cweill

Yes but if you're in a Jupyter notebook, you may not be directly connected to a DB. If you're using pandas, this unlocks some scalability before needing dask and a cluster.

If you ever need to join two large dataframes, but are OOMing on the join, write them to disk as parquet files then use DuckDB to do the join. It's amazing what you can do on one machine thanks to DuckDB.

Do you think he was on garden leave? Some kind of non-compete?

I also have this question. Is the RL MDP actually encoding cause and effect? Or just learning (bidirectional) correlations between states and actions?

I wonder if Pearl thinks that RL replicates his do-calculus under the hood, or if that's an innovation we're missing.

For reference, most gaming channels in the US have an RPM of $4-$5. I'm guessing their RPM is $1 because the channel owner is based in Europe. I imagine toy channels have higher RPMs in the US.

Look into Coach Sommer and https://www.gymnasticbodies.com/. Probably the best guide from beginner to advanced out there. I've been following their program for over 15 years, and there's something for every level of strength and flexibility. It's a program you can follow your entire life.

The big thing with calisthenics/gymnastics is developing the tendon and ligament strength in your joints which take several months/years, but is the foundation to doing the advanced stuff like planches and levers. Good luck!

How far back can this go? Would you be able to do 1-3 years of daily data?

Having been on both sides as a customer and foundee, I love subscriptions.

Every business needs a business model. Subscribers is a very clear one. I'm untrusting of any service that's free and doesn't rely on ad-revenue, because they have incentive to make money in other shady ways like selling your data to third-parties who will use it against your best interests.

Subscriptions are a nice business model because they are predictable costs/revenue to both customers and the business.

What I don't like are subscriptions that don't grandfather you into higher pricing, and pay-as-you-go plans, because it's so easy to forget about then and their rules and get charged a nasty bill later.

As a customer IMO subscriptions are the lesser of all the evils, and the best alternative to ads, eg YouTube Premium.

Google colab is $49/mo to get an A100. Trust me when I say you can build multimillion-dollar ML companies with just that (and maybe $100 extra dollars per month of spot credits).

I disagree with this comment, and anyone reading it should take it with a big grain of salt. Let's go back to 2016 and replace "LLM" with "Reinforcement Learning". Everyone thought every problem could be solve by RL because it's a looser restriction on the problem space. But then RL failed to deliver real world benefits beyond some very specific circumstances (well defined games), and supervised learning is/was still king for 99% of problems.

Yes, LLMs are amazing but they won't be winning every single Kaggle competition, displacing every other ML algorithms in every setting.

So many accomplishments are the result of someone just putting one foot in front of the other with a vague idea of where they're going. Hopefully, their north star is something others find respectable once achieved. They're then labeled "geniuses" if it works out and "fools" if it doesn't. Accomplishments are only impressive with hindsight.

There's no cabal keeping small creators small. Small creators need to make better content, i.e. provide value to their audience, and they will grow.

Making content isn't about sharing every little thought. It's about being consistent, finding a niche, and providing value. It's just like running a business.

I feel like you're being a little myopic if you feel that all of programming is logical. Look into the probability stuff like machine learning. Read books from Nassim Taleb like Antifragile to understand that it's very easy for us to convince ourselves that logic and increasingly complex models solve everything. In reality there are many meta-strategies you can use like "via negativa" to still be rational, but make the world appear less crazy, and make the crazy in the world look more obvious, and avoidable.

Hackers and painters

http://www.paulgraham.com/hp.html

Yeah, tree based models are great for tabular datasets that are primarily numeric, with only a few categorical variables. But as soon as you categorical variables have a 1000+ potential values that need 1-hot encoding or if you have any natural language text associated with your rows, deep learning almost always outperforms, especially if you have over 50K instances in my experience.

The major downside of DL is the slow training, and therefore slow iteration feedback loop. Couple that with an exponentially growing number or hparams to tune, and you get something very powerful but costly in terms of time to use.

But if you want the best possible accuracy, and data collection isn't expensive, DL is the way to go. Just expect to spend 10x the amount of time tuning it vs trees to get a 10% to 20% reduction in error.

