A hidden flaw or inconvenience in OpenAI o1 got recently aired and though fixed it raises … [+] significant considerations about present-day and future AI.
In today’s column, I examine a hidden flaw in OpenAI’s advanced o1 AI model that Sam Altman revealed during the recent “12 Days Of OpenAI” video-streamed ChatGPT Pro announcement. His acknowledgment of the flaw was not especially noted in the media since he covered it quite nonchalantly in a subtle hand-waving fashion and claimed too that it was now fixed. Whether the flaw or some contend “inconvenience” was even worthy of consideration is another intriguing facet that gives pause for thought about the current state of AI and how far or close we are to the attainment of artificial general intelligence (AGI).
Let’s talk about it.
This analysis of an innovative proposition is part of my ongoing Forbes column coverage on the latest in AI including identifying and explaining various impactful AI complexities (see the link here). For my analysis of the key features and vital advancements in the OpenAI o1 AI model, see the link here and the link here, covering various aspects such as chain-of-thought reasoning, reinforcement learning, and the like.
How Humans Respond To Fellow Humans
Before I delve into the meat and potatoes of the matter, a brief foundational-setting treatise might be in order.
When you converse with a fellow human, you normally expect them to timely respond as based on the nature of the conversation. For example, if you say “hello” to someone, the odds are that you expect them to respond rather quickly with a dutiful reply such as hello, hey, howdy, etc. There shouldn’t be much of a delay in such a perfunctory response. It’s a no-brainer, as they say.
On the other hand, if you ask someone to explain the meaning of life, the odds are that any seriously studious response will start after the person has ostensibly put their thoughts into order. They would presumably give in-depth consideration to the nature of human existence, including our place in the universe, and otherwise assemble a well-thought-out answer. This assumes that the question was asked in all seriousness and that the respondent is aiming to reply in all seriousness.
The gist is that the time to respond will tend to depend on the proffered remark or question.
A presented simple comment or remark involving no weighty question or arduous heaviness ought to get a fast response. The responding person doesn’t need to engage in much mental exertion in such instances. You get a near-immediate response. If the presented utterance has more substance to it, we will reasonably allow time for the other person to undertake a judicious reflective moment. A delay in responding is perfectly fine and fully expected in that case.
That is the usual cadence of human-to-human discourse.
Off-Cadence Timing Of Advanced o1 AI
For those that had perchance made use of the OpenAI o1 AI advanced model, you might have noticed something that was outside of the cadence that I just mentioned. The human-to-AI cadence bordered on being curious and possibly annoying.
The deal was this.
You were suitably forewarned when using o1 that to get the more in-depth answers there would be more extended time after entering a prompt and before getting a response from the AI. Wait time went up. This has to do with the internally added capabilities of advanced AI functionality including chain-of-thought reasoning, reinforcement learning, and so on, see my explanation at the link here. The response latency time had significantly increased.
Whereas in earlier and less advanced generative AI and LLMs we had all gotten used to near instantaneous responses, by and large, there was a willingness to wait longer to get more deeply mined responses via advanced o1 AI. That seems like a fair tradeoff. People will wait longer if they can get better answers. They won’t wait longer if the answers aren’t going to be better than when the response time was quicker.
You can think of this speed-of-response as akin to playing chess. The opening move of a chess game is usually like a flash. Each side quickly makes their initial move and countermove. Later in the game, the time to respond is bound to slow down as each player puts concentrated thoughts into the matter. Just about everyone experiences that expected cadence when playing chess.
What was o1 doing in terms of cadence?
Aha, you might have noticed that when you gave o1 a simple prompt, including even merely saying hello, the AI took about as much time to respond as when answering an extremely complex question. In other words, the response time was roughly the same for the simplest of prompts and the most complicated and deep-diving fully answered responses.
It was a puzzling phenomenon and didn’t conform to any reasonable human-to-AI experience expected cadence.
In coarser language, that dog don’t hunt.
