alibero

> Well first we should rob a bank...

This is sometimes called "instrumental convergence" in the AI safety world. Certain things (money, compute resources, safety from being turned off) are generally useful for an AI, and so we'd expect that a misalligned AI would attempt to acquire these things if given almost any substantial task.

Thank you for this excellent post series. It reminds me a lot of the pre-LLM days, though I was mostly using LSTMs back then. When the original GPT paper came out, I thought the future would be using LLMs to generate tons of synthetic labeled data and then training specialized LSTM or transformer models per-task.

Had a couple of questions:

1) You note that ARC-AGI is a meta-learning task, have you tried any meta-learning algorithms such as MAML?

2) Do you think this approach could extend to ARC-AGI 3? Or do you think the interactive environments require a higher level of complexity than what can be achieved with a small model?

The current trajectory is also a mass dying of humans so hopefully we can thread the needle there.

I've been looking into this. There seems to be some mostly-repeating 2D pattern in the LSB of the generated images. The magnitude of the noise seems to be larger in the pure black image vs pure white image. My main goal is to doctor a real image to flag as positive for SynthID, but I imagine if you smoothed out the LSB, you might be able to make images (especially very bright images) no longer flag as SynthID? Of course, it's possible there's also noise in here from the image-generation process...

Gemini really doesn't like generating pure-white images but you can ask it to generate a "photograph of a pure-white image with a black border" and then crop it. So far I've just been looking at pure images and gradients, it's possible that more complex images have SynthID embedded in a more complicated way (e.g. a specific pattern in an embedding space).

Later down the line, if you want to have separate behaviour for task deadlines vs payment deadlines, you're going to have to go through your codebase and look at every call to set_deadline and figure out if it's being used to set a task deadline or payment deadline. If you have an inkling that the deadlines might need a different behaviour, the “good example” can save you an annoying refactor in the future.

I used to work at Yelp, which had something that I think it similar to what you are describing called Data Pipeline (https://engineeringblog.yelp.com/2019/12/cassandra-source-co...).

I remember it being pretty simple (like, run one or two bash commands) to get a source table streamed into a kafka topic, or get a kafka topic streamed into a sink datastore (S3, mysql, cassandra, redshift, etc). Kafka topics can also be filtered/transformed pretty easily.

E.g. in https://engineeringblog.yelp.com/2021/04/powering-messaging-... they run `datapipe datalake add-connection --namespace main --source message_enabledness`, which results in the `message_enabledness` table being streamed into a (daily?) parquet snapshot in S3, registered in AWS Glue.

It is open source but it's more of the "look at how we did this" open source VS the "it would be easy to stick this into your infra and use it" kind of open source :(

Check out figures 1 & 2 in the Llama-2 paper :) They benchmark against ChatGPT for helpfulness and harmfulness

https://ai.meta.com/research/publications/llama-2-open-found...

It seems like VR is less than half of the investment by RL. In Meta's 2022 annual report, they say "Many of our metaverse investments are directed toward long-term, cutting edge research and development for products that are not on the market today and may only be fully realized in the next decade. This includes exploring new technologies such as neural interfaces using electromyography, which lets people control their devices using neuromuscular signals, as well as innovations in artificial intelligence (AI) and hardware to help build next- generation interfaces. ... *in 2023, we expect to spend approximately 50% of our Reality Labs operating expenses on our augmented reality initiatives, approximately 40% on our virtual reality initiatives, and approximately 10% on social platforms and other initiatives.*"

I'm not sure if Horizon falls into "virtual reality" or "social platforms" but it seems to be the latter: "For example, we have launched Horizon Worlds, a social platform where people can interact with friends, ..."

As one of the unsupervised dorks working on LLMs at Meta (not one of the authors here) I took it in a positive way :)

I feel like this is a boring answer but for me, I had to make a habit of it, and then it didn't feel so hard any more. I started doing a master's degree in my free time: when I started I barely had the energy to do anything outside of work, but now I feel like setting aside time for coursework is pretty natural.