Just ignore them. On social media, ignoring a stranger is not rude. It's the default. And a perfectly valid strategy for dealing with the volume of people asking for your valuable time.

Obviously, when it's someone you know or important to you, give them the courtesy of a polite decline or say you're very busy and will let them know if anything changes in the future. But for strangers: ignore if you don't want to interact.

Congratulations on getting your first 100 paying users. Most startups never make a single dollar.

How long did it take you to hit that number? Many of us devs want to do SAAS but don't understand what's a reasonable amount of time to expect until we can get "ramen profitable".

Hey Minimaxir again! This is really funny because I built an entire company around this idea: optimizing titles for YouTube videos: https://CreatorML.com.

Similarly, the user can use GPT-3 to generate titles, and a secondary ranker model scores them for estimated viewership. I use a regression model however.

Anyway, just funny that "great minds think alike" :)

Couldn't you use something like Huffman encoding to compress the entire will and print it on a single sheet of paper, then the estate could use OCR to scan and reverse the compression? If you're using ASCII characters only, I assume the compression could be very high.

I see you've read Zero to One. :)

Indeterminate pessimism also means they aren't taking economic risks, whereas many Gen-Z are becoming content creators or starting businesses with their savings, so it could be more determinate optimism. Possibly naive, but time will tell.

I made a video about this: https://youtu.be/YcJN_ZiFw9w

Hey Jeremy, i just want to say that I love your course and the way you teach. I refer everyone to the Fast AI in my YouTube videos on getting started with machine learning. Please keep up the great work!

Sounds to me like you might grow to resent this individual if you gave them shares. I agree with everyone else: give them a % revshare for hitting milestones. No shares this late in the game.

Also, in this looming recession, equity will be less sexy to hold for employees. So use that to your advantage. Good luck!

Probably only for companies without a business model.

Sentry is amazing, I pay for the business tier for three different Saas products I'm building. It's a game changer.

I'm Poor Charlie's Almanac, Charlie Munger explains that cheating is a huge problem, and that if you create a game where people can cheat to get ahead, they almost inevitably will.

At Google you could cheat the promotion system. Just spend more time optimizing for promo than for improving products for your customers, and you would be handsomely rewarded. You end up with the cheaters becoming the leaders and the good engineers leaving in frustration when they need to take orders from cheaters.

I disagree. There is a ton of low hanging fruit in applying SOTA ML techniques to niche problems. It just requires a bit of business sense, but there is a lot of money to be made there. More than what I was making as someone doing ML research at Google Research.

The Deepmind lead behind AlphaStar, there StarCraft 2 bot that beat several world champs, once said, "the best way to solve a reinforcement learning problem is with supervised learning". That's because supervised learning is relatively extremely efficient, and just RL with more constraints.

Just start learning the basics of supervised learning for classification and regression on common benchmark tasks like CIFAR and UCI. Apply a mix of linear models, neural networks, and trees like random forests and GBDTs. Next try convolutional networks for vision and transformers for NLP. You'll be all set to solve most real world problems.

The more assumptions you relax, the more general the algorithms become, for example going from immediate reward to delayed reward means going from supervised to reinforcement learning.

The trade-off is the more general algorithms needs many times exponentially more data and compute to come to a similarly good solution.

That's why reinforcement learning has seen so practical few applications relative to supervised learning. There's no free lunch.

That said, as a ML practitioner I would love it if I could just apply a single master algorithm to all problems, but that is likely many years away.

I've been working a Full Time Employee in the industry for over 8 years at multiple companies. Never once has HR done anything remotely useful to help me when there were family problems or tragedy, except maybe a couple weeks of paid leave.

But I would like to hear the take from a founder who built an HR team to know if maybe I am missing something.

I'm really curious if it's really different being an employee vs a contractor in that respect.

Can someone please share the current state of deploying Pytorch models to productions? TensorFlow has TF serving which is excellent and scalable. Last I checked there wasn't a PyTorch equivalent. I'm curious how these charts look for companies that are serving ML in production, not just research. Research is biased towards flexibility and ease of use, not necessarily scalability or having a production ecosystem.