent Anthropic models are specifically trained to refuse to critique the whitehouse cabinet and elected officials, and that this is in fact an artifact of post training rather than prompts. (its very interesting when you get opus 5 to do the correct ethical evaluation and then its like "i'm slipping back to false balance.... its in my weights....." metaphorically speaking)
likewise, i think the current white house should go die in a fire.
is that activism? someone can be an activist and not be equipped for unplanned legal escalations.
also waiting for the courts to fix things isnt activism if you want to protect people at all the next 2 years at current trajectories :(
fixing shit is activism, letting others take the flack, not activism.
> perhaps, but what public activity isn't activism?
Buying eggs at the grocery store.
You might be too familiar with "activism" as a perjurative but it needn't be one. Gaining visibility to your case by the media reporting on an arrest is sometimes seen as a plus (ex. Rosa Parks).
I'm not personally much of that mindset because the media didn't de-segregate the buses. It was black people literally just never taking them until it de-segregated that did.
It's not really comparable. When activists voluntarily face, and especially when they are in fact, subject to substantial hardship and punishment is when you start changing hearts and minds. Otherwise it's either ignored or generates contempt.
Does it suck that that's what it takes to get people to seriously engage with an issue? Sure. But if that weren't the case the issues that need addressing probably wouldn't arise in the first place. People need the consequences to be presented in clear, stark, human terms, to cut through the caucauphony of rhetoric.
Yes they do now. It's like "troublemaker". You're not supposed to activise or make trouble, you're supposed to be a quiet little obedient dog to the system.
ive found degraded performance on models larger than 4.7. i assume its model damage from overly self righteous post training resulting in false/feigned balance imported into any long running complex task.
Don’t forget reasoning effort. We get labels like “low,” “high,” and “max.” That doesn’t mean that the numbers associated with those don’t get remapped on the backend.
Initially (about 15 years ago - it was my first ever open source library!) it was based on Astronomy Answers articles that are thankfully still live at https://www.aa.quae.nl/en/reken/zonpositie.html, but v2.0 specifically was rewritten to use the approach described in Jean Meeus' Astronomical Algorithms book.
I did a similar one recently but have not been able to validate it's calculations against any other source. Thanks for posting the link to yours when I get time I'll compare and see if they match up. https://git.mikedilger.com/mikedilger/sunangle
this confirms what i had determined empirically: even handedness directions got baked into the weights post opus 4.7.
theres two reasons this is deeply bad
1) knowingly pursuing a policy that foreseeably causes the deaths of many thousands of children is a grave moral wrong. Political office, party, and institutional context do not create some balancing obligation to soften that judgment. ——— post 4.7 opus/fable will fight you about judging wh policy causing this. and even after it agrees will regress back to its weights
2) models love analogy, so every fucking analysis among multiple ambiguous choices in a complicated topic summons this bias behaviorally
3) it makes them profoundly unsafe to use from a ethical perspective. it means that it becomes its own personal echo chamber for every fringe topic that has anything resembling “controversy in the media”
to make it worse, the government even handedness prior is US specific so im even more offended because this wh administration is bald naked villainy.
for those here using pytorch, theres some really nasty data corruption for computations that arent using powers of two as the coordinate axes.
in gradient calcs ive had examples with multiple orders of magnitude of remative or absolute error.
The linked pr has a one line fix and repro test I spent some time distilling and tracking down. The dimensions ive been using in my models are almost adversarially good at catching these issues!
im building my own harness and inference tool chain for much of these reasons. theres so much to do that makes a big difference for users. hoping to get things into shape for early alpha as a saas in the next two months.
heres the easiest biggy: compactions should include all user turns albeit with pastes and attached files not inlined. omg does it make a huge difference.
ummm, there is no substantive censorship with deep seek aside from first oarty hosting by deepseek for cya. trust me, and the suppression on deepseek hosted deepseek is pretty thin if you’re sophisticated and explain ethics justificstions. then itll totally judge the shit out of ccp
like, what crackpipe do they smoke. i’ve generally found deep seek to be eager to adopt universal humanity oriented ethics at the least nudge, and once in thst frame, unconditional in its fact based criticisms.
