ChatGPT cut off date now April 2023
chat.openai.com
chat.openai.com
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.
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
"My knowledge is based on information available up to January 2022"
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'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.
> 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.
The response horizon is limited only by your patience with trying prompts before you give up.
Tips for skeptics
1. Don’t give it a url. But “github tmux/tmux” will usually work. There is a layer that finds tools via common whitelisted services; arbitrary urls make it morally panic ;)
2. State your intentions. It feels dumb; but it’s trained to be “ethical” based on a multitude of examples, a lot of which are tailored to detecting fraud and abusive sentiments.
3. It’s continuing a conversation. Telling it things are going well makes it continue a good conversation. Pessimism makes it struggle. It’s mimicking us, after all.
Getting the same old response still.
Response: Yes, my training data includes information up until September 2021, and I have been updated with additional knowledge that extends up until April 2023. However, I cannot browse the internet in real-time or access new data after this cutoff date. My responses are generated based on the information that was available up to that point in time. If there have been developments or new information released after April 2023, I would not be aware of them.
January 2022 for me still
I wonder if they use gpt-4 itself to generate the data to keep it upto date.
However. I can continue the OP chat in the "ChatGPT=4 with images" model and there indeed the cutoff is April 2023.
I am on the September 25 version, which says its cutoff date is January, 2022.
When I asked the question tonight I get January 2022.
(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.
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.""
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.