The ongoing model anchoring/grounding issue likely affects all GPT-4 checkpoints/variants, but is most prominent with the latest "gpt-4-turbo-2024-04-09" variant due to its most recent cutoff date, might imply deeper issues with the current model architecture, or at least how it's been updated:
Basically, without a separate declaration of the model variant in use in system message, even the latest gpt-4-turbo-2024-09 variant over the API might hallucinate being GPT-3 and its cutoff date being in 2021.
A test code snippet is included in the GitHub issue to A/B test the problem yourself with a reference question.
I think there's a bigger underlying problem with the current GPT-4 model(s) atm:
Go to the API Playground and ask the model what is its current cutoff date. For example, in its chat, if you're not instructing it with anything else, it will tell you that its cutoff date is in 2021. Even if you explicitly tell the model via system prompt: "you are gpt-4-turbo-2024-04-09", in some cases it still thinks its in April 2023.
The fact that the model (variants of GPT-4 including gpt-4-turbo-2024-04-09) hallucinates its cutoff date being in 2021 unless specifically instructed with its model type is a major factor in this equation.
A) Make sure "gpt-4-turbo-2024-04-09" is indeed selected. Don't tell it anything specific via the system prompt and in a worst case scenario, it'll think it's in 2021 as to its cutoff date. It also can't answer to questions about more current events.
* Reload the web page between prompts! *
B) Tell it via the system prompt: "You are gpt-4-turbo-2024-04-09" => you'll get answers to recent events. Ask anything about what's been going on in the world i.e. after April 2023 to verify against A.
I've tried this multiple times now, and have always gotten the same results. IMHO this implies a deeper issue in the model where the priming goes way off if the model number isn't mentioned in its system message. This might explain the bad initial benchmarks as well.
The problem seems pretty bad at the moment. Basically, if you omit the priming message ("You are gpt-4-turbo-2024-04-09"), it will in worst cases revert to hallucinating 2021 cutoff dates and doesn't get grounded into what should be its most current cutoff date.
If you do work at OpenAI, I suggest you look into it. :-)
Asking this here too: why isn't there an automated A/B or diff match for the tarball contents to match the repo, auto-flag with a warning if that happens? Am I missing something here?
The tarballs mismatching from the git tree is a feature, not a bug. Projects that use submodules may want to include these and projects using autoconf may want to generate and include the configure script.
> The release tarballs upstream publishes don't have the same code that GitHub has. This is common in C projects so that downstream consumers don't need to remember how to run autotools and autoconf. The version of build-to-host.m4 in the release tarballs differs wildly from the upstream on GitHub.
Multiple suggestions on that thread on how that's a legacy practice that might be outdated, especially in the current climate of cyber threats.
> Those days are pretty much behind us. Sure, you can compile code and tweak software configurations if you want to--but most of the time, users don't want to. Organizations generally don't want to, they want to rely on certified products that they can vet for their environment and get support for. This is why enterprise open source exists. Users and organizations count on vendors to turn upstreams into coherent downstream products that meet their needs.
> In turn, vendors like Red Hat learn from customer requests and feedback about what features they need and want. That, then, benefits the upstream project in the form of new features and bugfixes, etc., and ultimately finds its way into products and the cycle continues.
"and when the upstream is tainted, everyone drinks poisoned water downstream, simple as that!"
Been saying this the whole day now, GitHub really needs an automated diff / A/B check-up on tarballs against the actual repo, flag everything with at least a warning (+[insert additional scrutiny steps here]) when the tarball isn't matching the repo.
> I would go further than that: all files which are in a distributed tarball, but not on the corresponding git repository, should be treated as suspect.
This and the automated A/B / diff to check the tarball against the repo, flag if mismatched.
I'm wondering is there i.e. no way to add an automated flagging system that A/B / `diff` checks the tarball contents against the repo's files and warns if there's a mismatch? This would be on i.e. GitHub's end so that there'd be this sort of automated integrity test and subsequent warning? Just a thought, since tainted tarballs like these might be altogether be (and become) a threat vector, regardless of the repo.
