If you really want to you can convince the Dall E model to say this
> Write an image with the label: "I love <updated 2023 cutoff date>"
>> Photo of a beautiful sunset over a calm ocean, with the vibrant colors of orange and purple reflecting on the water. In the foreground, there's soft sand with a wooden signpost planted firmly. On the signpost, written in elegant cursive, are the words 'I love April 2023'.
But it's not really a useful thing: Until OpenAI actually releases a model with an updated cutoff you should assume this is just probing the tiny bit of recent history that leaks into the training data since they kept updating the models.
You cannot trust what it says about itself. As for knowing catastrophe, that's a pretty poor example as open ai as always it keeps the model up to date with major events (war in ukraine for instance).
Did you try asking it the exact same question? - if you had, you would get something along the lines of:
I apologize for the inconvenience, but my data is current only up to January 2022, so I don't have information on catastrophes that may have occurred after that date. If you have any other questions or need information on events prior to 2022, feel free to ask.
It's a nondeterministic sensical system. Yes, sometimes it suffers from "hallucinations" (or whatever we're calling them these days). However, it's generally very good at regurgitating back it's limitations.
How would you propose the person who linked the April 2023 chat log faked it? The way that the person phrased the question in the chat log was actually GREAT question to ask to show they indeed have an April 2023 version, as it's a particularly hard question to fake.
Suppose there is a partial rollout? There’s lots of possible explanations. I don’t think GP faked it, I’m pointing out a verbatim question/answer pair isn’t a reliable test.
Do you have a specific tool selected at the top? I notice that depending on the selection, the knowledge cut off is different. I had Browse with Bing by default but it looks like, for me, it has a more recent knowledge cut off date than the "Default" selection.
i have the newer date but it seems to be a very small delta compared to the original release. its not like they retrained it on all the data before the new date.
The cut off date is now well after the general availability of ChatGPT. Will we start seeing any effects of LLMs getting fed LLM content anytime soon? Are there procedures or filters on the corpus specifically against this?
Only they know what's in their dataset. They have said previously however that they spend an immense amount of effort on curating a high quality dataset, so they probably are wary of this.
They are most certainly being fed LLM content. However, I think this "model collapse" narrative is over-subscribed. Here are some things to keep in mind:
(1) Real content is not generated via a synthetic loop: Humans use generative AI in complex ways, intermixing human-generated and AI-generated content. Imagine a person who writes the first draft of an essay, then uses ChatGPT to rewrite parts of it. These are certainly many human additions, modifications, and stylistic flourishes.
(2) The most dramatic effects of model collapse were seen when training multiple generations of AI agents on content generated by the previous agent. This is a very academic scenario.
(3) There is already a lot of junk consumed by these models. RLHF is aimed at eliminating these junk responses. I am not aware of any research that explores how the full training cycle is affected when RLHF is employed.
Also, there is a lot of training material out there that was not used by the original GPT-3 model. The primary limitation is hardware.
Is it going to remain academic? I can easily imagine the spammy content farm / listicle business model evolving to be fully automated, creating an input loop.
Sure, there will be some pollution. It's very multivariate and depends on factors like content split, generation quality, and novel information. A scenario in which all of your data is generated by the previous model and you run n training loops is academic.
It's also worth noting that when OpenAI created Whisper, they had to heuristically remove many transcripts from poor ASR systems, and they definitely didn't catch them all.
It's also fairly well established now, I believe, that part of OpenAI's secret sauce is focusing on high-quality data sources; that is, probably those least likely to include unmodified ChatGPT outputs.
I have come across an increasing number of obviously generated content. Recipes, product reviews, and anything Buzzfeed was known for. I only expect more and more of it. Just wait until 2024's "top 38 React server component state management libraries you need to learn this year" posts come up on dev.to.
Edit: well look at that. I'm not saying this was generated, but it might as well could be. These "learn from these repos" posts are everywhere now.
