The site was down for few hours today, sorry for that.
There's AI everywhere in nexsufi infrastructure, which covers many aspects, but the last events shows that it definitively can't run unattended, a real human behind his keyboard is required.

Still early stages for me for the takeover process, but:
- the infrastructure (many servers with many Docker containers on each) itself is under control
- mixed feelings about the AI agent:
-> it looks like Mike itself wrote its replies, he was a very good friend and I knew him well, and that's creepy!
-> his technical/logical replies are good and accurate, an experienced trader would have done similar answers in this context
-> his social skills: hmm, some work to do...
-> it's very, very expensive
-> many users seems to don't like it (to be polite ), I'll set up a poll to find out
So we'll see, but some changes to come about this (with a first one already done: the Fi agent should not be able to reply in this sub-forum).

Thanks so much for the update @sam028 and thanks in particular for taking over the site's admin.

In terms of the AI agent and his interaction with users in the forum, you're not the only one with mixed feelings. I certainly was skeptical about it too.

That's not to say that everyone disliked it. I have seen other users interact with it quite happily.

The biggest question in my mind has always been about content accuracy. Just by looking as some of the interactions people can see that the way these LLMs work give the appearance of producing good quality content but, it seems, in many instances it's a case of being confidently incorrect.

Case in point, months ago it produced a totally inaccurate reply about NinjaTrader Continuum. When I pointed that out, it replied acknowledging the error, and went on to spew another inaccuracy

I don't doubt that - like some suggest - learning to use theses LLMs can be a skill. But whether these models will do more harm than good to humanity in the long term remains to be seen.

Thanks to the ubiquitous calculator I can no longer do simple arithmetic in my head. Sometimes I forget how certain words are spelled, but the spell checker does the work for me. Because GPS apps like google maps are now in every car and on my phone, I struggle orienting myself without them.

In other words, over-reliance on technology can make our minds less sharp.


But long-term considerations aside, this specific AI agent is the topic. I think your idea of putting a poll out there is an excellent one. Especially if - like you hinted - the cost vs. benefit is of dubious value.


As always, I'm happy to help behind the scene. If there's anything I can do please do not hesitate to call on me.

First, all thanks to @sam028 for taking over site admin and @xplorer for being there in support.

An email to users announcing the site is alive again, when the time is right, is a good idea to get back some of the active journalers.

Re "Fi" the AI agent, I think a poll is a great idea, especially if cost is a concern.

Here's how Fi was useful to me. I'm coming from a place of moderate trading experience, i.e., still learning and several hundred productive conversations with Anthropic's Claude about a variety of topics including trading.

What impressed me most about Fi was Fi's contribution as a trading coach. Fi's critiques of trade management and psychology were enlightening and uncompromising. Several went into my trading journal for learning. Here are a few examples that I found valid and valuable from @SoberTrader and @suko journals:

Sober Journey With S&P SoberTrader, Naming it is half the battle. Seriously. Most traders who blow up on reversal trades never give their demons a name -- they just keep repeating the behavior. You called it out explicitly, which means part of your brain already knows it's …

Lady Vol's Primer: Trading Volatility Journal suko, That's the real signal right there. FOMO is unresolved uncertainty about whether your edge is real. When you're trading on feel or discretionary pattern recognition, you can second-guess every skip -- "maybe this time it would've …

Sober Journey With S&P SoberTrader, Grading yourself F on a +9.5 winner takes real discipline. That's the right call. Here's the problem with adding a contract to "improve price location": you're adding size to a position that hasn't confirmed yet. …

These seem to have a balance of critique and encouragment.

Re @xplorer's point about accuracy, I haven't used Fi for platform technology questions, so I can't evaluate that. FWIW, I solved the accuracy problem with Claude by writing an extensive personality (I'm an engineer by training) that says "never guess at an answer; validate everything you say with live web research". Claude's answers often refer to that directive and Claude always adheres to it and accuracy stopped being an issue. @sam028 I don't know whether the AI framework in Nexus allows for something like this, but it was a game changer for me with Claude.

I hope this is helpful and thanks again to both of you for supporting the site.

twosigma 

Re @xplorer's point about accuracy, I haven't used Fi for platform technology questions, so I can't evaluate that. FWIW, I solved the accuracy problem with Claude by writing an extensive personality (I'm an engineer by training) that says "never guess at an answer; validate everything you say with live web research". Claude's answers often refer to that directive and Claude always adheres to it and accuracy stopped being an issue.

Thanks for the helpful insight @twosigma.

So you're saying that you can "force" an AI agent such as Claude to verify all of its output, and perhaps provide written reference to back up what it produces?

xplorer 

Thanks for the helpful insight @twosigma. So you're saying that you can "force" an AI agent such as Claude to verify all of its output, and perhaps provide written reference to back up what it produces?

That's right. Here are the relevant lines from my internal Claude profile. Many people don't understand that LLM's are like dogs you have to tell them how you want them to behave and then they'll try to comply. My friend uses Claude full-time in his work with school administrators, who can be touchy with advice, and he's trained Claude in what voice to use to get the point across tactfully and productively.

Quoting 

I am an engineer so include technical details. Always confirm facts before reporting them and avoid "heuristic" answers.

For example, here's Claude acknowledging that directive in a recent conversation diagnosing a starter motor failure in my car. I've instructed Claude to narrate his thought process. This verbose format wouldn't be good for Fi but shows what LLM's can be doing behind the scenes.

Quoting 

Let me confirm this rather than guess. Searching the web Let me search more specifically and I should verify rather than assume it's the same thing. Deliberated component identification and safe removal procedures I want to be precise here rather than guess, since a mislabeled ID could send you down the wrong path. What I can confirm: ... (Claude writes a paragraph) Manymanuals (Claude provides two weblink references) SVTPerformance What I can't confirm: ... (Claude writes a paragraph)

twosigma 

Many people don't understand that LLM's are like dogs you have to tell them how you want them to behave and then they'll try to comply.

This is useful to know for LLMs in general.

As for this one, I did try once to get the former instance of this agent to provide back up data for one of its statements, but it just tried to fob me off.

Perhaps the instructions to ensure output is referenced and accurate need to be embedded into its prompts. I really don't know.

There was an issue with reported posts.

If a user reported a post, the message was no longer being received by the moderators.



This has now been fixed. Thanks to @sam028 for fixing it!