I think that ChatGPT's success might be partly attributable to its chat interface. For whatever reason, a lot of people - including me! - are much more forgiving of inconsistencies, slip-ups, and inaccuracies when in a conversational format. Kind of like how you might forgive a real human for making a mistake in conversation.
I don't think that's necessarily good, and might not have much connection to attempts to build new non-conversational products on top of LLMs, but maybe it has some explanatory power for the current situation.
- copilots are useful
- chat is entertaining and useful
- future tech is coming
- investment moneyWhich isn't to say that the tools aren't useful. But I have to add this fact because many people conflate any criticism with being dismissive of the technology. But technology is about progress, not completion. Gotta balance criticism and optimism. Optimism drives you and criticism directs you.
Completion models are obviously wrong very often. Instruct model was kinda ok, but you know it's a dumb machine.
Chat was a bit of an uncanny valley. I treated the instruct model like a child, but chat felt like having a conversation with someone of 80 IQ. It felt frustrating, and you ended up going "no no no, what I meant WAS ..." It felt like dealing with an incompetent colleague.
But I guess there's lots of views on it. Some expected it to be an oracle, even a god. Some treated it like Stack Overflow, then got frustrated that it was giving poor quality answers to poor quality questions. Some were just abusive to it. I suppose it's a mirror in a sense.
The key term here is "conversation". If I query something from the machine and it disappears and rumbles and then prints off something like a 1980s mainframe, with paper that has those holes on the side that you tear off... and then it's wrong, it's wasted time.
Meanwhile with the conversation I'm watching it in real time, and can stop it, refine it, or ask or clarification immediately and effectively. There is an expectation of give and take and "talking through" things to get to an answer, which I find is effective. I don't need it to be 100% right all the time, just 80% and then start parsing answers out of it to refine it to 90% accuracy with high confidence.
Now I am curious, what are some tasks that can accept a model that is at 80% as good as a human, but is 100x cheaper?(or, 100x faster?)
1: In my country, after ChatGPT launched last year, when you call customer support you are now prompted to “just say in a few words” what you want instead of going through tap-this-number menus (they exist as a fallback) and I believe the backend is an LLM. The user flow and voice recordings are still programmatically determined though, but I can easily see one streamlined model calling APIs and whatnot, handling it all.
Transcription, specifically Whisper, is one of those ML models where the accuracy is basically on-par with humans. So, I really expect a lot more to come out of real time voice/LLM integrations.(the ChatGPT voice thing is a good glimpse, but very janky)
Also I think this has been around for longer than chatgpt. It is often accompanied by a fake keyboard clicking noise.
Now imagine how well it works for people with non-"native" accents (even for native, I'd guess that a good Scouse/Glaswegian/Kiwi accent might confuse the hell of those systems as well). It's a disaster and I hate those.
See also: replacing physical buttons with convoluted swipe gestures on mobile devices in the never-ending quest to make screen as large as possible. When was this ever a user ask?
I feel sometimes like the present UX design is one large LLM-like hallucination.
Generating fake pictures also seems like it would be more ordinary false advertising.
but yeah liability would scare me, esp. because without it you can put the liability squarely on the restaurant or in some cases the person ordering / asking / buying.
The volume of helpdesk tickets large enterprises deal with is very easily and vastly underestimated. If you can even route 30% away from the central triage with 90+% accuracy and drop everything else back to the central triage... you suddenly safe 2 FTEs in that spot in some places. And increase customer satisfaction for most of those tickets because they get resolved faster.
Or, as much as people hate it, chatbots as a customer front. Yes, everyone here as an expert in a lot of tech has had terrible experiences with chatbots. Please mark your hate with the word "Lemon" in the comments. But decently implemented chatbots with a few systems behind them can resolve staggering amounts of simple problems from non-techies without human interaction from the company deploying them. It remains important to eventually escalate to humans - including the history from all of these interactions to avoid frustrations, sure.
Or, ticket/request preprocessing. Remember how spelling that 10 digit account number to a call center agent hard of hearing sucks? Those 4 retries because of you not using a better way to communicate that number also costs the company. Now, you can push a few of these retries into an AI system. If you mail them, an AI system can try to extract information like account numbers, intent, start of the problem, problem descriptions and such into dedicated fields to make the support agents faster.
Companies are certainly overdoing it at the moment, I'm not denying that. But a lot of the support/helpdesk pre-screening can be automated with current AI/ML capabilities very decently. Especially if you learn to recognize and navigate it.
A UK company, Octopus has been doing some interesting work on GenAI <> customer support, which is helped by their "Energy provider in a box" software called Kraken (https://octopusenergy.group/kraken-technologies), which gives a single unified view over their operations.
They even have support agent level of personalisation - i.e; the agent will talk in the tone of voice of a given agent via fine-tuning of their chat history.
I have yet to see a chatbot in a customer service function that isn't strictly worse than a button. Usually, the button is "request refund / return" for whatever reason. It's like captcha stuff, web site owner is too dumb (no offense) to figure out how to handle spam so they offload that task onto the customer.
It doesn't have to be perfect. It's not as if the actual data in there is perfect. It just has to be in a form where I can search it, ideally with named entities mapped.
Quality - like deciphering the writing on scrolls buried in volcanic ash in Herculaneum - gets all the attention. But what I really want is quantity - I want to be able to search through those 5000 pages of 200 year old mildly damaged cadastral records in dense handwriting. I want to relieve the army of kind retirees who currently transcribe these sorts of documents one by one based on their own needs.
