I'm confident that Muse can do lots of agentic tasks succesfully for normies on outdated models. Like buying sneakers, booking appointments, dealing with government forms.
It would be more interesting to compare trading agents with index tracking ETFs. The better version of an ETF could maybe be a model where you zoom in on the companies and add/remove to your portfolio on the company related news, but keeping a broader portfolio.
Maybe agentic trading still performs worse than ETFs. But alternatively, if it were meaningfully better then it would be okay to opensource, similarly how ETFs are publishing their portfolios.
When I started working no the trading agent I mentioned above I wanted to see if it can be just a better investor over the long run. The intention was not to do high-frequency trading. As you can see most of the days it is not taking any actions. The losses where down to mistakenly setting the stop losses too close to the top. If it wasn't so careful it might have made some money tbf.
My gut feeling is that AI agents will be able to manage a long-term portfolio much better than a human. Though it is just a gut feeling.
Lmao, you llm people have some crazy delusions. You realize markets are zero sum, and if you're using a public model that everyone else also has access too, you llm psychos will destory eachothers "agentic" edge (not that there ever was one). Not to mention all the other obvious flaws with llms, lime having an effective memory of ~200k words and no ability to judge whats actually going on in the real world.
Markets are only zero-sum in any given trade. Allocating capital to assets with higher growth rates (on the marginal dollar) creates value in the long run.
So, very simply, if AI can actually do better at picking a better long-term winner then it will increase growth.
Yeah and you probably have to. An ETF easily has >1000 different stocks and even being weighted. So it has a completely different risk appetite by being so averaged.
But ETFs do have to follow particular rules defined by their product description. So it is still interesting to benchmark against.
You really have no idea what you are talking about.
I have been running an intermittent experiment with a multi agent "investment firm" for over a year now across model releases.
They certainly can beat indexes, BUT.. the model families have some biases that you have to design around. The stop loss that bit the parent is certainly one. The models like to create rules. Often rules, one of those is making all kinds of exit conditions.
Another big one from my experience is the bias to inaction in a scenario with risk. This means a model without structure around it will bias to keeping too much cash.
This would be nice if it was supported by some hard data. But even if it was, one year is too short of a timeline to make any kind of reasonable conclusions about its efficacy.
Buying stock based on coin flips can beat indexes short term too, that does not mean it is a better strategy or that it works over the long term.
There must be some room for some anti-llm agent that can profit from specific behaviors of these models when deployed against actual markets.
The idea that somebody here came up with idea that all professional algo traders didn't explore to the last penny a year if not more ahead of others is funny... but its not my money adding liquidity to the markets.
Come to me when you have 500 trades and can beat Vangaurd's-VOO over a multi year time frame. Ill bet my entire networth and all future earnings for the rest of my life that your bot doesnt beat it. Your llm induced Dunning Kruger is going to get you in trouble one of these days I promise.
Look, it isn't fool proof and it is dangerous. With the current models you need to understand both markets and model biases and dynamics.
However with that said they are a huge multiplier and can tirelessly analyze the market for you.
They certainly can be used to beat sp 500 quite easily, but again that requires some understanding of risk on your part because the models will do what you ask them. If you go all in on options or something without clear risk management you will lose your ass.
It makes no sense to train frontier models from scratch anymore. The best frontier models are only a half year ahead of Chinese open models. In this regard Anthropic and OpenAI are also in a bad spot when they waste so much compute on training models.
An important factor is that fine tuning existing open models is incredible cheap. You can easily change any cultural biases if you want a model to be 'sovereign'. And Mistral could combine that with their custom data sets for their enterprise customer needs. Mistral still trains their own models, but they also seem to offer fine tuning existing models.
With model weights being commoditized, another differentiator could be deploying efficient inference chips, especially if you combine it with a developer ecosystem for vendor lock-in. That is why it is interesting that both Samsung and ASML are investors, since they are companies that could make a difference in this area.
It only makes sense to train a frontier model if you are trying a different architecture to one that is available from an existing frontier model. This is because the different model architecture will learn the weights differently.
It may make sense to train a frontier model on an existing architecture if the base model is not available and the instruction trained version doesn't fit with what you want. There are techniques like ablation, but those could have other effects on the model, and there can still be lingering effects of the instruction training in the model that surface less frequently (e.g. on an input not covered by the ablation training).
Otherwise, fine tuning is definitely the way to go. However, you need to be careful not to over-tune the model such that it is only tuned to the data you are training it on.
On a higher level it might make sense to build the expertise that comes with base training.
I dont know enough about the process to estimate these gains, but china has been doing it in manufacturing for decades.
All the money in the world is useless when no one knows how to do the thing
So many moving parts that can break. I am already sad my robot vacuum cleaner broke (it's the moving lidar that broke). I have to send the entire thing including dock back to the manufacturer (very expensive).
My preference would be ifixit level guides where you can easily replace parts yourself. Then it would be perfect if the same servo is used everywhere so I can have 3 or 4 stored somewhere.
In general, a cost optimized robotic platform for a consumer market, has inverse specifications compared with reliable industrial robotics.
Industrial robotics like all machinery can be dangerous, and could still get your firm slapped with an export restriction. A lose-lose product class from a business perspective. =3
Alternatively, AI teaches you lessons about technical debt, code maintainability, and architecture faster too. Traditionally it took at least one or two years to really bump into those problems. Now any student can get into these problems within one or two weeks.
My analysis is that Stripe wants to own the costs of running a startup. They would get a complete picture of both sides of the business: what money comes in and where it goes.
There are also a couple of advantages. They can take money directly from revenue before it leaves Stripes and without any processing costs. They can also invest into startups through credits and financing. And finally, their exposure to bankruptcy risk can drop as well.
reply