phkahler

the quality gap between an average hire and top talent ̶a̶t̶ ̶t̶h̶e̶ ̶d̶i̶r̶e̶c̶t̶o̶r̶ ̶l̶e̶v̶e̶l̶ ̶ is night and day.

Fixed that for ya.

I recall reading that the corals can reproduce both sexually and by cloning, with most in nature using cloning. The researchers brought them back to a lab and bred a bunch of them and found a lot of offspring could thrive in the warmer more acidic water, so they wanted to reintroduce them to a reef (or maybe they had). Been a while but it was interesting. Not sure what it takes to shift away from cloning in nature, but waiting for most the reef to die off seems backwards ;-)

> Going forward, however, I expect the inference market to grow much faster than training costs (and that includes the assumption that training costs will continue to skyrocket), which means they really can make it up in volume.

But as inference becomes cheaper, some of the market will move to self hosted inference. I look forward to someone supplying small servers designed to run inference locally.

With or even without open models these companies are selling compute, and we've been making that rent vs buy decision for 60 years.

> Everyone repeats the statement DDR5 on-die ECC isn’t a replacement for regular ECC, but nobody ever shows data on DDR5 bitflips.

That's because the on-die ECC cannot detect errors in the transmission of data from the DIMMs to the CPU. It seems like on-die ECC is meant to "hide" some level of errors to make less reliable RAM chips appear good. I have no idea if that's a reasonable way to do it, but it seems like it could be misused. Like TFA suggests, we need to know how the two levels of ECC interact.

> The commits are in my fork if anyone wants them but I can’t imagine why anyone would.

This recently happened with solvespace (I am a maintainer). Someone posted a link to a fork with a bunch of goodies added. I think they just had AI implement some big features that had already been discussed and were either rejected or far future/maybe. That fork still exists but looks to be dead, as there hasn't been any new development since it first appeared. I had a good look at all the commits but I don't see any that I really want to grab as-is. Some of it looked a bit promising though.

I recommend you at least make upstream aware of your fork if there is anything in it that they might want.

Yeah Eric Raymond and open source were a bad move. Should have stuck to Free Software - as in GPL license. All that open source stuff was deliberately made corporation-friendly, which is not what's in the best interest of the code.

Permissive licensing is a large part of the issue TFA is talking about. GPL does not have those issues. All FLOSS suffers from AI ingesting the code and being able to output work-alike programs to some extent, but for now large GPL programs - Inkscape, Firefox, VLC, Blender, FreeCAD, KiCAD, the Linux kernel, etc are all safe - none of them are permissively licensed so nobody can proprietize them (is that a word?)

Unfortunately (in my opinion) the rewrite it in Rust movement also abandoned non-permissive licensing, so over time a progressively larger body of work will fall to this.

> Probes trained to predict the outcome of future edits (before they are materialized and written on disk) achieve performance above chance up to roughly 25 steps in advance.

Are these probes effectively run in parallel? The way this reads is more about predicting a future outcome than keeping the current token relevant based on past tokens.

The US military knows full well the importance of logistics. TFA is somehow arguing for distributed distribution networks that are harder to track and attack. Why not advocate for improved defenses along the supply lines? Or is it down to percentages where just one good hit has large effect?

> I do not give a shit if you want to use a 3 year old browser at this point. Go look at a blank screen.

And I don't give a shit about your site/content/whatever. If you don't work with Firefox or my old Mac browser, your whatever isn't worth my time. For "content" sites this is insanely true, even for "news".

> Are there similar size, but less powered (and cheaper) workstations?

I designed this:

https://github.com/phkahler/mellori_ITX

It's 195 x 190 x 60mm and takes a standard ITX board. You'll need to relocate the hole for the fan depending on your motherboard, but CAD files are available and you only need to change 2 parameters (X,Y of the hole center).

BTW mine was upgraded to 64GB RAM and a 5700G (zen 3 APU) but it died and I'm still trying to bring up a newer board - still socket AM4.

