ChatGPT help me solve a refactoring bug today. I had spent hours messing around trying to figure out what the issue was until I realized, via asking ChatGPT, that I had misunderstood a piece of the code and the docs. It was able to answer and provide examples (until it had error and crashed) in a way a senior engineer might have been able to.
The funny thing is I had tried just pasting in code and saying "find the bug" and it wasn't helpful at all, but when I posted in a portion and asked it to explain what the code was doing I was able to work backwards and solve the issue.
Its nice anecdote where the AI felt additive instead of existentially destructive which has been a overbearing anxiety for me this last month.
If you’re willing to overlook the small detail that it is the duck explaining to you how your code works rather than you explaining to the duck then sure
It’s been said that the best way to get the correct answer from somebody is not to ask for it, but to instead give them an obviously incorrect answer. This annoys people enough for them to do the necessary research to come up with the correct answer, in order to prove the incorrect answer to be wrong. It might be that ChatGPT can be used to weaponize this tendency.
Maybe Stack Overflow should have an automatic pinned answer from ChatGPT (clearly labeled as such) on all questions, in order to goad people into providing the actually correct answer?
From Star Trek: First Contact: "When you build a machine to do a man's job, you take something away from the man."
You surrendered the need to think to the machine. You are lesser for it. I don't think these AIs are just removing drudgery, like, say, a calculator. They actually do the work. Or more correctly, they produce something that will pass for the work.
Wholesale embracing of this sort of technology is bad for us.
I disagree with this being an example of your point. OP tried to be lazy “find the bug”, but that failed because the machine could not think for them. Then OP used a tool that helped them better understand the code, and found the bug. That’s not lazy, that’s smart.
That’s the message of Frank Herbert in the Dune universe regarding machines. They’re useful up until you start letting them do the thinking for you. Which leads to laziness, stagnation and control by an elite. So did reliance on spice for that matter.
> Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.
- - - -
This is good and well for existing minds, but I think a lot of people will let these machines raise their children and that might cripple them. Like if you let your kids use a wheelchair instead of learning to walk? It's an imperfect metaphor.
The new movie M3GAN raises that question. It's not a great movie, but even without the eventual horror elements, it would raise disturbing implications for creating doll robots that can do the work of parenting so parents can focus on what they really care about, like work.
I don't know about you but if I had the ability to dictate requirements and to get a program out the other side that matches those requirements, the process of coding has become mere busywork that can be eliminated for the benefit of me and everyone else.
I'm sure the buggy whip makers had pride in their work as well.
But none of what you are talking about has happened today and even the buggy whip reference doesn't make sense because the buggy whip market disappeared because we stopped using horses. The buggy whip equivalent would be the IDE.
GPT3 has shown how ML can be trained on multiple unstructured data sources to produce structured information on demand.
Iterate a few more versions from here, so that the models are stronger at producing the correct structured data, and the impact on every office job will be profound.
I.e. instead of training a generative model on text from the internet, train it on every single excel file, sql database, word document and email your company stores. Then query this model asking it to generate Report X showing Y and Z.
When you step back and consider it, 99% of office jobs are about producing structured data from unstructured data sources. The implications of this are being hugely underestimated.
Nah. When AI is able to do all what you have said, requirements will just get harder and humans will still have to put hours to make something done. Just like 30 years ago it wasn't feasible to implement streaming music over the internet in a weekend, and now any teenager can do so by just 'npm install'ing... AI will only open the door to even more complex problems to solve.
Both of those things can be true... it's not just one or the other. There are plenty of jobs right now that are more menial tasks which will definitely be replaced by this, and I feel for those older folks in those positions because they'll most likely lose their jobs / be impacted by this.
Exactly. People forget that there used to be (up to about thirty years ago) a large number of people employed as "file clerks". Their job was to file and retrieve on demand paper documents in filing cabinets. When relational databases became practical, this information was stored and retrieved electronically and the entire job of "file clerk" was made obsolete.