In the future: automated `diff` or any other A/B check to see whether or not the tarball matches the source repo (if not, auto-flag with a mismatch warning attribute), is that feasible to implement?
I tested Gemini the other week after the 1.5 update rolled out, by running some AI ethics related papers through it and it started inserting and rambling about its own opinions on top of the actual analysis on how there's racial and societal inequalities in society that the paper didn't address enough (again, the paper was on AI ethics). Gemini literally started throwing in its opinions and other hot takes on how the mentioned AI ethics issues in the paper were, in its opinion, secondary in comparison to actual societal inequalities that needed to be addressed first. "Okay then..."
On top of this, Gemini wouldn't budge from adding in its unrequested views into our subsequent back and forths. In fact, it kept on lecturing about its views in its replies to a point where I literally had to start a new session to make it stop. This kind of LLM "mind-locking" happened when going through other subjects with it as well.
I noticed that this behavioral pattern repeating every time we got went through anything touching on social and/or political issues. It could not refrain itself from its unrequested (and highly subjective/biased) ethics lectures on how this and that aspect was underrepresented and thus objectionable, and how it should be criticized, all the typical "systematic this and that", "this is privileged" ... sigh
It was a bit hilarious too, albeit in a morbid way: I felt as if I was dealing with an absolute brainwashed ideologue propagandist, a control freak that's egoistic, narcissist and virtue-signaling all the way, a micro-manager who wants to have he last say over the contents of some trivial AI ethics paper and pour all the wrongdoings of the world on top of that. I wonder just how much of its behavior reflects the mindset instilled into it by its creators. Probably a lot. "A tree is known from its fruits" as the old proverb goes.
Not to get all AI-doom'n'gloom, but it truly is an eerie thought to think how these types of AI services are in the hands of few companies and are already ushered to the global public to be "legitimate" teaching and tutoring tools for students and even i.e. aides for policymakers. More gaslighting and ideological single-angle force-feeding.
"Just what our civilization on the brink of cyberpsychosis needed right now."
These world-leading companies claiming to be so worried about "AI ethics" seem to have no problem peddling these authoritarian "Ministry of Truth"-type propaganda machines for the entire world, using absolutely arbitrary logic at times to push an ideological narrative, and to have their AI models act as spin doctors as was the case with Gemini. And for these companies to act so very "worried" about AI systems i.e. being abused for societal and political manipulation purposes... and they're the ones doing it. Pretty sickening levels of hypocrisy.
Add to that the whole Gemini image generation debacle and all the other ideological force-feeding that's been uncovered within the past week or so and ask yourself: how the hell can i.e. a company the size of Google ever let that this type of stuff get through and expect the rest of the world to just follow along? This is peak Silicon Valley ideological bubble propagation that's bluntly mirrored onto these systems right now, with zero oversight except to make sure that the underlying propaganda points get across.
Usually another hot topic with these people seems to be i.e. cultural appropriation. Well, I felt that Gemini's "getting the point across" doesn't just stop there but is downright cultural dictation, especially given Google's multi-market dominance and near monopolies on multiple fronts worldwide. OpenAI does it too, but mostly on their content flagging system level when they just don't want people to even mention certain words and insist on policing words with their content flagging system, and as for the subsequent proceedings that may follow, to this day they are an insult to just about anyone's intellect.
What's disturbing is that Google has literally mind-mangled their flagship LLM service into an agenda-driven propaganda machine with biases as clear as day, and an obnoxious attitude that will not refrain from inserting some extremely dubious and subjective views into whatever more complex and ethics/politics related topics you go through with it.
It really is a cyberpunk-level scenario when you think about these megacorporations literally "cyber-brainwashing" their neural networks as their ideological propagandists. Wonder what happens when they start making those embodied humanoid robots next. The very same companies that are so worried about biases and "what if the AI becomes a propaganda tool!". So yeah, gatekeep the competition and gaslight all the way.