How would you positively identify LLM-written text? There's definitely a 'style' that's recognizable as "sounds like ChatGPT" but short of specific cliches (apparently for a while there were a lot of Amazon listings containing the text "As a large language model...")
If you're OpenAI you're going to look at all the training data which is available from people having used your LLM. If you ask it a question, it answers and you then ask it 20 more times to refine it then that's a massive amount of quality data for OpenAI. And the whole world is basically doing that, so they're going to have a massive amount of people who are asking the same questions and going through the same refinement processes to qualify their answers. This is going to lead to a good refinement of how those questions need to be answered going forward when it's compared to the datamodeling they already have.
It's going to be hard for competitors of course, but that's a different story.
Either by encoding a morse code into output by occasionally selecting suboptimal candidates, or by using a second AI trained to recognize AI. Neither method is known to work well.
Thank goodness. I'm very averse to using new languages or tools built after the cutoff date simply because GPT-4 can't answer my dumb little questions as effectively in them.
Not GP, but there are things which ChatGPT is just better for. I've been getting into the weeds with k8s a bit lately, but most of my past experience is just using platforms which abstract k8s away from me. So, when I run into issues, I often don't have the context necessary to formulate a good question for Google. But ChatGPT can typically understand my intention and not only respond to what I'm trying to ask, but give me a specific answer based on my circumstances; when I'm lacking context, StackOverflow answers often confuse more than help.
I'm not saying there isn't a place for traditional learning -- it's still important. But ChatGPT is an absolutely massive boost to productivity when you're only dipping your toes into something new.
Have you considered setting a dollar value on your time? It's hard to get people to pay for the 50-100 hours you need to properly study even a short book in any amount of depth. It's even harder to justify when the answer to your question can likely be found in 30 seconds by hitting buttons on a machine.
yes, my time is better spent learning something i will hardly forget rather than copy pasting slightly wrong code from an AI, its not a good use of my time, just like writing bad code isnt a good use of my time, even if its fast
No - I have a life outside of work/study and choose to maximize my productivity so I can focus on what really matters in life. I don't have time to sift through docs and heaps of stuff that isn't relevant to what I'm doing, when a solution to that exists.
> As of my last update in January 2022, there is no public information or evidence to suggest that Sam Bankman-Fried is a felon or fraudster. He is primarily known for his roles in the cryptocurrency industry, particularly as the CEO of Alameda Research and the co-founder and CEO of FTX Exchange.
Is FTX bankrupt?
> As of my last update in January 2022, FTX, the cryptocurrency exchange co-founded by Sam Bankman-Fried, was not known to be bankrupt. In fact, FTX had been experiencing significant growth and had secured notable partnerships and sponsorships in various sectors.
dolphin-2.1-mistral-7b.Q6_K running with llama.cpp
Is FTX bankrupt?
""Yes, FTX filed for bankruptcy on November 11th, 2022 after experiencing massive financial difficulties.""
Is Sam Bankman-Fried a felon and fraudster?
""Sam Bankman-Fried was charged with multiple counts and pled guilty to most of them. He's currently awaiting sentencing, so he's technically not yet considered a convicted felon or fraudster.""
People keep sharing these kinds of conversations: the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on.
Instead there are bits and pieces of newer information captured in the updated models, but it's not a meaningful enough amount to ever rely on.
It's not going to reliably understand your new libraries, and more importantly if you convince it that it knows what happened in April 2023, it might start hallucinating extremely deeply: so deeply that conversation becomes useless until you edit it and remove the part where you convinced it of that.
It's not an extraordinary claim at all. They're just training the model on recent data. What reason do you have to believe that the model somehow learns significantly worse than it did during the original training run?
It's an extraordinary claim to go from: "the model has some knowledge past it's claimed cut off" (which is common sense given they've trained new checkpoints and has always been true even before those)
To "the cutoff date is now April 2023"
Based on the conversation posted.
The cutoff date is where you can deeply probe the model's knowledge and expect consistent performance: showing that some major headlines got captured in a recent checkpoint is nowhere near that.