Consider "report generator" as one category, throw in "Buzzfeed writer" and "Stack Overflow copy/paster", and a clear bimodal distribution emerges. The human touch is still necessary to add context and distinguish fact from plausible hallucination, but experts can now scale their contributions 10x as a result of immediate access, minimal latency, and reduced communication costs.
We're moving towards a world of chair-fillers at one end, and maestros at the other. The clearest difference between labor in 2022 and 2026 will be the hollowing-out of the middle.
The value of a human is in reacting to changing requirements, considering context and in understanding other humans. AI cannot do any of that reliably.
Some office tasks can be automated and those that can don't need AI anyway - they need properly labelled data, databases and some coding.
AI will be very good at creating the illusion of competence. AI cannot actually ensure competence or verify it. That will remain the domain of humans.
>I.e. instead of training a generative model on text from the internet, train it on every single excel file, sql database, word document and email your company stores. Then query this model asking it to generate Report X showing Y and Z.
This has already been possible for decades using old-fashioned automation (Python scripts etc.), assuming the data entry is designed for this.
Honestly, I think the reason managers have teams of people reporting to them is not just to give them unbiased information.
Part of it is probably ego stroking, but I suspect the humans in the loop are doing some sort of analysis too, and reporting qualitative patterns that an AI might not pick up on.
Deep thinking work will not go away. From trades to tech. It'll change work for sure, but it's not going to obliterate workers jobs.
I'm no luddite, but I've seen enough rocky digital transformations to know that human beings don't operate like manufacturing pipelines. Automation and AI assisted automation will be harder to generally implement.
But what I do feel confident about is that there will be a large mass of consultants who'll sell and expensive dream to a lot of mid tier businesses. The next big flex for business IT will be to have a notch on your belt for a failed AI automation project.
> Widespread adoption of generative AI will act as a lubricant between systems,
I largely agree with this article, but I feel like you have to be careful with these general predictions. Many technologies have purported themselves to be this "business lubricant" tech (ever since the spreadsheet), but the actual number of novel spreadsheet applications remains small. It feels like the same can be said for generative AI, too. Almost every day I feel the need to explain that "generation" and "abstract thought" are distinct concepts, because conflating the two leads to so much misconception around AI. Stable Diffusion has no concept of artistic significance, just art. Similarly, ChatGPT can only predict what happens next, which doesn't bestow it heuristic thought. Our collective awe-struck-ness has left us vulnerable to the fact that AI generation is, generally speaking, hollow and indirect.
AI will certainly change the future, and along with it the future of work, but we've all heard idyllic interpretations of benign tech before. Framing the topic around content rather than capability is a good start, but you easily get lost in the weeds again when you start claiming it will change everything.
That's fair enough, I've seen some pretty cool things in spreadsheet software too.
My larger point, though, is that most people end up using spreadsheets to do the same thing. It's fun to imagine novel uses for a spreadsheet, like a DAW or video game, but ultimately it's not very useful for that. Similarly, ChatGPT is great for writing convincing text - that's what everyone uses it for. Can it solve math though? Not very well. Future applications of the tech are more likely to be specialized, in that sense.
Mostly, I'm a curmudgeon and I despise these "flying car of the future" articles. Popular Mechanics printed them for decades, and half a century later nothing has changed (not even the culture writing them).
I think we knew from the get-go that spreadsheets would be used for pretty much anything to do with numbers. That there aren't any new applications past that understates their general applicability.
I agree though, chatGPT isn't a real flying car. Imagine if someone revolutionized the paper clip. The day-to-day of millions would be forever and irrevocably changed; and almost nothing would happen.
>Mostly, I'm a curmudgeon and I despise these "flying car of the future" articles.
When you understand how the sausage is made it is hard to be overly excited. I fail to be "mind blown" by ChatGPT because every time someone claims it can do task X they only managed to scrape by within its significant limitations.
