Georgy Khalin (@gkhalin) header
Georgy Khalin (@gkhalin) profile picture

Head of Quantitative Development at https://teapawt.net/t.co/HuN94SnxTi Scaling research into impact

Amsterdam, The Netherlands
Joined March 2012
Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)Photo by Georgy Khalin (@gkhalin)
When I started my career in trading, my compensation grew exponentially for a couple of years. I felt like I had made it and that it would always stay at least this high Then one day our strategies crashed, and it took years to recover. It was an incredibly humbling experience When my stress and anxiety hit rock bottom, I reached out to a coach. She started asking hard questions: Why do you need this much money? What’s the worst-case scenario if you don’t make it? A few months later, after doing the math and thinking it through, I realized three things: 1. I don’t need that much money to live a good life 2. I’m skilled enough to find another job if I ever need to 3. Skills compound. That’s what I should optimize for. Money is just feedback that I’m on the right track Since then, I’ve kept my lifestyle relatively cheap, and my money anxiety has largely disappeared Ironically, worrying less about money made it easier to focus on what actually mattered: getting better at my craft and building things that work Things have been going great again lately
Any professional advice that asks you not to think of money while pursuing a career is horrible professional advice. Let me explain: - Building a career should be a monetarily rewarding experience. If you are doing everything right, it is vital that you earn appropriately. - In a capitalistic world, where businesses pursue profits, it should be normalized for professionals to pursue money. Of course, I am not implying that all unjustified salary demands should be fulfilled or even entertained. In a free market, it all boils down to one's skill and contribution to the organization's success/ profit. Ultimately, we are living in a permissionless world. Build skills and play the game of wealth. If you are great at what you do, you will win anyway.
2
20
2,224
Amazing what you can do with $30 of inference and an inspiration board I wanted to see what my apartment could look like after renovation I’ve been collecting a home design inspiration board for months, just because I enjoy doing that I used xai grok 4.7 and fed it my apartment floor plan via grok bot. First, we reconstructed a 3d model. Then I gave it the 3d model and a link to my inspiration board on Savee After 5 iterations using voice input, I’m happy with the result Details matter a lot in interior design, and I’m sure a professional would still have plenty to figure out. But I think I can now communicate exactly what I want with much less ambiguity
1
3
88
AI adoption questions I use when interviewing candidates: 1. What's the median time your agents work autonomously for? 2. What are agents good and not good at right now? 3. What know-how tricks do you use to get the most out of agents? 4. What's the most ambitious thing you've achieved with agents? This also reveals reflection skills and how big people think If someone says that most of the time their agents work autonomously for 3 minutes, well, that's totally unacceptable
1
6
181
Personal agentic digests have the highest signal-to-noise ratio compared to alternatives I’ve tried I have three different pipelines built on top of @bot, with customized ranking and information compression based on my preferences, via the xAI API One of the pipelines sources companies raising money, including investment size, their top-1 scenario, moat, beliefs, and sometimes revenue growth. An agent sends me the report every morning Not saying I’m bullish on personal agents yet, but they definitely look promising
3
2
140
Some thoughts on the human interface to AI research in the near future: 1. In an agentic world, the human role is to set goals and manage the search space by prioritizing directions and adding constraints. That's why prompt engineering matters 2. Research questions should be reframed to be verifiable. Otherwise, agents can't reliably tell whether they're making progress 3. High-quality validation is extremely important. Agents will happily overfit to your validation setup if you let them 4. You should keep a record of all experiments, outcomes and conclusions. This is how both you and the agent learn over time and avoid repeating mistakes 5. The agentic loop should be mostly autonomous. It should keep working even when the human goes on vacation, with appropriate stopping criteria and compute budgets 6. Humans should focus on understanding what the agent is doing wrong and improving the research process rather than grinding towards the outcome themselves 7. There should be a separate agentic role focused on optimizing experimentation speed. If you can learn the outcome of an experiment faster, that's a huge lever
1
1
122
Friction used to be a great filter on low-belief ideas. When friction goes to zero, you stop thinking that hard
This quoted post is unavailable.
2
105
One reason progress toward ASI could slow down is that the better models get, the harder it becomes to build evals they can’t solve yet And as problems get harder, feedback loops tend to get longer, slowing down iteration
5
93
Having access to private information is going to be a big edge post-AGI Right now there is too much noise which is hard to make sense of. GPT-9 would be able to process all the information you have and advise you on the most optimal next step toward your goals Talk to people, expand your network, get access to places few people can enter
4
103
The question is what makes you exceptional if intelligence is out of the equation If you understand that, you know what to double down on post-AGI
3
100
Agents get more ambitious when you say “I’m sure X is possible” Over the last few weeks, I’ve had a few ridiculous requests like “speed it up 100x” or “this metric must cross Y” The agents narrowed their search space toward these outcomes and eventually managed to find solutions
2
106
One piece of advice for eval builders would be to write hints to distinguish between impossible evals and those that are just too hard for the current generation of models E.g. if astra fails an eval but passes with hints, keep it until a better model is released that can solve the problem without hints
2
145
You should have evals testing specific knowledge about your business in Q&A form These are much easier to build and maintain than problem-solving evals This will let you measure whether models understand your business well enough to help optimize it
3
104
If you don't care enough about other people's problems, don't be surprised when nobody helps you next time you're stuck It's an interesting trade-off: how much time do you spend helping others to build enough trust that they'll help you when you need it? I have a skill issue with saying "no", which turns out to be quite useful for building trust
4
125
To be naturally productive at work, you need to understand what you want out of your career and ask for it Sounds simple, but most people either don't know what they want or their preferences change as they try different things Keep reflecting on what feels interesting, what gives you energy, and what drains you Over time, you'll figure it out
2
118
You don't have to review every line of code as long as 1. Abstractions are clear and interfaces look good 2. You are happy with the quality and understand which metrics changed and why 3. Tail risks are taken care of by extra tests/checks/safeguards 4. You monitor performance
1
2
129
The Lindy effect seems to predict pretty well which jobs will remain relevant post-AGI Think of the oldest profession you know. I’m pretty sure it’s not going anywhere
2
85
Demand for therapists is going to blow up in the next couple of years So many people will face an existential crisis when AI starts threatening the identities they've built around their work
3
96
You’re good at delegation as a manager if you can spend at least 25% of your time exploring what’s next rather than constantly patching holes
3
116
Georgy Khalin retweeted
Replying to @staysaasy
Atomic Habits baited way too many people into learning how to do more of what they don’t necessarily want to do We need a bestseller on how to figure out what you actually want out of life
1
1
9
255
You're better prepared for AGI than you think you are It's not deep knowledge of C++, the skill to debug complex issues, or the ability to quickly solve tasks given by your manager that made you successful It's often about meta-skills like the ability to focus for long periods of time, high tolerance for failure, taking ownership, and getting things done no matter what it takes LLMs extend intelligence, but they don't own the outcomes or take the risks If you understand deeply enough what actually made you successful, you're in a much better position to adapt even if intelligence gets solved. You'll figure out new ways to provide value and have a fulfilling career
6
113
Sales and marketing will be the new software engineering after AGI When the number of products explodes, distribution becomes the scarce resource You’ll have to understand human nature & behavior really well to succeed
5
106