For me I find I usually fall into a 2/2/2 pattern for forming habits*: The first two days are super hard, after about two weeks it starts to feel doable, after two months the habit is pretty set and I don't have to worry as much about falling off the bus.

* This entire pattern is probably a placebo but that's fine by me

Of course, your energy is not infinite. If you are trying to work crazy hours and fit in other taxing activities, you are going to fail at some point.

> Finally, RLHF, or "RL with Human Feedback". This is a fancy way of saying that the model now observes two humans in a conversation, one playing the role of a user, and another playing the role of "the AI", demonstrating how the AI should respond in different situations. This clearly helps the model learn how dialogs work, and how to keep track of information across dialog states (something that is very hard to learn from just "found" data). And the instructions to the humans are also the source of all the "It is not appropriate to..." and other formulaic / templatic responses we observe from the model. It is a way to train to "behave nicely" by demonstration.

I think this misses a big component of RLHF (the reinforcement learning). The approach described above is "just" supervised learning on human demonstrations. RLHF uses a reinforcement learning objective to train the model rather than maximizing likelihood of human demonstrations. In fact, you can then take the utterances your model has generated, collect human feedback on those to improve your reward model, and then train a new (hopefully better) model -- you no longer need a human roleplaying as an AI. This changed objective addresses some of the alignment issues that LMs struggle with: Open AI does a pretty good job of summarizing the motivation in https://arxiv.org/abs/2009.01325:

> While [supervised learning] has led to markedly improved performance, there is still a misalignment between this fine-tuning objective—maximizing the likelihood of human-written text—and what we care about—generating high-quality outputs as determined by humans. This misalignment has several causes: the maximum likelihood objective has no distinction between important errors (e.g. making up facts) and unimportant errors (e.g. selecting the precise word from a set of synonyms); models are incentivized to place probability mass on all human demonstrations, including those that are low-quality; and distributional shift during sampling can degrade performance. Optimizing for quality may be a principled approach to overcoming these problems.

where RLHF is one approach to "optimizing for quality".

I was under the impression you got a 60 day grace on the TN as well: https://www.ecfr.gov/current/title-8/chapter-I/subchapter-B/...

"An alien admitted or otherwise provided status in E-1, E-2, E-3, H-1B, H-1B1, L-1, O-1 or TN classification and his or her dependents shall not be considered to have failed to maintain nonimmigrant status solely on the basis of a cessation of the employment on which the alien's classification was based, for up to 60 consecutive days or until the end of the authorized validity period, whichever is shorter, once during each authorized validity period. DHS may eliminate or shorten this 60-day period as a matter of discretion. Unless otherwise authorized under 8 CFR 274a.12, the alien may not work during such a period."

This is also what I've been told by my company's lawyers.

What I’m seeing is that some big tech companies these days (including my own) have standardized rubrics [1] for leetcode-style questions that focus on a few areas (like DS&A, communication and coding style) so I can’t really give any marks for these other positive behaviours.

There’s some benefits to rubrics (reduces differences between interviewers and is more fair to minority candidates from what I’ve seen) but I’m it definitely impacts or senior hiring.

Unfortunately there are also so many experienced developers in the hiring pipeline who actually can’t code, so doing at least one coding interview seems inevitable. I’d give less “tricky” questions but, like rubrics, the questions are standardized too :/

[1] https://blog.tryexponent.com/google-coding-interview-rubric/...

Yes, I skipped or paraphrased those parts of the bill to keep things short in a way that I thought made sense. But I think the message is unchanged with the full text. Namely that Section 230 would be amended to also state:

"NO EFFECT ON CHILD SEXUAL EXPLOITATION LAW. Nothing in this section [NB Section 230] (other than subsection (c)(2)(A)) shall be construed to impair or limit ... any charge in a criminal prosecution brought against a provider of an interactive computer service under State law regarding the advertisement, promotion, presentation, distribution, or solicitation of child sexual abuse material, as defined in section 2256(8) of title 18, United States Code;"

And so Section 230 protections to content providers would cease to apply* in cases of child secual exploitation law, I think.