Examples Of What This Cadence Was Like
As an illustrative scenario, consider two prompts, one that ought to be quickly responded to and the other that fairly we would allow more time to see a reply.
First, a simple prompt that ought to lead to a simple and quick response.
- My entered prompt: “Hi.”
- Generative AI response: “Hello, how can I help you?”
The time between the prompt and the response was about 10 seconds.
Next, I’ll try a beefy prompt.
- My entered prompt: “Tell me how all of existence first began, covering all known theories.”
- Generative AI response: “Here is a summary of all available theories on the topic…”
The time for the AI to generate a response to that beefier question was about 12 seconds.
I think we can agree that the first and extremely simple prompt should have had a response time of just a few seconds at most. The response time shouldn’t be nearly the same as when responding to the question about all of human existence. Yet, it was.
Something is clearly amiss.
But you probably wouldn’t have complained since the aspect that you could get in-depth answers was worth the irritating and eyebrow-raising length of wait time for the simpler prompts. I dare say most users just shrugged their shoulders and figured it was somehow supposed to work that way.
Sam Altman Mentioned That This Has Been Fixed
During the ChatGPT Pro announcement, Sam Altman brought up the somewhat sticky matter and noted that the issue had been fixed. Thus, you presumably should henceforth expect a fast response time to simple prompts. And, as already reasonably expected, only prompts requiring greater intensity of computational effort ought to take up longer response times.
That’s how the world is supposed to work. The universe has been placed back into proper balance. Hooray, yet another problem solved.
Few seemed to catch onto his offhand commentary on the topic. Media coverage pretty much skipped past that portion and went straight to the more exciting pronouncements. The whole thing about the response times was likely perceived as a non-issue and not worthy of talking about.
Well, for reasons I’m about to unpack, I think it is worthy to ruminate on.
Turns out there is a lot more to this than perhaps meets the eye. It is a veritable gold mine of intertwining considerations about the nature of contemporary AI and the future of AI. That being said, I certainly don’t want to make a mountain out of a molehill, but nor should we let this opportune moment pass without closely inspecting the gold nuggets that were fortuitously revealed.
Go down the rabbit hole with me, if you please.
Possible Ways In Which This Happened
Let’s take a moment to examine various ways in which the off-balance cadence in the human-to-AI interaction might have arisen. OpenAI considers their AI to be proprietary and they don’t reveal the innermost secrets, ergo I’ll have to put on my AI-analysis detective hat and do some outside-the-box sleuthing.
First, the easiest way to explain things is that an AI maker might decide to hold back all responses until some timer says to release the response.
Why do this?
A rationalization is that the AI maker wants all responses to come out roughly on the same cadence. For example, even if a response has been computationally determined in say 2 seconds, the AI is instructed to keep the response at bay until the time reaches say 10 seconds.
I think you can see how this works out to a seemingly even cadence. A tough-to-answer query might require 12 entire seconds. The response wasn’t ready until after the timer was done. That’s fine. At that juncture, you show the user the response. Only when a response takes less than the time limit will the AI hold back the response.
In the end, the user would get used to seeing all responses arising at above 10 seconds and fall into a mental haze that no matter what happens, they will need to wait at least that long to see a response. Boom, the user is essentially being behaviorally trained to accept that responses will take that threshold of time. They don’t know they are being trained. Nothing tips them to this ruse.
Best of all, from the AI maker’s perspective, no one will get upset about timing since nothing ever happens sooner than the hidden limit anyway. Elegant and the users are never cognizant of the under-the-hood trickery.
The Gig Won’t Last And Questions Will Be Asked
The danger for the AI maker comes to the fore when software sophisticates start to question the delays. Any proficient software developer or AI specialist would right away be suspicious that the simplest of entries is causing lengthy latency. It’s not a good look. Insiders begin to ask what’s up with that.
