janalsncm

3 ms·

31.
by janalsncm 20d ago
The problem is this entire argument depends on “chain of thought” being a faithful representation of a model’s actual plan. That doesn’t need to be the case. Thinking tokens are just tokens at the end of the day. Do you know what is a faith
32.
by janalsncm 22d ago
Basically your argument relies on two claims, both of which must be true. 1. Competitors to OpenAI and Anthropic are good because of distillation. 2. OpenAI and Anthropic will come up with some methods for preventing distillation in the fut
33.
by janalsncm 22d ago
Maybe another way to say it is to reframe the idea of “human in the loop”. Humans are always in the loop, because we can always expand the definition of loop to include the humans that pushed the button and built the system and processes th
34.
by janalsncm 22d ago
Yes as a teenager I made many mistakes, not least of them being making a Gmail account. I even used it to sign up for a few services. I do have a bit of a wild side so sometimes I log in to check it.
35.
by janalsncm 22d ago
Log scale should only affect how the years are displayed on the chart, not how they are sampled.
36.
by janalsncm 23d ago
If the implementation was correct it wouldn’t be hard to let the user choose either option. Yes, many births would be within the last 200 years but it’s still very different to be a man born in the year 1950 in New York vs a woman in 1950 i
37.
by janalsncm 23d ago
That would certainly be easier at this point, but I like the “random human life” idea more. I’ve reflected in the past about how lucky I am to have been born where I was, when I was, and be given the abilities that I have. Sure there are pl
38.
by janalsncm 23d ago
I think this would be cool if it actually worked as intended. This simulation doesn’t even draw birth years correctly. The right way would be to bucket people by something like 1 million. Everyone in the bucket has the same sex, birth year
39.
by janalsncm 23d ago
That’s not what it says on the tin though. “Drawn at random from everyone who has ever lived” means years will not be uniformly distributed.
40.
by janalsncm 23d ago
If you are telling the truth you might want to check your network for any weird connections to Chinese LLM transit stations.
41.
by janalsncm 23d ago
Ok so the correct comparison would be to fix the harness on the old model and re-compare. Now they are comparing a new model to an old crippled one.
42.
by janalsncm 23d ago
It’s more like if Zitron was a doctor his prognosis would be “you’re going to die on Tuesday” and you come back on Wednesday (after not dying on Tuesday) and he defends it by saying “everybody dies”.
43.
by janalsncm 23d ago
Maybe just a big Mitch Hedberg fan?
44.
by janalsncm 23d ago
I will give Anthropic credit for standing up against the department of war. The bar is incredibly low, but not doing domestic surveillance and not creating autonomous weapons are laudable. That doesn’t mean I like them pirating books and be
45.
by janalsncm 24d ago
Oh nice, I came here and made a very similar comment. I like the Easter eggs that reward readers who are paying attention.
46.
by janalsncm 24d ago
> February 2025: "I will keep writing this stuff until I’m proven wrong." Wrong (Zitron continues to write despite repeatedly being proven wrong) This was a good one.
47.
by janalsncm 26d ago
That’s fair, but my impression is that AI is a very industry-driven field. When commercialization fails, industry funding dries up and drags gov funding down too, resulting in an AI winter. And industry really cares about solving practical
48.
by janalsncm 26d ago
It’s both. Fine-tuning a BERT model changes its weights, which causes the embeddings to change. For example you might have one model which embeds a text query and another model which embeds an image. You also have a dataset of image + text
49.
by janalsncm 26d ago
In that case, you agree with me that you don’t have a right to privacy for all of the reasons I already stated plus the additional one that you can’t afford to fight it in court.
50.
by janalsncm 26d ago
Switching to end to end encrypted chat is a technical solution to what is ultimately a social/political problem. It doesn’t work: https://www.fire.org/cases/florida-international-university-...
51.
by janalsncm 26d ago
GPT2 and 3 work via autoregression. In chat bots, decoders work via autoregression. > people spend years just post-training BERTs in various ways Yes, I was one of them. That’s not called “post-training” it’s called fine-tuning.
52.
by janalsncm 26d ago
1) Except for specific things like third and fourth amendment rights, you don’t have a blanket “privacy” right, much less in public. 2) If your argument is that flock cameras are illegal, take them to court.
53.
by janalsncm 26d ago
My impression at the time was they were perhaps overly cautious but this was a bunch of researchers who wanted to self-regulate. Anthropic didn’t exist, deepmind was also much less product-focused and relatively cautious. Chinese models wer
54.
by janalsncm 26d ago
> [in 2020/2021] the dominance of autoregression was not as well-established as it is today: GPT-3 had turned some heads, but the ‘ChatGPT moment’ wouldn’t come until late 2022 I disagree with this. Decoders were absolutely dominant
55.
by janalsncm 29d ago
I’d imagine the choice would be Windows (+$120) Linux ($0) DIY ($0)
56.
Meet.hn(meet.hn)
2 points by janalsncm 29d ago|0 comments
57.
by janalsncm 1mo ago
On narrow domains, it is very common for small models to match or outperform larger ones at a fraction of the parameter count. For example in language, this is called the “curse of multilinguality”. Small models that handle a single transla
58.
by janalsncm 1mo ago
“Properly engineered” means the principle of least privilege and fitting the sandbox to the constraints of the problem. The test did not require internet. They gave it internet. Therefore it was not properly engineered. We do not need to de
59.
by janalsncm 1mo ago
If I test out my backyard cannon and blast a 10 foot hole in my neighbor's wall, “a cannon that can't smash through walls isn't useful” probably won't be a great defense in court.
60.
by janalsncm 1mo ago
It doesn’t matter how smart it is. 200 years of technology were not accomplished by thinking harder. It required empirical observation, new materials and tools, and supply chains. We could send a cracked team of scientists and engineers tha

More ›