* EDIT: Except for those points that would be added to Section 230 specifically regarding E2EE

We do background checks, so hopefully we're not actually hiring people who tell big lies on their resumes.

> IMO for the hopper it is also not easy I agree. I'm changing jobs soon and have some apprehension for the reasons you mention, plus don't forget the fear of underperforming and being deported for losing my job :/

> the goal of the job is to get paid and support your family Parents and people who pass the vibe check get strong yeses from me so no issue there /s

But in all seriousness I do feel really bad when interviewing parents or anyone with other obligations that stop them from grinding practice questions if I have to fail them because they can't solve some BS leetcode question. Unfortunately my intuition if someone is a good engineer or not doesn't matter much if they can't crank out some leetcodes in 35 minutes :/

It's ok if my employer's competitors are advantaged because I'm switching jobs soon to get that sweet sweet pay bump anyways :)

Guess it's time to start giving strong no hires to job hoppers then /s

The part of the bill that mentions E2EE (Section 5) is an amendment to the Communications Act of 1934, namely the famous Section 230 which contains: "No provider or user of an interactive computer service shall be treated as the publisher or speaker of any information provided by another information content provider."

So the EARN-IT act would seem to me to modify Section 230 to not apply in cases of child sexual exploitation law, importantly "any charge in a criminal prosecution brought against a provider of an interactive computer service under State law regarding the advertisement, promotion, presentation, distribution, or solicitation of child sexual abuse material". However despite this amendment, using E2EE would not "serve as an independent basis for liability of a provider", whatever that means.

This seems more notable to me than the whole "creating a committee to create best practices" sections but I could be misreading or misinterpreting the bill honestly, I'm no expert.

In my experience, being familiar with inductive proofs is pretty useful for programming. For most non-trivial code involving recursion or loops the way I personally "know it works" usually has the flavour of an inductive proof.

For example, I might not remember how to code bisection search but I can figure out the loop invariant and from there it's easy enough to code a working binary search. And even if you have right bisection search down to muscle memory, you can modify the loop invariant to create left bisection search if you need it, while if I tried to keep two binary search algorithms in my head I feel like it would be more error-prone.

Waterloo also offers a BCS which has fewer math requirements than the BMath: https://cs.uwaterloo.ca/current-undergraduate-students/major...

The BMath option is only required if you want to have a double major in CS and a different area of math. Otherwise the BCS is strictly more flexible: you have 5 fewer math courses and 5 more electives (which you may use to take math courses if you really want to). From my experience most people are graduating with BCS unless they really like math.

The relationship is that computing the SHA hashes takes polynomial time. One way you can think of NP is that it is the set of problems whose solutions that can be deterministically verified in polynomial time. So for any hash computable in polynomial time, the pre-image problem is in NP.

Note that this doesn’t prove the security of SHA256, it just says that to prove it secure would be to prove P != NP. You could still prove SHA256 insecure and that proof could be totally separate from P =? NP.

You’re correct that there is no proof of the security of SHA1. The existence of any one-way function would imply that P != NP.

And if it turns out that P = NP then it will turn out that most of the cryptographic guarantees we rely on today will be unrealizable on classical computers.

Quantum computers may not help us as it is currently unknown if quantum computers are more powerful than classical computers in terms of time complexity (it’s strongly suspected that this is the case though).

I had a similar experience but I don’t really regret choosing not to skip grades. I didn’t really bloom socially until I hit university either but I think being a 14 year-old freshman would have probably changed how that played out.

Yeah, the MNIST model seems to have issues. I can get it to make mistakes for all line widths and most digits. Maybe the issue with small digits is that there is no cropping happening and the model was trained on MNIST where all the digits are uniform size.

Actually this is probably a good demo for a ML visualization tool as we were pretty easily able to find these issues. If cropping is the issue, this is also something that a test set might not catch.