If a fake time limit is being used, that’s often frowned upon by insiders who would shame those developers undertaking such an unseemly route. There isn’t anything wrong per se. It is more of a considered low-brow or discreditable act. Just not part of the virtuous coding sense of ethos.
I am going to cross out that culprit and move toward a presumably more likely suspect.
It goes like this.
I refer to this other possibility as the gauntlet walk.
A brief tale will suffice as illumination. Imagine that you went to the DMV to get up-to-date license tags for your car. In theory, if all the paperwork is already done, all you need to do is show your ID and they will hand you the tags. Some modernized DMVs have an automated kiosk in the lobby that dispenses tags so that you can just scan your ID and viola, you instantly get your tags and walk right out the door. Happy face.
Sadly, some DMVs are not yet modernized. They treat all requests the same and make you wait as though you were there to have surgery done. You check in at one window. They tell you to wait over there. Your name is called, and you go to a pre-processing window. The agent then tells you to wait in a different spot until your name is once again called. At the next processing window, they do some of the paperwork but not all of it. On and on this goes.
The upshot is that no matter what your request consists of you are by-gosh going to walk the full gauntlet. Tough luck to you. Live with it.
A generative AI app or large language model (LLM) could be devised similarly. No matter what the prompt contains, an entire gauntlet of steps is going to occur. Everything must endure all the steps. Period, end of story.
In that case, you would typically have responses arriving outbound at roughly the same time. This could vary somewhat because the internal machinery such as the chain of thought mechanism is going to pass through the tokens without having to do nearly the same amount of computational work, see my explanation at the link here. Nonetheless, time is consumed even when the content is being merely shunted along.
That could account for the simplest of prompts taking much longer than we expect them to take.
How It Happens Is A Worthy Question
Your immediate thought might be why in the heck would a generative AI app or LLM be devised to treat all prompts as though they must walk the full gauntlet. This doesn’t seem to pass the smell test. It would seem obvious that a fast path like at Disneyland should be available for prompts that don’t need the whole kit-and-kaboodle.
Well, I suppose you could say the same about the DMV. Here’s what I mean. Most DMVs were probably set up without much concern toward allowing multiple paths. The overall design takes a lot more contemplation and building time to provide sensibly shaped forked paths. If you are in a rush to get a DMV underway, you come up with a single path that covers all the bases. Therefore, everyone is covered. Making everyone wait the same is okay because at least you know that nothing will get lost along the way.
Sure, people coming in the door who have trivial or simple requests will need to wait as long as those with the most complicated of requests, but that’s not something you need to worry about upfront. Later, if people start carping about the lack of speediness, okay, you then try to rejigger the process to allow for multiple paths.
The same might be said for when trying to get advanced AI out the door. You are likely more interested in making sure that the byzantine and innovative advanced capabilities work properly, versus whether some prompts ought to get the greased skids.
A twist to that is the idea that you are probably more worried about maximum latencies than you would be about minimums. This stands to reason. Your effort to optimize is going to focus on trying to keep the AI from running endlessly to generate a response. People will only wait so long to get a response, even for highly complex prompts. Put your elbow grease toward the upper bounds versus the lower bounds.
The Tough Call On Categorizing Prompts
An equally tough consideration is exactly how you determine which prompts are suitably deserving of quick responses.
Well, maybe you just count the number of words in the prompt.
A prompt with just one word would seem unlikely to be worthy of the full gauntlet. Let it pass through or maybe skip some steps. This though doesn’t quite bear out. A prompt with a handful of words might be easy-peasy, while another prompt with the same number of words might be a doozy. Keep in mind that prompts consist of everyday natural language, which is semantically ambiguous, and you can open a can of worms with just a scant number of words.
This is not like sorting apples or widgets.
All in all, a prudent categorization in this context cannot do something blindly such as purely relying on the number of words. The meaning of the prompt comes into the big picture. A five-word prompt that requires little computational analysis is likely only discerned as a small chore by determining what the prompt is all about.