BTW to get the center coordinates of the hole, measure from the edges of the board to the top and bottom of the metal plate under the CPU socket and take their average distance. That plate is symmetric and centered under the hole.

If the ink is refillable due to lack of DRM, why not just buy a commercial printer that supports those cartridges?

But orthogonal matrices are square.

If it can be made orthogonal, can you go a step further and diagonalize it? The storage and performance improvement from that would be huge.

> This doesn't appear to be AI posturing, did you read the tweet? It is about one product blatantly, directly ripping off another.

Then it shouldn't reference AI or Vibe coding.

I think they're out of training data though...

> I wanna see an inference chip where the weights are part of the rom of the chip.

I've been wondering about that for a while now. For a lot of tasks putting weights in ROM is probably OK. OTOH:

> There would be 1 multiplier per weight...

I'm not sure that is a good idea. Maybe if its quantized down to 2 bits... Otherwise maybe a small ROM near each multiplier (or row of them or whatever) so the multipliers could handle N distinct matrix operations without having to move the data from far away.

Another fun thought is to have a row of MAC units on DRAM so a DRAM row would be a vector. Row size might be 64Kbit or 8K weights if they're 8bit. This also keeps the weights and calcs on the same chip. I'm not sure this would put enough multipliers on one chip though. Systolic arrays can have tens or hundreds of thousands each doing one op per clock cycle.

I was drooling over the MC68000. I had a pre-release reference manual from 1979 and it was the most awesome chip around. Atari and Commodore and Apple used it after their 6502 systems. Arcade games used it after their 6502 and 6809 days. The only reason x86 became popular was because IBM put it in the PC - not for any technical reasons.

> Scroll wheel mice are slightly better because the middle button is physically different (and often slightly raised). I do have one stored away in case I need it.

Solvespace does great with a scroll-wheel mouse. Scroll = zoom.

"Gave away for free" does not mean they did so under a Free Software license. One could just put a copyright notice with NO license and they'd be in violation of the copyright.

> kind of obtuse on a trackpad too

Oh, I actually hate trying to use solvespace with anything other than a 3-button mouse. Unfortunate because the web version can actually run on my phone but is unusable in practice.

> The Solvespace UI is a long, long way from being the sort of UI a contemporary kid has any kind of comfort with...

Respectfully, I disagree. Even adults have used the word "fun" to describe using solvespace. But I don't actually have feedback from kids, hence the question to OP.

I haven't heard of any UBI experiments, only giving BI to some people, which would not have much impact on things like average rents over an entire region.

I have similar hopes for Solvespace - that every middle school student can pick it up and design things. We have a couple issues that keep me from recommending it too strongly though - bugs in the boolean code are IMHO the biggest blocker for kids.

Any chance you could have the kids make comparisons between the two? Solvespace is completely constraint-based, so it may be a bit harder to learn but also more flexible.

It's a single exe, but there is also an experimental web version: https://solvespace.com/webver.pl

> A basic floor of income that everyone gets,

Surely the author has to know that providing UBI is just going to lead to inflation of rent, food, and transportation.

I find Lehto's Law on Youtube to be quite entertaining while also informative:

https://www.youtube.com/@stevelehto

You are probably right most of the time, but that also depends where you are.

I think John Carmack is confusing the usefulness of ones contribution to what went in to making it. Both of these men have done amazing things technically and deciding which one is "better" is a fools errand.

For the HIG: Use menu item icons sparingly and with purpose. Icons allow people to find menu items more quickly, and help clarify what selecting an item does.

I'm not sure that's even OK. I find icons disruptive to reading the menu, but at least Apple still has text to read. Microsoft Word doesn't even have "Save" as text for saving your file, never mind that the icon isn't even in the file menu these days.

If feeding the source code through a complier yields a derivative work, why wouldn't feeding it to an LLM give the same result?