christina97

5 ms·

31.
by christina97 3mo ago
This should be divided into three parts: marketing and selling people questionable combo drugs at insane cost (bad), the case of oral phenylephrine (idiotic + bad), and the efficacy of the other drugs in the mix (guaifanesin, etc) (unclear)
32.
by christina97 4mo ago
This post seems quite far fetched. Amazon is well aware of the paradox of choice, and the vast majority of UI changes I have seen recently are exactly those that guide and reinforce you to buy one option, without the decision paralysis. Ite
33.
by christina97 4mo ago
This may apply to B2C firms, but B2B firms are more likely to be run by people similar to them.
34.
by christina97 4mo ago
I thought this comment would go in a slightly different direction: the body of work that is mathematics has plenty of “bugs”; proofs with mistakes or other human errors. Yet we take the body to be correct (we believe it “works”) in aggregat
35.
by christina97 4mo ago
Start with a quant, you can run the Qwen 27B model at 4-bit on one 3090, presumably 6/8-bit on 2x3090.
36.
by christina97 4mo ago
I am not sure which part you are interpreting as underestimation or whatever? Quite the opposite: I claim the difference arises from a difference in strategies, not from intrinsic differences in ability. Also I was responding to a claim abo
37.
by christina97 4mo ago
I think you are making a distinction between pre training and later stages? The value on eg Fable output is exactly the careful preference optimization embedded in those responses. Not all data is the same (sorry if my first comment was slo
38.
by christina97 4mo ago
The article makes a very specific claim with a clear deadline less than 6 months ago. I do not underestimate the Chinese labs and their capabilities, if they wish they can retool to start overtaking the US labs with a different strategy. My
39.
by christina97 4mo ago
The Chinese models will not overtake the frontier US ones given the current way things are going. The US models derive their lead from incredible efforts to source more and higher quality (mostly synthetic data) via great feats (eg generati
40.
by christina97 4mo ago
Well operated, mature nuclear power plants can easily achieve 90%+ uptime. I don’t think this is a huge issue.
41.
by christina97 4mo ago
GP is saying that the goal should be something entirely different from gathering lots of data. Do you think the average person wants a higher resolution time series of their weight, or better access to a higher quality doctor, cheaper?
42.
by christina97 4mo ago
They do not claim to have migrated 50 million lines of Ruby. Simply that some migration took place in such a codebase.
43.
by christina97 4mo ago
The basic message is sound. Bregman himself seems a bit more confused than the (moral) clarity he attempts to project. This is fine.
44.
by christina97 4mo ago
As others pointed out, it only sucks for the buying side. The actual customers instead get price gouging and taking-the-blame as a service.
45.
by christina97 4mo ago
Right but that’s not what the article argues. The article argues that strawberries have been destroyed and now you only get the synthetic flavor and no grandma nostalgia.
46.
by christina97 4mo ago
The point is that that’s a valuation on paper, made up by a small group of investors that have a strong incentive to have the number be high. Would the number have gone up like this if it were a public company?
47.
by christina97 5mo ago
It seems worse in all aspects to the 26B A4B? I would have thought dense models beat MoE still on many benchmarks? Is the entire point of this model then that it runs if you don’t have enough GPU memory to load the 26B? That one runs faster
48.
by christina97 5mo ago
I think the key thing that depreciates is all their models. You train one at crazy cost and 6 months later it’s worth $0. If you ignore that depreciation you look much more profitable.
49.
by christina97 5mo ago
The SSD would wear out in days while the laptop generates two responses a day. This is like saying you could power your home with AA batteries, yes technically you could but in practice entirely infeasible.
50.
by christina97 5mo ago
My earlier research suggests NVIDIA does not actually cap spikes, it caps the average over short periods of time. So setting the power limit is no guarantee.
51.
by christina97 5mo ago
But it’s okay to be down when the whole internet is down.
52.
by christina97 5mo ago
While I’m sympathetic to this argument (it would be great if the internet were a safe place), in practice this thinking leads to governments trying to impose legislation that hurts legitimate uses but does little to protect from the long ta
53.
by christina97 6mo ago
I recently set up the 26B A4B model up on vLLM on an RTX3090 (4-bit) after a hiatus from local models. Just completely blown away by the speed and quality you can get now for sub-$1k investment. I tried first with Qwen but it was unstable a
54.
by christina97 6mo ago
You’re absolutely right!
55.
by christina97 6mo ago
By and large cooling, just like a data center.
56.
by christina97 6mo ago
If it’s so easy then why don’t we have a high quality classifier?
57.
by christina97 6mo ago
So in other words… they internalized a number of heavily subsidized externalities of individual transportation by car?
58.
by christina97 6mo ago
This is frankly one of the most frustrating things about LLMs: sometimes I just want to drive it into a corner. “Why the f** did you do X when I specifically told you not to?” It never leads to anything helpful. I don’t generally find it ne
59.
by christina97 6mo ago
Right, but the question is whether the companies producing foundation models will capture that value or not. Right now it seems like tokens might end up just being a commodity sold at cost plus, and companies higher up in the supply chain w
60.
by christina97 7mo ago
No the US government legitimately sells this kind of insurance (and reinsurance): https://www.dfc.gov/what-we-offer/our-products/political-ris...

More ›