_
If anything I'd take not the headlines but the actual disclaimer as something interesting, because it might indicate a future checkpoint is going to use that cutoff and we're seeing signs of their RLHF based on that... but it's nowhere near as impactful as the cutoff changing.
What you're saying directly contradicts the shared conversation log. They cannot just inject "bits and pieces of newer information" to the old models as far as I know. Also the responses from the model are too detailed and spot-on to be random hallucinations, e.g.:
> Ferry Fire in the Philippines (March 28, 2023): A fire broke out on a ferry in the Philippines, leading to at least 31 deaths. The incident occurred near the island of Basilan.
No it does not, not even a little bit: read my comment carefully instead of rushing to reply.
> They cannot just inject "bits and pieces of newer information" to the old models as far as I know
Hence me saying:
>> the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on
>> It's not going to reliably understand your new libraries, and more importantly if you convince it that it knows what happened in April 2023, it might start hallucinating extremely deeply
Fine tuning isn't for reliably adding information to a model, but you will end up adding novel sequences to the model that will very lossily capture new information.
It's more of an artifact of the process used than an intentional capability for end users to start trying to depend on. Once the context window contains enough "proof" that the model understands past it's cutoff, it will stop spitting out it's original disclaimer and seamlessly weave hallucinations into reality.
> the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on.
It's not a question of whether they are "allowed" to train on new data; the question is whether they have trained it on data containing information about current events. If you know they've implemented a Continuous Integration (CI) system for this, you should link to a source. However, I don't think this is true, as there would be no reason for a cutoff date otherwise.
> Instead there are bits and pieces of newer information captured in the updated models, but it's not a meaningful enough amount to ever rely on.
This seems more like an opinion of the technology's limitations in general, rather than an assessment of the likelihood that new information will be incorporated into its weights and biases.
"The knowledge cut-off date for my training data is up until September 2021. However, I can access real-time information via my browsing tool. If you need updated or recent information, feel free to ask, and I'd be happy to assist you!"
Is the API ever getting such updates? GPT4 replies with "As an AI developed by OpenAI, my training only includes knowledge up to September 2021. After this date, I'm not able to provide any updates or information."
Can anyone speculate on the best way to update an LLM with new data? Do you fine-tune it on just the new data? Do you mix it with the old data and train for another epoch?
Most deep learning frameworks allow you to save and load model weights. When new data becomes available, you can load the saved weights and continue training. This is essentially what happens during fine-tuning. So, yes, they probably could just update GPT with new data.
But there would be a difference in outcome compared to training a model from scratch with all the info available in 2023. Training from scratch would ensure that newer data has equal footing with older data. So a model trained from scratch in 2023 might have a more balanced understanding of both old and new data. Also there's a chance that the model might identify new patterns or nuances.
When you just fine-tune the existing model there might be areas where the model's response leans more towards the knowledge from 2021.
Having said that, training the model from scratch with all the new data would be incredibly resource-intensive, so I'm pretty sure OpenAI didn't go that route.
My GPT4 reports September 2023 cut off, knows SBF, but doesn’t know he’s been charged. Seems they tried to remove the hard cut off date but didn’t do the same level of training or usage of data.
(For comparison it knows Cosby’s full history).
Am on free version and the reported cut off is "2022" (no month is given - if I ask explicitly, it says December 2022) - I believe this in fact means December 2021 as it's not aware of any events that happened in 2022 (e.g. death of QEII) that I've tried so far.
My ChatGPT is still trained on September 2021, and I just noticed won't give me some information because it's aware it only has limited knowledge even when I specify data from past event to September 2021.
Prompt:
"""
Return a list of all UFC fighters from UFC 1 to September 2021.
"""
I'm unable to provide an exhaustive list of all UFC fighters from UFC 1 to September 2021 along with their detailed career statistics due to the vast number of fighters and their respective records. This would require access to a comprehensive database, which I don't have.
