redredrobot

Are you complaining that when they cut their pay from 400k to 32k that wasn’t enough? Or are you complaining that they didn’t cut their 32k salaries down further?

They can both be true. It's can be easy to make the switch and still difficult to convince people to do it if the alternatives are simply crappier.

The argument is - engineers are expensive so why pay for the expertise to setup and run machines? There's just so many better things for your company to be spending the money on.

I think there are a bunch of people who consider this work unethical or at least deeply in the grey. The negativity isn't that surprising

If they are used in a tool that lets you generate someone's likeness as part of user-specified new content, yes. But unlike 15.ai that isn't their core purpose and no such tool exists.

I would imagine it will end with a similar outcome to video game likenesses - a person owns their likeness and you can't create products that includes their likeness without their consent.

You're right about the compute part being wrong. I never said it wasn't legal, just that they took someone else's work to train it. I would hope that voice synthesis is illegal without permission from the voice's owner, but I imagine it is untested so far.

But it's not just about the popup - it 's more that when your work is fundamentally about using reusing someone else's character, it feels pretty hypocritical to be so focused on making sure you get credit.

The author took someone else's IP as training data, trained a model on someone else's compute, and then gets extremely bent out of shape when others use the model without crediting them?

I don't see how this can possibly go anywhere if employees aren't willing to put their names on it.

Here is a 1997 interview with Jeff Bezos that I like where he talks about why he created Amazon - https://www.youtube.com/watch?v=rWRbTnE1PEM. He describes a purely logical explanation of seeing data about internet growth and identifying books as being a uniquely good product to sell on the internet.

Which is not to say he wasn't "willing to dedicate [himself] entirely to solving the problem", but I think it agrees with your take about finding market opportunities rather than following a passion.

You're pretty quick to act like there is a lot of blame to be allocated, but GTA 5 is arguably the most successful game in history. With games that truly have technical problems so bad they ruin the game, people don't complain - they just don't play it.

Maybe a long load time just wasn't that important for the success of GTA.

You tell the interviewer you are familiar the question instead of pretending to come up with a solution on the spot. Is that not obvious to everyone?

But don't we already know that composition exists in DALL-E? Don't the points shown in the tweet indicate that some form of composition exists? The 3D renders are clearly render-like, the painting and cartoons are clearly in the appropriate style.

So his argument is that the text clearly maps to concepts in the latent space, but when composing them the results are unexpected, so it isn't language? Why isn't this better described as 'the rules of composition are unknown'?

Are you sure about that? I tend to think that Evernote's differentiation is the vast feature diversity. IMO old software like this often continues to succeed because it can support the long-tail of use cases in a way that newer software does not.

Yeah - it’s probably unfair of me to say it doesn’t scale at all. But between large data and 2 extra orders of magnitudes of rows, the single SQLite file approach quickly breaks down, even if you don’t store the large content in-db.

Datasette is pretty cool.

But AFAICT, it just doesn’t scale whatsoever. That SQLite db is both the dataset index and the dataset content combined, right? So you're limited by how big that SQLite db can realistically be. The docs say "share data of any shape or any size", but AFAICT it can't handle large datasets containing large unstructured data like images and video and multi-billion data point datasets are hard to store in a single machine/file.

Not really a criticism, but more wondering if there are scale optimizations in Datasette I'm not aware of since the docs do say any shape or size.

Completely untrue. Just finished up my interview process and my experience was that comp is higher than ever at the strong FANGs (I.e not Netflix). Microsoft and Amazon just massively bumped their comp maximums. Google was good as ever. Pre-ipos cash comp is still very strong although they are currently relying on previous stock valuations so the comp isn’t as good as they claim right now

I actually think it's probably pretty simple underneath the facade - Musk offered to buy Twitter for a lot of money. The market tanked and Musk doesn't want to pay that price any more. He thinks he can strong-arm Twitter and either not go through with the purchase or force them to accept a lower price. The law does not support this, but Musk has a history of being unbothered by the law and there are examples (not Musk) where this tactic has worked to reduce the purchase price by hundreds of millions of dollars.

Everything else is just window dressing.

