cheald

Abliteration and fine-tuning remain standard operations on open weight models.

A robust open source profile is my single favorite hiring profile indicator. However, with the current state of things, if I get a whiff of AI-driven "contribution" it becomes an instant black mark against the candidate.

So far I've been impressed enough with the HW4 Teslas that I haven't had them do anything that I had to intervene to correct or prevent. It's pretty amazing at how well it handles all kinds of things - construction, weird merges, road debris. This morning, there was a tire in the middle of the road, which caused traffic ahead of me to slam to a halt. Mine had to brake hard enough that ABS engaged, and then navigated around the tire. I was impressed.

Very fun! There seems to be an exception when using the 10-roll function that might need some attention.

    index-BwvgcC1c.js:62 Uncaught TypeError:
      C is not a function at zi (index-BwvgcC1c.js:62:150798)

Those lines of code are:

    for (var oe = [], ce = 0; ce < 10; ce++)
      oe.push(C());

"Applied theology with inadequate safety margins" might just be my favorite difficulty description ever.

If you wouldn't give it to an enthusiastic junior dev, don't give it to AI, period.

Hang out at a BJJ or MMA gym for a bit, and you'll find plenty. Peptides are really popular in combat sports circles, with good reason.

I used a combo of low-dose retatrutide, tesamorelin, and ipamorelin and lost about 15lb over 45 days, including 60% of my visceral fat, and put on 4lb of muscle, per before-and-after DEXA scans. I lifted regularly, ate well, and prioritized protein, and while I definitely under-ate protein, I was very pleased to find that I was able to increase muscle mass while cutting the fat. My visceral fat was the primary target here, since I'd been unable to get it to budge despite consistent training and diet. Very pleased.

It's actually very easy to cool in deserts, because low humidity makes it very easy to move heat into the ambient air. You have to contend against ambient temperatures, but that's what insulation is for. The other big things you need for datacenters are reliable power and a low probability of infrastructure-disrupting natural disasters.

I've done a lot of exploratory work with Stable Diffusion LoRAs, and I actually do buy that there's some juice here, though it's almost certainly not nearly as good as other techniques can be. In particular, this technique will likely avoid the intruder dimension problem which plagues naive LoRA. SVD is expensive, but you only have to do it once at the beginning of training.

I haven't done much research lately, but when I was working on it, I was having substantial success training an adapter of the form U_k @ P @ A, where U_k was the top k left singular vectors of the underlying weight, and then P and A were your typical LoRA projection matrices.

The 13 parameters are kind of misleading here; the real juice is going to be in the P_i fixed random matrices. My suspicion is that they are overfitting to the benchmark, but they almost certainly are observing a real gain in model capacity that is largely due to avoiding the intruder dimension problem.

Have you not seen Instagram lately? It's overwhelmingly clogged by AI content, much of it masquerading as real people.

It's best to think of instruct-tuned LLMs as mirrors rather than intelligences. They generally reflect what you're putting into them, but they do it in a way that can easily masquerade as wisdom. I think this makes it really easy for people to self-delude.

Remember that many people are heavily are happy-path biased. They see a good result once and say "that's it, ship it!"

I'm sure they QA'd it, but QA was probably "does this give me good results" (almost certainly 'yes' with an LLM), not "does this consistently not give me bad results".