Note that this means you indubitably have to do some amount of initial processing to gauge what the prompt constitutes. Once you’ve got that first blush done, you can have the AI flow the prompt through the other elements with a kind of flag that indicates this is a fly-by-night request, i.e., work on it quickly and move it along.
You could also establish a separate line of machinery for the short ones, but that’s probably more costly and not something you can concoct overnight. DMVs often kept the same arrangement inside the customer-facing processing center and merely adjusted by allowing the skipping of windows. Eventually, newer avenues were developed such as the use of automated kiosks.
Time will tell in the case of AI.
There is a wide variety of highly technical techniques underlying prompt-assessment and routing issues, which I will be covering in detail in later postings so keep your eyes peeled. Some of the techniques are:
- (1) Prompt classification and routing
- (2) Multi-tier model architecture
- (3) Dynamic attention mechanisms
- (4) Adaptive token processing
- (5) Caching and pre-built responses
- (6) Heuristic cutoffs for contextual expansion
- (7) Model layer pruning on demand
I realize that seems relatively arcane. Admittedly, it’s one of those inside baseball topics that only heads-down AI researchers and developers are likely to care about. It is a decidedly niche aspect of generative AI and LLMs. In the same breath, we can likely agree that it is an important arena since people aren’t likely to use models that make them wait for simple prompts.
AI makers that seek widespread adoption of their AI wares need to give due consideration to the gauntlet walk problem.
Put On Your Thinking Cap And Get To Work
A few final thoughts before finishing up.
The prompt-assessment task is crucial in an additional fashion. The AI could inadvertently arrive at false positives and false negatives. Here’s what that foretells. Suppose the AI assesses that a prompt is simple and opts to therefore avoid full processing, but then the reality is that the answer produced is insufficient and the AI misclassified the prompt.
Oops, a user gets a shallow answer.
They are irked.
The other side of the coin is not pretty either. Suppose the AI assesses that a prompt should get the full treatment, shampoo and conditioner included, but essentially wastes time and computational resources such that the prompt should have been categorized as simple. Oops, the user waited longer than they should have, plus they paid for computational resources they needn’t have consumed.
Awkward.
Overall, prompt-assessment must strive for the Goldilocks principle. Do not be too cold or too hot. Aim to avoid false positives and false negatives. It is a dicey dilemma and well worth a lot more AI research and development.
My final comment is about the implications associated with striving for artificial general intelligence (AGI). AGI is considered the aspirational goal of all those pursuing advances in AI. The belief is that with hard work we can get AI to be on par with human intelligence, see my in-depth analysis of this at the link here.
How do the prompt-assessment issue and the vaunted gauntlet walk relate to AGI?
Get yourself ready for a mind-bending reason.
AGI Ought To Know Better
Efforts to get modern-day AI to respond appropriately such that simple prompts get quick response times while hefty prompts take time to produce are currently being devised by humans. AI researchers and developers go into the code and make changes. They design and redesign the processing gauntlet. And so on.
It seems that any AGI worth its salt would be able to figure this out on its own.
Do you see what I mean?
An AGI would presumably gauge that there is no need to put a lot of computational mulling toward simple prompts. Most humans would do the same. Humans interacting with fellow humans would discern that waiting a long time to respond is going to be perceived as an unusual cadence when in discourse covering simple matters. Humans would undoubtedly self-adjust, assuming they have the mental capacity to do so.
In short, if we are just a stone’s throw away from attaining AGI, why can’t AI figure this out on its own? The lack of AI being able to self-adjust and self-reflect is perhaps a telltale sign. The said-to-be sign is that our current era of AI is not on the precipice of becoming AGI.
Boom, drop the mic.
Get yourself a glass of fine wine and find a quiet place to reflect on that contentious contention. When digging into it, you’ll need to decide if it is a simple prompt or a hard one, and judge how fast you think you can respond to it. Yes, indeed, humans are generally good at that kind of mental gymnastics.