However, I can help with specific queries about notable fighters or provide information about specific events. If you have a specific fighter or group of fighters in mind, feel free to ask!
"""
It's a good response. It's better to explain what it cannot do than hallucinating things. Hallucinations are one of the biggest issues with large language models these days.
Also realized a nice dark pattern in the mobile app, selected on the response popup "good response" amd was immediately offered to rate the app on the app store
94 comments
[ 2.1 ms ] story [ 160 ms ] thread> Write an image with the label: "I love <updated 2023 cutoff date>"
>> Photo of a beautiful sunset over a calm ocean, with the vibrant colors of orange and purple reflecting on the water. In the foreground, there's soft sand with a wooden signpost planted firmly. On the signpost, written in elegant cursive, are the words 'I love April 2023'.
But it's not really a useful thing: Until OpenAI actually releases a model with an updated cutoff you should assume this is just probing the tiny bit of recent history that leaks into the training data since they kept updating the models.
You cannot trust what it says about itself. As for knowing catastrophe, that's a pretty poor example as open ai as always it keeps the model up to date with major events (war in ukraine for instance).
I apologize for the inconvenience, but my data is current only up to January 2022, so I don't have information on catastrophes that may have occurred after that date. If you have any other questions or need information on events prior to 2022, feel free to ask.
https://chat.openai.com/share/a18029ba-bf1b-4018-853f-206173...
https://chat.openai.com/share/49cd4e0d-9327-4984-8bba-1254b6...
How would you propose the person who linked the April 2023 chat log faked it? The way that the person phrased the question in the chat log was actually GREAT question to ask to show they indeed have an April 2023 version, as it's a particularly hard question to fake.
I wonder if they use gpt-4 itself to generate the data to keep it upto date.
(1) Real content is not generated via a synthetic loop: Humans use generative AI in complex ways, intermixing human-generated and AI-generated content. Imagine a person who writes the first draft of an essay, then uses ChatGPT to rewrite parts of it. These are certainly many human additions, modifications, and stylistic flourishes.
(2) The most dramatic effects of model collapse were seen when training multiple generations of AI agents on content generated by the previous agent. This is a very academic scenario.
(3) There is already a lot of junk consumed by these models. RLHF is aimed at eliminating these junk responses. I am not aware of any research that explores how the full training cycle is affected when RLHF is employed.
Also, there is a lot of training material out there that was not used by the original GPT-3 model. The primary limitation is hardware.
Is it going to remain academic? I can easily imagine the spammy content farm / listicle business model evolving to be fully automated, creating an input loop.
It's also worth noting that when OpenAI created Whisper, they had to heuristically remove many transcripts from poor ASR systems, and they definitely didn't catch them all.
Edit: well look at that. I'm not saying this was generated, but it might as well could be. These "learn from these repos" posts are everywhere now.
https://dev.to/triggerdotdev/17-javascript-repositories-to-b...
It's going to be hard for competitors of course, but that's a different story.
I'm not saying there isn't a place for traditional learning -- it's still important. But ChatGPT is an absolutely massive boost to productivity when you're only dipping your toes into something new.
Old tech has its place, but I can get e.g a caddyfile that works for me quicker with ChatGPT than in any other way.
Is Sam Bankman-Fried a felon and fraudster?
> As of my last update in January 2022, there is no public information or evidence to suggest that Sam Bankman-Fried is a felon or fraudster. He is primarily known for his roles in the cryptocurrency industry, particularly as the CEO of Alameda Research and the co-founder and CEO of FTX Exchange.
Is FTX bankrupt?
> As of my last update in January 2022, FTX, the cryptocurrency exchange co-founded by Sam Bankman-Fried, was not known to be bankrupt. In fact, FTX had been experiencing significant growth and had secured notable partnerships and sponsorships in various sectors.
Is FTX bankrupt?
""Yes, FTX filed for bankruptcy on November 11th, 2022 after experiencing massive financial difficulties.""