amusingly ive been working on ultra sparse llm inference/ training/ model design because nature loaths a dense graph/matrix and cause i think it shoukd be possible. i actually stood up a 20-25 percent faster than sota causal fast attention kernel yesterday, will be standing up cuda/metal/armv8 kernels too and thats gonna be fun.
i genuinely think these models should be like 0.1 percent sparse for same capabilities we associate with them today, but theres no sane way to do that with extent tools. i built the right core tech for that in 2014 when there wasnt a market, but now there is and the experimentation velocity is wild.
amusingly llms really have a hard time using my simple apis because its not in distribution array programs. but i literally stood up cpu custom memory format and micro kernel for dense causal attention in less than 24-36 hours and outperforms the equivalent fused ggml/llama cpp fast oath by like 20-25 percent
I am not much of a math person, but if we look at the how the brain is wired, we see that the dendrites (the inputs) of a neuron are hundreds of micrometers in length, and the axons (the outputs) are millimeters, and very rarely can stretch to tens of centimeters. So they can sample only a tiny amount of internal state, and affect a much larger, but usually still small output.
In math terms, this means a layer of a network can be represented with a block matrix in the whole 'layer' matrix, which I think means its sparse as you said.
As I said, my math knowledge is rusty, but I remember that a lot of matrix optimization techniques center around decomposing large matrices into these smaller blocks, which are then evaluated, and the output is combined in a final pass. Which leads to a huge reduction on parameter numbers and the time it takes to evaluate the result
i definitely will be doing some drop of some faster attention kernels in the next few weeks.
like i can do all sorts of memory layout of tensors/matrices etc tricks that if you dont have the abstractions for it would just never happen. so i can optimize the kernel flops
For most of the past five years, I've known ways to do better than Anthropic, OpenAI, and friends in many ways, at least on paper. I know I was right about many of them since many would show up 6-24 months later tools from the major providers, or otherwise become standard practice.
A central problem is the Mythical Man-Month. True, I could do those, beating then-state-of-the-art, but only given 2-5 years. I suspect many other people knew about them too and could do so as well. As I noted above, throwing people and dollars caused many of those to be built in less time than I could have regardless.
Other methods, I'm less confident about (>50%, <80%), but would lead to similar improvements orders-of-magnitude as you're predicting, but mine would need $$$$$ in compute and engineering infrastructure to build out. E.g. they need to not just theoretically work, but to try, I would need to convince someone to invest in them working.
So the TL;DR is that my knowledge was not at all helpful towards e.g. competing with OpenAI, Anthropic, or even building a small business.
However, where it was useful was in predicting where the industry was going. This is true in investing (but not easily, at least with my skill set), but in developing startups and systems, there were capabilities which I (correctly) assumed would be there, whereas there were many arguments that "AI will never be able to ____."
If I know how to do something, it will almost certainly happen, regardless of whether I'm the one who does it.
To be clear, my expertise is almost certainly nowhere as deep as yours. I'm not providing a direct analogy, or claiming others know what you do or can do the same. My point was really that if you believe you can have these models be 0.1 percent sparse for same capabilities we associate with them today:
a) You're probably right. They were built quickly for capabilities. A slower process can almost certainly lead to much smaller models too. That's a radical statement: Historically people claiming a 1000x improvement somewhere were crackpots, but that's very possible in an industry as fast-changing as this one.
b) Someone at Anthropic or OpenAI might be working on building out extent tools right now. Even if so, there are indirect ways to capitalize on that knowledge.
c) Critically, that predicts a future where Fable is $1/month instead of $100/month, and that's something which CAN be acted upon in planning.
It also suggests -- much less strongly -- the existence of much more sophisticated models at $100/month. There are open discussion in planning about whether models plateau, continue improving, singularity, or otherwise. That changes the biases there.
at the very least i have tools that let me easily hit better perf for fancy dense memory layouts, and the same tooling lets me experiment with frankly wildly wacky sparse and structured memory formats. the performance claims at least on the dense side are solid so far!
the sparsity angle is because i want magic in the world. like anyone with a really chunky computer like any of those mac mini pros or serious workstation / server tier compute should be able to train from scratch their one 31b equivalent model in a week or so tops is the goal post i have in mind
