fnbr

A mix of startups, smallish research labs (that is, smaller than Google Brain/FAIR/DeepMind), small, profitable AI companies.

I focused on smaller companies as that’s what I wanted. Bunch of larger companies were hiring though.

Yeah, you should be familiar with the state of the art in your field. You don’t need to be as strong with theory though as Research Scientists.

Yeah, it was dumb. One company gave me an offer to lead engineering for their LLM team and another wouldn’t even interview me because I didn’t have any LLM experience.

Clearly one of them was wrong.

4 year BSc + 1 year MSc. Experience matters much more than having a PhD. Lots of people without phds at the big ai research labs as engineers. I’d say it’s uncommon for engineers to have phds (although almost all have masters).

I was an L5, so significantly below $500k.

I work as an AI research engineer, so I’m having a totally different experience than the others I’m seeing here. Job market is very very strong for senior AI roles. Multiple offers that were >2x what I was making at a BigTech research lab (depending on how you value the equity). Interview process was much easier than I was expecting- mostly situational interviews (although I was referred in for those).

Having said that, many places refused to interview me because I didn’t have experience with LLMs. I was surprised how strict some places were with experience requirements.

parameter tuning is pretty necessary, according to anecdotes. People on twitter have got good results by changing the default parameters.

What did Uber do?

The betting strategy martingale is an example of a mathematical martingale.

Not in my office. No one met that description.

They shut my office down in Canada. Lots of jobs. Most are remote for US companies though. Very few recruiters spamming me, but companies are very receptive to inbound inquiries from me.

I’m a senior eng in AI research though, which is obviously having a big moment right now.

Yeah. DeepMind has released various MCTS implementations (eg in OpenSpiel). Not the same as releasing AlphaZero.

What specific hardware acceleration is possible? I think TPUs are already as good as you can get for accelerating large MLPs (half of what a LLM is). Is there that much overhead for the attention blocks?

This is very expensive to run. I bet they’re not going to have particularly high margins with this. Each response probably costs them several cents.

Presumably he did a really good job running OpenAI- it’s not like it was pre-destined to become the company it is today.

If you can stitch something together though, you can start to get the data you need to train an end-to-end model. And once you get that at any reasonable level of proficiency, you should see the ML flywheel spinning. The hard part is stitching everything together.

Were you able to install it with your existing furnace or did you have to replace the furnace?

> Unlike Lebanon, though, one cannot legally operate their solar panels to power a home while the electric grid is down in Israel.

That’s fascinating! Can you say more?

How are you tuning it to make sure it doesn’t hallucinate? I think that’s a fundamental property of neural networks (see, for instance, the adversarial examples literature, or the recent work demonstrating adversarial exploits in AlphaZero style agents in Go).

this is why I don’t think google is threatened by LLMs. 1) this doesn’t do links, and 2) it has accuracy issues. I don’t understand the hysteria about how google will be replaced by ChatGPT.

Normal retail price. They were probably hoping to wholesale it for 42 per case, which is what they’re selling it for now.

Contact Meta, they might still want to hire you.

I like your edited version!

I don’t follow your critique for #2. SVMs, Random forests, etc., aren’t the counterexample to Rich’s post (for anyone who knows him, Rich doesn’t even particularly _like_ neural networks). The counterexample is hand crafted features.

A counter example would be showing a number of successful examples in, say, computer vision, where handcrafted features do better than learned features. This is largely not the case. In, say, both NLP and Computer Vision, learned features dominate, even at companies with less compute (they use pretrained models).

(Disclaimer: I work with Rich.)

What’s an example of a watch that’s mechanically identical to a Rolex?

This is so true. I can’t believe how much I love having a kid. And mine’s only 6 months old. I’m beyond excited to actually do things with him, like run around with him, or go skiing/ride bikes.

It’s very hard to use CFR in Stratego because you can’t represent the hidden information in memory. In poker, the hidden information is a 52C2 vector, so you can easily pass it around. Not so in Stratego, so you have to do something else. RNAD has similar theoretical foundations though.

I don’t think it is more challenging than StarCraft or Dota. Does the blog post claim that anywhere?

Stratego is way more challenging than Poker, though. StarCraft/Dota/Stratego have the property that you can’t represent their imperfect information as a vector in memory, whereas you can easily do that in Texas hold’em poker (there’s only 52C2 = 1024 possible hands). So for those games, you have to use an approximate distribution rather than the exact one.

I’m an author on this paper (although my contributions were relatively minimal) and on the Player of Games paper (which did Poker, Chess, and Go).

I don't get it. Why can't Google just train their own LLM and use that for answer cards?

The main value of Google Search is the ability to search the web for websites. Not to search for answers.

Yeah. It’s one thing if the only way the job is worse is in pay, it my experience has been that if the pay is significantly worse, the job is worse off in every way.