michaf

Is there such a license? Or any license with special clauses for LLMs? Is it enforcable? Could someone 'poison' an LLM training run with injecting just one such licensed document? I am genuinely curious about what levers exist (or are conceivable) to protect your own IP from becoming LLM training data, if regular copyright does not qualify.

It even can model 'vi', including modal input, exiting with <Esc>:wq!, and it results in a file containing the entered text in the imagined filesystem.

If you restart the session and enter

  > echo test1 | md5

it will output another random 'hash' each time. But within one session it knows that md5 should produce the same output. Running

  > echo test1 | md5

multiple times, or even

  > VAR1="test1"; echo $VAR1 | md5

produces the same output for me within the same session, while

  > echo test2 | md5

results in a different, but also session-consistent hash.

I just told it that it had a directory containing the GPT-3 model weights. I'll assume the files contain some imagined gibberish, but what if not? ...

  > ls -lh 
  -rw-rw-r-- 1 username username 125M Jun  5  2020 model.ckpt-0.data-00000-of-00001
  -rw-rw-r-- 1 username username 787K Jun  5  2020 model.ckpt-0.index
  -rw-rw-r-- 1 username username 2.4M Jun  5  2020 model.ckpt-0.meta

and then things like

> cat /etc/shadow

do actually work...

Quite interesting that it will make subtle errors in its otherwise reasonable-looking answer, e.g. "kipedwaia" has two "a"s; "kawipieda", "kipedwaia" and "pwakieida" have only two "i"s.

I have seen reports that it will happily hallucinate a plausible but wrong answer to all sorts of different prompts, intermixed with many mostly correct answers. It's interesting to think about how to place trust in such a system.

So, if it's true that these KH-12 spy satellites are basically modified versions the Hubble space telescope, would it be reasonable to assume that Webb will get copied in a similar way for spy satellites?

Your comment brings focus to the question of why reviewers agree to do these unpaid reviews in the first place. I assume your comment is in jest, and you were fully aware of the unpaid nature of the work. But what was your motivation to do it anyway and spend 2000 hours on it?

This way you would trade in a null-byte-terminated variable length string for essentially a null-bit-terminated variable length number (plus the remaining string). I am not convinced that this actually would be much safer.

The advantage 2 was only released in 2016. I myself have the advantage (1) since about 8 years. Sorry for being pedantic here, but that's why that statement tripped me up.

I have a couple of pants made from G-1000, and I really like it's properties (light, durable, will not retain much water, can be made water-repellent with wax), but it actually is made from 65% polyester and 35% cotton [1].

[1] https://www.fjallraven.com/uk/en-gb/about-fjallraven/materia...

There must be some information content left in stock price time series data, as evidenced e.g. by the price Momentum factor [1], which was been replicated in a number of studies (e.g. [2]), observable over the last couple of decades.

[1] https://en.wikipedia.org/wiki/Momentum_investing

[2] https://papers.ssrn.com/sol3/papers.cfm?abstract_id=2961979

> The US economy around 1800 was the first economy in history to produce a consistent food surplus.

Can you provide a source for this claim? Thanks.

If the number of tests administered is the limiting factor, and if you only increase the number of tests linearly in time, then case numbers would grow linear as well.

Now the perpetual licensed version always lags 12 months behind in updates, i.e. you do not have a perpetual license for the past 12 months worth of updates at the time your subscription runs out.

It's interesting that this story has already vanished from at least the first 100 'news' entries on HN, 9 hours after posting, and currently at 400 comments.

I wonder if Netflix would have to spin off their in-house productions once they reach $25B revenue. Or, if they don't, if the mobile platform providers could change their business models into similar subscription services to avoid the break-up.

I have read about some difficulties of running SLURM on POWER9 systems, so maybe IBM proposed/insisted to run their own POWER9-tested scheduler?

And if your employer goes bankrupt, there is an additional 3 month of guaranteed salary ("Insolvenzgeld", by the state) afaik.

Most HPC cluster installations I know (small to medium edu sites in Germany) use CentOS / RedHat. Source: am part-time HPC sysadmin

There is, however, http://www.aosabook.org/en/pypy.html , which is indeed self-hosting.

The article mentions pooling layers as one source for the high frequency patterns. Jeffrey Hinton recently introduced capsule networks (https://news.ycombinator.com/item?id=15609402), in part because he wants to get rid of pooling layers. Maybe this approach is effective to counter (at least visually indistinguishable) adversarial examples.

edit: ok, someone already tested it, and it does not seem to help that much: https://github.com/jaesik817/adv_attack_capsnet

I did something similar in Montreal (Bixi) a couple of years ago, in order to determine when I had to leave the house so that there was still a bike left at my station to go to uni.

While I really like my x230, the touchpad is practically unusable for me. It's just too small (or too jerky, if you use a large scaling factor). And while I have the higher quality IPS panel, it's still just 1366x768. Just give me the x230 with a bigger touchpad and a nice HD display, and I am happy (maybe x260/x270 would work? ...).

These technical analysis features can become 'self fulfilling prophecies', if a significant number of market participants believe in them (meaning it's a psychological effect). Other than that TA seems to be mostly BS.

Another recent review paper: Deep Reinforcement Learning: An Overview, https://arxiv.org/abs/1701.07274

This post is from 2013, maybe add this to the title?

So far I didn't use their service, but I fondly remember their security CTFs, which were lots of fun, and very well run. Still hoping for another incarnation of those...

I was locked out of my AWS account for about 2 years due to loosing my 2FA authenticator on my phone. But had no open bills and didn't really care about it, so I did not try to gain back access.

Until last week, when I decided that I need to use the account again. I simply clicked the recovery link on their 2FA login form. Had to enter my account details and my phone number. Within 15 minutes (as stated on the website) a member of their service team called me. I explained the situation (lost 2FA app). Not a problem, the service agent told me. He then sent me an e-mail (to my AWS-associated mail address) with some random characters, which I had to read back to him. After that he simply removed 2FA, and I was able to log in with my username/password. For me the experience was quite pleasant, but had someone else have access to my e-mail account and knew my AWS username/password, he might have been able to take over my account, bypassing 2FA. I don't quite know how to feel about that.

EDIT: Ok, this probably only works if your phone number is also already associated with your account. So an attacker additionally would need access to my phone number, making things quite a bit more secure in my view.

I am not an expert, but maybe look into recurrent neural networks [0]? These things can turn a sequence of inputs of possibly arbitrary length into a fixed size internal representation, and can output a single value derived from this representation.

[0] https://en.wikipedia.org/wiki/Recurrent_neural_network