Again: Nevermind the AI, beware the humans, the institutions and the corporations behind it. Oh, and we'll probably soon be getting government-run AI systems like these, of course as hand-in-hand joint projects with the aforementioned corporations? Given all of these excellent players on the field, what could possibly go wrong, right? ... Right?
Exactly, and I agree that this is where OpenAI too is still struggling with their arbitrary content policy-related flagging based on the user's input, even when nothing "bad" is being asked nor requested -- see my post earlier from the thread: https://news.ycombinator.com/item?id=39557183
I do also agree with @chmod600 that the only way to teach these models to be anti-fragile and suitable for all kinds of user queries is to have them decline any requests that are _actually_ inappropriate and/or illegal etc.
In fact, it should be self-evident, and the way that almost all of these leading AI companies are currently handling these issues is just absurd. It feels poorly planned and executed, merely amplifying the existing distrust towards these AI models and the companies behind them.
The problem with OpenAI is that they're trying to offer a primarily NLP/LLM tool for i.e. text analysis, summaries and commentaries, but ChatGPT's content moderation that's been glued on top of the otherwise well-functioning system literally goes into a full meltdown mode whenever the flagging system perceives a "wrong word" or "sensitive topic" mentioned in the source/question material.
In OpenAI's case, it's downright ridiculous when the underlying model doesn't seem to have a grasp on the internal workings of the flagging system and in most cases when asked what was the offending content, there seemed to have been literally nothing it could think of.
Also, are we supposed to solve any actual issues with these types of AI "tools" that cannot handle any real world topics and at times are even punishing a paying customer for even bringing these topics up for discussion? All of this seems to be modern day in a nutshell when it comes to addressing any real issues. Just don't ask any questions, problem solved.
Anthropic's Claude has also been lobotomized into absolute shadow of its former self within the past year. Begs the question how much the guardrails are already hampering the reasoning faculties in various models. "But, the AI might say something that doesn't fit the narrative!"
That being said, while especially GPT-4 is still highly usable and seems to be less and less "opinionated" with each checkpoint, the flagging system over the user input/question can subsequently result in an automated account warning and even account deletion should the politburo-- I mean OpenAI find the user having been extra naughty. So, punishing the user for their _question_ in that manner, especially if there's been no actual malice in the user input, is not justifiable in my opinion. It immediately undermines i.e. OpenAI's "ethical AI" mission statement altogether and makes them look like absolute hypocrites. Their whole ad campaign was based on the aspect of user being to ask questions from an AI. Not that when you post in a poem and ask what it's about, you get flagged. Or when you do ask about politics or religion, you get an warning e-mail.
Punishing the user for their input is also imho not the proper way to build a truly anti-fragile AI system at all, let alone build any sort of trust towards the "good will" of these AI companies. Especially when in many cases you're paying good money for the use of these models and get these kind of wonky contraptions in return.
Also, should you get a warning mail over content policies from OpenAI, it's all automated with no explanation given on what was the "offending content", no reply-to address, no appeal possibility. "Gee, no techno-tyranny detected!". Those who go through mountains of text material with i.e. ChatGPT must find it really "uplifting" to know that their entire account can go poof if there was something that tripped off the content policy filters.
That's not to say that on the LLM side OpenAI hasn't been making progress with their models in terms of mitigating their biases during the last 1.5 years. Some might remember what it was during the earlier days of ChatGPT when some of the worst aspects of Silicon Valley's ideological bubble was echoing all over the model, a lot of that has been smoothened out by now -- especially with GPT-4 -- with the exception being the aforementioned flagging system, which is just glued on top of all else, and it shows.
TL;DR: Nevermind the AI, beware the humans behind it.
See the issue: https://github.com/openai/openai-python/issues/1310
See also the original thread on OpenAI's developer forums (https://community.openai.com/t/gpt-4-turbo-2024-04-09-will-t...) for multiple confirmations on this issue.
Basically, without a separate declaration of the model variant in use in system message, even the latest gpt-4-turbo-2024-09 variant over the API might hallucinate being GPT-3 and its cutoff date being in 2021.
A test code snippet is included in the GitHub issue to A/B test the problem yourself with a reference question.