Did you actual address your addictions or did you just fill that hole with something new and so moderation became easy? I would suggest stop focusing on the symptoms and focus on the underlying problem that is driving your addictions, whether that is unhappiness or persistent habit triggers or whatever it might be. If you "quit food" will you just be writing this same post next year but with a new addiction?

More concretely:

- Identify why you need an "escape valve". Understand that having an escape valve that dominates your life negatively is the problem, not the shape that valve takes

- Identify the triggers that push you to the escape valve. Both the long-term triggers (for example, it could be being stressed or unhappy) and the immediate habit triggers (for example, it could be seeing paraphernalia or being extremely hungry). Try to reduce the long-term triggers. Try to develop new habits around the immediate triggers (trigger still exists, but habit response is something you want to do). Being aware of your habit loop is IMO important for improving how you react to triggers (therapy can be really helpful here)

Elon didn't do any non-public due diligence of Twitter

You can't just quote a part of a contract and specifically cut out the legal caveats that exists to avoid the risk of being considered material misrepresentation. The correct quote is:

> We have performed an internal review of a sample of accounts and estimate that the average of false or spam accounts during the first quarter of 2022 represented fewer than 5% of our mDAU during the quarter. The false or spam accounts for a period represents the average of false or spam accounts in the samples during each monthly analysis period during the quarter.

> In making this determination, we applied significant judgment, so our estimation of false or spam accounts may not accurately represent the actual number of such accounts, and the actual number of false or spam accounts could be higher than we have estimated

I worked at a couple of dev tool startups that required you to invest some effort to adopt and would require effort to move away from (e.g. if the company went under) - open sourcing ended up being basically a necessity to sell the product (enterprise sales during seed and A). YMMV, but I would definitely encourage open sourcing enough to make potential adopters feel comfortable that they won't have to do a big, urgent migration and/or lose a bunch of engineering investment if you go under or decide to pivot.

I have no idea what you think that proves.

What I know of your experience shows a low number of years of experience, a lack of papers, and a lack of true hands-on experience at the small number of companies in the world that have the resources to truly investigate large models. How can you know so much about LLMs without ever having the resource to train one?

I'm obviously not going to dox you, so you can easily just dismiss what I'm saying. But even just reading through your HN comments shows arrogance in your own knowledge (across multiple domains).

A specifically memorable quote is:

> I frequently create unique on the internet [words]

This is very true. Your erudition is apparently only matched by the uniqueness of the words you use when on the internet.

> I have more expertise in deep learning than anyone else here

No, you don't. Looking at your experience, there is simply no way that you are the foremost expert in DL on HN.

Python is a great option because it is high-level, easy to pick up, powerful and has a library for anything they are interested in pursuing. Something like streamlit is cool because you can write a small amount of code to build a useful web app that you can interact with. For data viz related things, Jupyter + Python (maybe Pandas, not SciPy or NumPy unless that sort of math is specifically interesting to them)

Javascript and webdev is also a decent option since you can build such interesting and shareable web apps quickly, but the frameworks and scaffolding can be intimidating and painful.

I would work backwards from a project/interest and pick a high-level language where it's easy to get a good enough solution and create something that feels like an accomplishment. I would want to teach both programming as well as the joy of programming.

There was a Robinhood breach a while back - https://blog.robinhood.com/news/2021/11/8/data-security-inci...

Currently on the job market in the AI space in the Bay Area - 400k to 600k is senior level at FAANG + similar. Low 7 figures at staff wouldn't shock me (although I don't have any actual data on that)

There are some groups defining AGI in a way where a 20 year timeline is aggressive but not impossible. The real question is how is AGI being defined by the author.

> using AWS is fine, as long as you use cloud agnostic stuff and avoid AWS specific services

IMO this is exactly when you should not be using AWS. AWS is a premium offering and you can find cheaper alternatives for the commoditized offerings like EC2, RDS (the older, non-AWS-specific features) and (maybe) S3.

It's the high-level services where the real value is in AWS - the cutting edge technologies that only exist in a specific cloud (or multiple clouds but where it's difficult to port across clouds due to subtle differences). This is where the extremely high development cost is able to be amortized over many customers, allowing you to get scalable, easy-to-use, and easy-to-operate services at a much lower cost than would be possible outside of that cloud. A great example of that for AWS is Serverless Aurora.