Georgy Khalin

Quant dev lead at Deep understanding is the only real leverage

[Georgy Khalin @gkhalin]

@0xfdf The same thing happens in trading: someone from a top shop can be very expensive to hire, but the chances that they understand something useful deeply enough are often low. The market has adverse selection

[Georgy Khalin @gkhalin]

Most senior individual contributors haven’t realized they’d make great tech leaders until AI forced them to

[Georgy Khalin @gkhalin]

@maharshii To be truly exceptional at something, you better focus on what comes naturally to you

[Georgy Khalin @gkhalin]

@Flomerboy Lead the meeting and really try to understand other people

I still write meeting logs manually to make sure I understand everyone correctly and keep everyone in sync

[Georgy Khalin @gkhalin]

I don’t buy the game theory that you should never reveal full information

I was told I shouldn’t be saying some of the things I say

I believe radical honesty is important. Don’t say what you don’t believe. Be as honest as possible

That’s how trust is built. Being honest has never let me down so far

[Georgy Khalin @gkhalin]

@staysaasy Just make sure you don't have to lie

[Georgy Khalin @gkhalin]

@staysaasy I feel like it’s ok to share your beliefs, but you have to be honest about how confident you are. Just manage expectations properly

[Georgy Khalin @gkhalin]

@iamKierraD Looks like a lot of bright young people happened to be around experienced people they could learn from, especially their beliefs and vision. That’s a huge unlock

[Georgy Khalin @gkhalin]

@ivanburazin How do you build an RL dataset from your own data? Is there a guide?

[Georgy Khalin @gkhalin]

@Azaliamirh @JeffDean @Sanjay_Ghemawat @quocleix @OriolVinyalsML Jeff Dean has also left 😕 Tough times for google, wonder what's going on

[Georgy Khalin @gkhalin]

Trusted abstractions with clear invariants is the fastest way to verify LLM output

Otherwise you get a slop report and have to second-guess every result until you review all the new code that produced it

[Georgy Khalin @gkhalin]

@Gemini It’s a risk/profit trade-off. The answer depends on how you price risk

[Georgy Khalin @gkhalin]

@pwlot @VictorTaelin How do you use council? Is it role based with complicated prompts?

[Georgy Khalin @gkhalin]

Given that test time compute improves quality, it seems like a lot of math progress comes from some crazy harness, assuming labs use models similar to those available to general users: careful prompts, orchestration, best-of-k with a verifier, etc

Without that, agents don’t seem capable of making meaningful progress on difficult research problems

I wish someone would share a reproducible setup for solving a genuinely difficult math problem

[Georgy Khalin @gkhalin]

@0xfdf When I read AI processed meeting logs, I wonder how LLMs manage to do a decent job at coding given how little they seem to understand about the business

[Georgy Khalin @gkhalin]

@staysaasy I wonder if it’s a net positive strategy to feel that you are smarter than most people around you

The downside is that you might ignore genuinely good ideas from other people

[Georgy Khalin @gkhalin]

I wonder how open weight model labs are going to make money or what drives them

My bet is that at some point there will be unique paid products around problems like continual learning

Extra points if continual learning doesn’t require sharing private data with a third party, e.g. through local inference

[Georgy Khalin @gkhalin]

@ChShersh I don’t think people can become more productive all of a sudden

[Georgy Khalin @gkhalin]

@bschne Sometimes depth suffers because there are too many things you want to explore

[Georgy Khalin @gkhalin]

@TheGlobalMinima Clinging to old ways of doing things might not be the best use of curiosity right now

[Georgy Khalin @gkhalin]

Chinese labs' progress is the only reason I don’t care about optimizing $300 session costs right now