All you’re proving is that any specific domain will have a dozen mature solutions.
What separates one product from another is going to be constraints specific to the project you’re integrating this into. And none of that is defined in your prompts. So you’re effectively just asking an LLM to roll the dice and pick something at random and then acting surprised that randomness is results in different answers.
People think "AI" has all the facts to make objective decisions, but it seems to boil down to (SEO again?) what pages its search engine returned that day.
> If you're not a leader, this is the part that matters. Your visibility depends on which pages the engine reads that day. The way to make it stable is to be on more of those pages.
Ultimately a true facts knowledge graph independent of mentions is needed.
Looking at the questions they asked, they are quite broad. In my experience, part of the value of LLMs comes from its randomness, but you must understand it as a characteristic virtue of these tools, and must adapt accordingly.
Asking questions with tight scopes, clear constraints, and specific outcomes, always yields better results, in my opinion.
In general it seems like LLMs have a fundamental problemw with noise in their outputs such that randomness (what data they pull and how they aggregate it) has an outsized impact when asking it for opinions/recommendations/forecasts.
Newer models dont use temperature. It was really hard to make deterministic even with temp =0 due to numerical inconsistencies with whatever else is in your batch
I’m the author. Every day we ask ChatGPT, Gemini, Google AI Overviews and Google AI Mode the same 210 buying questions, like “best CRM for small business”, and record which brands they name and which pages they cite. This post compares the first two days.
84% of answers named a different set of brands the next day. On Gemini it was 95%.
The cause is easy to see in the data. Each answer is a summary of a few web pages, and the engine picks those pages fresh each time. Fewer than half of the pages cited on day one were cited again on day two. When the pages change, the brands change.
One example: asked for the “easiest CRM for a team that has never used one”, Google AI Mode read Zapier’s list and put Less Annoying CRM first. The next day it read a Slack blog post instead, and Less Annoying CRM wasn’t mentioned at all.
Who survives depends on how many pages you’re on. A brand ranked #1 was named again the next day 90% of the time. A brand ranked sixth or lower: 61%. Big brands are on most of the pages, so whichever ones the engine reads, they’re there. Smaller brands are on a few, so they come and go.
You are asserting that these changes are somehow tied to time, by the very construction of your experiment in that you ask the same questions sequentially day-by-day. However, that leaves you unable to be sure whether the changes are actually tied to the time because it is not being controlled for, as near as I can tell.
As near as I can tell from your "Method and Limitations" from the previous day's run, you are only asking a given AI a given question once. You need to ask the same AI the same question a number of times, preferably as close to simultaneously as possible, to see if there's as much variance within a single day as there is in a week.
As your experiment is constructed, it is not possible for you to collect the data adequate to prove the claim you are trying to make. As I think it is likely that the variance you are seeing is the randomness in the answer process itself and not tied to time at all; two people can probably ask the same question at the same time and get two different answers.
If nothing else, have an AI go back over your experiment, perhaps feed it this comment as well, and ask it what else may be flawed in your experiment. That doesn't mean you have to adopt every suggestion; every experiment can always gather more data and always have an expanded scope and always cost more money. However this is a more-or-less fatal problem with your experiment and there may be others, this is just all I'm considering. An AI can bang on the problem in a more dedicated fashion than I'm doing here.
I don't think "84% of answers changed" is a useful way to describe it. It looks like the content stayed pretty much the same, with sometimes one or two brands switching position or dropping off the list.
If you asked me what my top six favorite shows were, and then again 2 days later, I'm sure I would have one or two series that were new to the list or changed spots.
Edit: If this is something you're actually interested in measuring, there are several statistical tools meant for comparing the similarity of rank-ordered lists. I'd probably reach for a rank-biased overlap measure first.
All you’re proving is that any specific domain will have a dozen mature solutions.
What separates one product from another is going to be constraints specific to the project you’re integrating this into. And none of that is defined in your prompts. So you’re effectively just asking an LLM to roll the dice and pick something at random and then acting surprised that randomness is results in different answers.
People think "AI" has all the facts to make objective decisions, but it seems to boil down to (SEO again?) what pages its search engine returned that day.
> If you're not a leader, this is the part that matters. Your visibility depends on which pages the engine reads that day. The way to make it stable is to be on more of those pages.
Ultimately a true facts knowledge graph independent of mentions is needed.
I don't think this is inherently surprising.
Looking at the questions they asked, they are quite broad. In my experience, part of the value of LLMs comes from its randomness, but you must understand it as a characteristic virtue of these tools, and must adapt accordingly.
Asking questions with tight scopes, clear constraints, and specific outcomes, always yields better results, in my opinion.
In general it seems like LLMs have a fundamental problemw with noise in their outputs such that randomness (what data they pull and how they aggregate it) has an outsized impact when asking it for opinions/recommendations/forecasts.
??? Set the temperature to zero then ???
Newer models dont use temperature. It was really hard to make deterministic even with temp =0 due to numerical inconsistencies with whatever else is in your batch
I’m the author. Every day we ask ChatGPT, Gemini, Google AI Overviews and Google AI Mode the same 210 buying questions, like “best CRM for small business”, and record which brands they name and which pages they cite. This post compares the first two days.
84% of answers named a different set of brands the next day. On Gemini it was 95%.
The cause is easy to see in the data. Each answer is a summary of a few web pages, and the engine picks those pages fresh each time. Fewer than half of the pages cited on day one were cited again on day two. When the pages change, the brands change.
One example: asked for the “easiest CRM for a team that has never used one”, Google AI Mode read Zapier’s list and put Less Annoying CRM first. The next day it read a Slack blog post instead, and Less Annoying CRM wasn’t mentioned at all.
Who survives depends on how many pages you’re on. A brand ranked #1 was named again the next day 90% of the time. A brand ranked sixth or lower: 61%. Big brands are on most of the pages, so whichever ones the engine reads, they’re there. Smaller brands are on a few, so they come and go.
You are asserting that these changes are somehow tied to time, by the very construction of your experiment in that you ask the same questions sequentially day-by-day. However, that leaves you unable to be sure whether the changes are actually tied to the time because it is not being controlled for, as near as I can tell.
As near as I can tell from your "Method and Limitations" from the previous day's run, you are only asking a given AI a given question once. You need to ask the same AI the same question a number of times, preferably as close to simultaneously as possible, to see if there's as much variance within a single day as there is in a week.
As your experiment is constructed, it is not possible for you to collect the data adequate to prove the claim you are trying to make. As I think it is likely that the variance you are seeing is the randomness in the answer process itself and not tied to time at all; two people can probably ask the same question at the same time and get two different answers.
If nothing else, have an AI go back over your experiment, perhaps feed it this comment as well, and ask it what else may be flawed in your experiment. That doesn't mean you have to adopt every suggestion; every experiment can always gather more data and always have an expanded scope and always cost more money. However this is a more-or-less fatal problem with your experiment and there may be others, this is just all I'm considering. An AI can bang on the problem in a more dedicated fashion than I'm doing here.
have you tried doing the same with web search disabled (to see what is actually in the training, and see if answers vary due to temperature)
I don't think "84% of answers changed" is a useful way to describe it. It looks like the content stayed pretty much the same, with sometimes one or two brands switching position or dropping off the list.
If you asked me what my top six favorite shows were, and then again 2 days later, I'm sure I would have one or two series that were new to the list or changed spots.
Edit: If this is something you're actually interested in measuring, there are several statistical tools meant for comparing the similarity of rank-ordered lists. I'd probably reach for a rank-biased overlap measure first.