cma

47 ms·

661.
by cma 1y ago
Google began as a search engine with a popup blocker extension for a competitor's browser. Now they're a display ad company with a browser that includes a built in popup blocker extension blocker.
662.
by cma 1y ago
I think he's just trying to choose a point in time when mental healthcare was more primitive to go along with saying the diagnosis is more sophisticated now. A main thing is that people with autism would just be classified as generally
663.
by cma 2y ago
He also tweeted in approval of this tweet putting forward the "Jewish people planned it" antisemitic form of great replacement theory with "you have said the actual truth": > Jewish communties have been pushing the e
664.
by cma 2y ago
There may be something like that that does make it counterintuitive. Usually those kind of Malcom Gladwell paradoxes end up overstated. There would be other factors, like drunk people are probably safer with their lights on too. Lane keep
665.
by cma 2y ago
You can give the rebate based on prior year or estimated usage at the start of the year, and then repay at the end of the year if it was too much, like with healthcare subsidies.
666.
by cma 2y ago
Those eventually degrade to CO2 so the increased warming from them compared to co2 by mass is temporary, like with methane.
667.
by cma 2y ago
I'd rather have my Bluetooth speaker stolen than an installed stereo stolen where they just gut parts of the car and rip things up. But it will be a bigger target since it's easier to resell.
668.
by cma 2y ago
Why wouldn't they use 2.5 flash first, and then if an identical query is made by lots of people rerun it with 2.5 pro? Sometimes it seems much more error prone than 2.5 pro or even 2.0 even on common searches.
669.
by cma 2y ago
If black and white or some other cases, but there are typically subpixels that can make things like sparse enough text stroke transitions get 3X horizontal resolution.
670.
by cma 2y ago
Hallucinations heavily are correlated with poor performance on factual benchmarks, or at least I was referring to hallucinating false facts it was never trained on. There's a subset of hallucinations though where it hallucinates real f
671.
by cma 2y ago
nice that's a huge difference
672.
by cma 2y ago
> Google has wanted to block third-party cookies for a long time and they can't because they're not allowed to, legally. There's no reason they couldn't allow add-ons to do it: but instead with manifest v3 I think it
673.
by cma 2y ago
Just since any continued training tends to cause catastrophic forgetting if the old info isn't regurgitated again. Not specifically showing catastrophic forgetting, but hallucination for o3: > From the results of this evaluat
674.
by cma 2y ago
There have already been good results there with DeepMind's math Olympiad work. I think the LLM portion there was only for translating from informal to formal in the training process and in the final process they still used a manual tr
675.
by cma 2y ago
At least with Deep Seek math (with the same RL technique as the later R1) they noted similar things in their paper in the "Why RL Works?" section. Around the 1:04:00 mark of this Yannic Kilcher video review of the Deepseek math p
676.
by cma 2y ago
I think this is definitely not true of catastrophic forgetting from finetuning. And with other related types of forgetting from model abliteration there are often extreme increases hallucination. The InstructGPT paper also showed that RLHF
677.
by cma 2y ago
I'm pretty sure RL causes catastrophic forgetting of its base knowledge and that's why o3 hallucinates so much more. If you mess around with trained weights you're going to delete some base knowledge, as least the knowledge t
678.
by cma 2y ago
With a laptop 7735HS, using WSL2, I get 15ms for the interpreter to start and exit without any imports.
679.
by cma 2y ago
What's the startup time now with 9950X3D, after a prior start so the pyc's are cached in RAM?
680.
by cma 2y ago
That's how it works, everything is recomputed again every additional prompt. But it can cache the state of things and restore for a lower fee, and reingesting what was formerly output is cheaper than making new output (serial bottlene
681.
by cma 2y ago
Why not put all post surgery patients through a cheap metal detector? Would at least catch stainless steel instruments.
682.
by cma 2y ago
Wouldn't reasoning training be expected to cause catastrophic forgetting of ground truth random fact stuff learned in the main training? Do they keep mixing in the original training data?
683.
by cma 2y ago
That's what my comment tried to say but flubbed it a little bit
684.
by cma 2y ago
Nice, I wouldn't want it to be automatic after pinch but that looks promising and I'm sure could be modded, maybe pinch zoom and then if you keep holding the two fingers without moving much for X amount of time it reflows
685.
by cma 2y ago
The original android browser had something much better, zoom until the size was right, double tap to reflow to fit
686.
by cma 2y ago
As long as you get the model weights without agreeing to the license, there has been no case in the US yet at least where model weights have been ruled to be subject to copyright and it is likely they are a mechanical transform of data. FB
687.
by cma 2y ago
Reading that article closely it says nothing about an NVD budget cut, only a NIST one. They were trackijg the changes after NIST's budget was cut, not NVD's. As pointed out below, CISA announced a cut and then NIST more than mad
688.
by cma 2y ago
That says nothing about a funding cut, see my comment below
689.
by cma 2y ago
Can search these for the links 2023 > CISA had previously been supporting the NIST NVD program with approximately $3.7 million per year in interagency funding, which they have discontinued 2024 > While NIST has since reallocated $8.5
690.
by cma 2y ago
I tried to look over the history and I only see a funding increase, CISA cut $3.7 million at the end of 2023 for the next year and in response NIST reallocated extra funding to NVD: $8.5 million in 2024 A funding shortfall and strain isn&#x

More ›