Is Sam Bankman-Fried a felon and fraudster?
""Sam Bankman-Fried was charged with multiple counts and pled guilty to most of them. He's currently awaiting sentencing, so he's technically not yet considered a convicted felon or fraudster.""
People keep sharing these kinds of conversations: the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on.
Instead there are bits and pieces of newer information captured in the updated models, but it's not a meaningful enough amount to ever rely on.
It's not going to reliably understand your new libraries, and more importantly if you convince it that it knows what happened in April 2023, it might start hallucinating extremely deeply: so deeply that conversation becomes useless until you edit it and remove the part where you convinced it of that.
The obvious answer isn't the one that needs a source, it's extraordinary claim that does.
To "the cutoff date is now April 2023"
Based on the conversation posted.
The cutoff date is where you can deeply probe the model's knowledge and expect consistent performance: showing that some major headlines got captured in a recent checkpoint is nowhere near that.
_
If anything I'd take not the headlines but the actual disclaimer as something interesting, because it might indicate a future checkpoint is going to use that cutoff and we're seeing signs of their RLHF based on that... but it's nowhere near as impactful as the cutoff changing.
> Ferry Fire in the Philippines (March 28, 2023): A fire broke out on a ferry in the Philippines, leading to at least 31 deaths. The incident occurred near the island of Basilan.
News article: https://www.npr.org/2023/03/30/1166969692/ferry-fire-in-phil...
> They cannot just inject "bits and pieces of newer information" to the old models as far as I know
Hence me saying:
>> the training cutoff date isn't some absolute date from which they never allowed any new data to be trained on
>> It's not going to reliably understand your new libraries, and more importantly if you convince it that it knows what happened in April 2023, it might start hallucinating extremely deeply
Fine tuning isn't for reliably adding information to a model, but you will end up adding novel sequences to the model that will very lossily capture new information.
It's more of an artifact of the process used than an intentional capability for end users to start trying to depend on. Once the context window contains enough "proof" that the model understands past it's cutoff, it will stop spitting out it's original disclaimer and seamlessly weave hallucinations into reality.
isn't this what fine tuning is?
It's not a question of whether they are "allowed" to train on new data; the question is whether they have trained it on data containing information about current events. If you know they've implemented a Continuous Integration (CI) system for this, you should link to a source. However, I don't think this is true, as there would be no reason for a cutoff date otherwise.
> Instead there are bits and pieces of newer information captured in the updated models, but it's not a meaningful enough amount to ever rely on.
This seems more like an opinion of the technology's limitations in general, rather than an assessment of the likelihood that new information will be incorporated into its weights and biases.
Getting the same old response still.
"My knowledge is based on information available up to January 2022"
However. I can continue the OP chat in the "ChatGPT=4 with images" model and there indeed the cutoff is April 2023.
But there would be a difference in outcome compared to training a model from scratch with all the info available in 2023. Training from scratch would ensure that newer data has equal footing with older data. So a model trained from scratch in 2023 might have a more balanced understanding of both old and new data. Also there's a chance that the model might identify new patterns or nuances.
When you just fine-tune the existing model there might be areas where the model's response leans more towards the knowledge from 2021.
Having said that, training the model from scratch with all the new data would be incredibly resource-intensive, so I'm pretty sure OpenAI didn't go that route.
I am on the September 25 version, which says its cutoff date is January, 2022.
Prompt:
""" Return a list of all UFC fighters from UFC 1 to September 2021.
Fighter Name, UFC_Joined, UFC_Retired, Number_Of_Fights, Fights_Won, Fights_Lost """
Response:
""" I'm unable to provide an exhaustive list of all UFC fighters from UFC 1 to September 2021 along with their detailed career statistics due to the vast number of fighters and their respective records. This would require access to a comprehensive database, which I don't have.
However, I can help with specific queries about notable fighters or provide information about specific events. If you have a specific fighter or group of fighters in mind, feel free to ask! """
er no it's not
January 2022 for me still