1) I couldn't see it, is there any indication of if the qwen math improvement caused degradations elsewhere?

2) the negative result on making similar improvements in Gemma 3 270M is also fascinating.

Such an unexpected result...

"No human can go edit a brain to become better at math" -- I'd offer this is exactly what "learning math" is.

I know I'm being pedantic, and I understand you mean "edit" like "surgically", but it does still feel worth noting.

The equivalent to "human learning math" would be: Astra fine tuning a Qwen model. It's distinctly different from surgically editing weights.

human brains are black boxes to humans. the fact you can learn math doesn't change that. you have no idea how your brain is learning math down to the neuron level.

LLM weights are _white boxes_ to LLMs.

Perhaps this is analogous to what our brains are doing? Maybe the different regions of the brain operate like semi-independent "models" that edit the other regions' "weights", and that's where the learning comes from?

Inducing the changes to become better at math requires years of study. This is "I know kung fu".

That's what I was thinking of, except it's not even dealing with the corpus of kung fu and doing the super-speed rounds of learning. It's directly reconfiguring the neurons. It's even better.

I thought that was what was happening to Neo when they jacked him into the learning pod?

"Learning" now feels like breaking down your whole Lego build and building it up again, in order to change a few fundamental blocks, while this feels like making only the required changes.

Orca was too heavy and, ultimately, too slow for me too keep using. Switched to herdr haven't looked back.

Calling people Nazis online ... and a state sponsored billion dollar surveillance program ... Are what you're choosing to equivocate?

"Calling people Nazis online" has resulted in countless murders, violent riots. Do not downplay it like that.

> countless murders

We can very easily count the murders that happen in the US every year. With the exception of COVID, which was largely due to unemployment and school closures, homicide has dropped precipitously since the invention of “calling people Nazis online”

Where are all these murders caused by calling out racism and fascism?

This was great to read even if just to name the stuff that I find so irritating!

I have thoughts about this pedagologically. I think these kinds of tools can be helpful, but the number of rhythm students I have seen who need to break out of the dependence on a visual metronome is rough.

Part of the struggle is really listening to yourself, and learning how to pick up on your own mistakes.

I hope the tool is helpful, and I hope I don't sound too curmudgeonly when I say don't let the time making and maintaining tools for practice detract from the most important thing, the ur-tool of all learning: more practice.

(Ask me how I know! A common trap for programmers!)

Makes sense! You can enable mini mode (top right) so that you can sing without the pitch track and program at the same time :D... then listen back to your take along with the original and see how you're tracking score wise.

Agree, a good teaching tool is one that understands the scaffolds are only temporary and is constantly seeking ways to reduce the presence of scaffolds as learners increase in skill. Most tools don't take on a pedagogical lens so essentially view the user's skill as static, and thus seek to maximally reduce friction because that is what is assumed to add value. This often comes at the expense of the learner.

A more generalized view is that the design of software should aim for some optimal level of friction between the person and the task (a two-way optimization), rather than simply try to diminish it (one-way).

Also, this is a neat product :)

I feel beyond being able to play precise rhythms and pitches, there is great value in being able to be off-grid in a meaningful, purposeful way.

While it is useful to be able to be precisely on grid if you want to be so — be it with the timing of your notes or the pitches played — many musicians I know are so overly fixated on the metrum of click track or the accurate pitch the forgot how to feel the ebs and flow that comes with great music that is just played in its own space and time. And pitch is no different here, sliding into or out of pitch of being able to sing/play slightly of grid is a very powerful skill.

But technique as the ability to pull it off is one thing, the knowledge when to use it and how much only comes with years of practise.

Nowadays many beginning musicians tie themselves so much to these grids, they can't imagine playing without them. To quote a musician that was asked by the sound engineer to play without click to make the track less mechanical: "That is allowed?"

I had an idea some time ago for a pitch training app that works aurally rather than visually. Basically it's a reverse auto tune. You wear headphones and hear your pitch errors exaggerated, forcing you to sing more accurately.

Wow, this is is incredibly important temporal context.

I hate how the role of the human scientists involved is so diminished, with credit being given to "Claude," which is at best a family of products.

Well Kvrocks is on disk, not in memory.

That's true, however the intention to run the benchmark would be against most of the widely used commands and see if there is any command or use case where kividb might be a better fit than kvrocks. strictly speaking from hardware performance, it would certainly beat kvrocks in terms of sheer throughput, but in terms of volume where does it stand would be something interesting to experiment.

Kvrocks is both iirc, since it can cache the rocksdb data in memory.

This was really loud in my mind reading this post. It seems so obvious!