You are assuming that the probability of failure is independent, which couldn't be further from the truth. If a digit recogniser can recognise one of your "hard" handwritten digits, such as a 4 or a 9, it will likely be able to recognise all of them.
The same happens with AI agents. They are not good at some tasks, but really really food at others.
Parent comment doesn't understand the concept of expectation, and this comment is apparently unfamiliar with the fact that SotA for digit recognition [0] has been much higher than 90% even in the 90s. 90% accuracy for digit recognition is what you get if you use logistic regression as your model.
My point was that numbers that look good in terms of research often aren't close to go enough for the real world. It's not that 90% works for zip codes, it's that in the 90s accuracy was closer to 99%. You have validated my point rather than rejected it.
And a LLM who only needs to call to a few API calls isn't hard.
Very little need perfect accuracy and for that we still have classical software.
CharGPT for writing emails to bigger audiences, Gramma correction.
Image generator for fun.
LLM for Feature Extraction from random text like a website or PDF (llama).
Another opportunity is that you can have less steps or more shared context. One interesting thing about Whisper is that it's not just straight speech recognition but can also be prompted and given context to understand what sort of thing the speech may be about, increasing its accuracy considerably. LLM Vision models also do this with things like OCR. This might not help it with the individual digits in an account number, but it does help with distinguishing an account number from a street address on a check.
Or to take another old-style ML technique, you probably shouldn't be doing sentiment analysis in some pipeline, because you don't need to: instead you should step back and look at the purpose of the sentiment analysis and see if you can connect that purpose directly with the original text.
All that said, you definitely can write pipelines with compounding errors. We haven't collectively learned how to factor problems and engineer these systems with LLMs yet. Among the things I think we have to do is connect the tools more directly with user intention (effectively flatting another error-inducing part of the pipeline), and make the pipelines collaborative with users. This is more complex and distinctly not autonomous, but then hopefully you are addressing a broader problem or doing so in a more complete way.
It works through an incredibly novel mathematical technique. You simply go to the patient and tell them they don't have cancer.
Since 99.5 percent of people don't have cancer, this classifier is 99.5 percent accurate.
Completely bullshit classifier for a completely bullshit metric.
Use sensitivity/specificity or precision/sensitivity instead
The handwritten automations have performed better and the issues are reproducible, so even when there are issues, there's some sense of forward progress as you fix them. With handing it all over to an agent, it really feels like running around in circles.
I think there's probably something here, but it's less trivial than just tossing a webpage at chatGPT and hoping for the best.
Ex. Content generation + zero-shot classification/mapping are powerful, and with a human in the loop (somewhat) responsible for accuracy, they can move much faster.
Language is essential for human civilization, so are tools. We wouldn't get far without either.
maybe a language model can understand what it needs to do but not how to do it, so you give it a tool.
Humans can get pretty far without 100 percent accuracy, we can get a lot from AI models before they reach 100 percent, but being that at some point AI will be able to improve itself even remake itself daily with 2x the abilities, 100 percent or at least 99.7 percent is attainable.
Right now I can take any YouTube video summarize it and turn it into a podcast, short form videos, and a blog post.
There's definitely a lot of marketing uses right now for ai agents. If you think about embodied AI, it's only as good as it's body. if it doesn't have good grippers it will struggle to pick things up.
Also with a lot of things, accuracy is subjective one person might think ad copy is great and maybe their manager thinks it's shit. One person could give it a 100 percent score and another a 70 percent.
My point is we're so close here, and it's already amazing technology and we can augment failures by creating larger toolboxes.
Up until fairly recently many systems used non-LLM models for making decisions based on natural language. Their performance would have been far worse but they still did useful work. Examples would include content policy enforcement, semantic search and so on.
There are very many cases where a system will make an automated decision on a heuristic or random basis for lack of better options. ML improved those decision points and spawned new ones. LLMs improve a subset of those decision points and spawn new ones.
It's way more nuanced than this. Of course, you need a decent "accuracy" (not necessarily the metric), but in many business cases, you don't need high accuracy. But you need a solid process: you can catch errors later, you can cross references etc, you need to failsafe, you need to have post-mortem error handling, etc...
I shipped stuff (classical ML) that was nothing more than "a biased coin flip," but that still generates value ($) due to the process around it.
what do you think would help people consider this before going down that path?
> not what is most likely to reflect reality.
Shouldn't there be a strong statistical correlation between the two? And, isn't that, fundamentally, more about intent of the training? If I train a model to predict what comes next in reality, it's through next word prediction, but it is predicting what reflects reality the best.
One thing that is still confusing to me, is that we've been building products with machine learning pretty heavily for a decade now and somehow abandoned all that we have learned about the process now that we're building "AI".
The biggest thing any ML practitioner realizes when they step out of a research setting is that for most tasks accuracy has to be very high for it be productizable.
You can do handwritten digit recognition with 90% accuracy? Sounds pretty good, but if you need to turn that into recognizing a 12 digit account number you now have a 70% chance of getting at least one digit incorrect. This means a product worthy digit classifier needs to be much higher accuracy.
Go look at some of the LLM benchmarks out there, even in these happy cases it's rare to see any LLM getting above 90%. Then consider you want to chain these calls together to create proper agent based workflows. Even with 90% accuracy in each task, chain 3 of these together and you're down to 0.9 x 0.9 x 0.9 = 0.73, 73% accuracy.
This is by far this biggest obstacle towards seeing more useful products built with agents. There are cases where lower accuracy results are acceptable, but most people don't even consider this before embarking on their journey to build an AI product/agent.