If you want superhuman intelligence you are going to need to break through the short term memory limitation of humans. If the AI can memorize a 1 million line code base then it will be practical but everyone is only working with small code snippets or generates the entire code from scratch and then extrapolates that to a million line codebase even though that isn't possible. That is the height of impracticality.
And before anyone accuses me of moving goalposts. I'm not the one moving them. It's the people telling me it can make manual programming obsolete. Why not just stick with what it can do instead of making these claims?
It's not just that a system built in MS Access facing scale concerns needs a rewrite from an engineer's perspective.
It's that the business will also accept that it needs a rewrite. As opposed to the current status quo where they'll ask what's wrong with continuing to use $Slick_and_Fancy_Tool (then act surprised when it stops scaling with regards to whatever business, performance, or compliance barriers you've then reached).
I always tell people that spreadsheets should be used to display data, not store it. As soon as you use spreadsheets to store data you've begun the descent into darkness.
> Our collective awe-struck-ness has left us vulnerable to the fact that AI generation is, generally speaking, hollow and indirect.
This totally resonates with me. This is absolutely correct. Thinking about the future of work, there's much of what I do every day in my job that is hollow and indirect. And I would be totally okay if I could have something like ChatGPT do it for me.
Wall-E was about as much about post-scarcity as it was about escaping reality. To me it looks like we've focused on the second part and we got pretty good at it.
I don't have a problem with the main point of the article, but there is a huge terminology confusion that is rapidly gathering force to confuse people. The key breakthroughs of GPT3 et al are not primarily about generative AI. People had been building generative models long before GPT3, and it was generally found that discriminative models had better performance.
They key to the power of GPT3 is that it has billions of parameters, AND those parameters are well-justified because it was trained on billions of documents. So the term should be something like "gigaparam AI" or something like that. Maybe GIGAI as a parallel to GOFAI. If you could somehow build a gigaparam discrimative model, you would get better performance on the task it was trained on than GPT3.
Good point on the terminology. What do you think the right terminology should be? LLMs is too much of a mouthful and is not as informative for the general public, imo. People are also using Foundation Models, which I rather like.
+1 to Foundation Models. I don't share your concerns about LLMs, though, and often refer to the future involving LXMs where X could be images, audio, bioinformatics data, etc
I don't like "Foundation Models" because it's a term invented by Stanford and they're pushing it hard while not really doing all that much in the field.
ChatGPT and generative AI will not
who write these things for a living will definitely have fewer clients. But this is a small percentage of paid writing, and not the most lucrative or desirable.
I do not think that the world is changing because of large language models. That seems to be a controversial opinion so I won't get into it here. But these are powerful new tools, no question. The way I work has changed and I'm very glad to have ChatGPT.
I do believe that in the coming years knowing how to use ChatGPT or similar products will be as important as knowing how to use Google is now. People that know how to leverage LLMs going forward will simply have an advantage over those who don't. It won't be long before it isn't optional for executives and knowledge workers. This will be a big change for many people. But we adapted to Google in the early 2000s and people will adapt to this as well.
Or worse, subtly inaccurate. The problem I have with generative AI right now, its product looks like it makes sense and sometimes it does, but there is always the risk of total nonsense hidden somewhere in the middle. So you still need someone capable to check and correct for most professional work, and sometimes that is harder or more time consuming than making the product itself.
The same sort of problem with self driving cars, they are often correct but not often enough, and staying alert to correct the AI is worse than driving yourself which is more work, paradoxically enough.
AI might manage to push through these barriers, but I remain skeptical with the technology in the current state: statistical machines that are good in the common cases but sketchy at the edges.
Which is similar to correcting subpar human results:
Rather than meticulously correcting the works of a subpar programmer it is much more efficient to let a proficient developer produce the code. Or even better, engage a 10x developer.
If an inefficient programmer can solve a real world business problem in a fraction of the time, but it requires slightly more computing resources, I’d still pay for it. Efficiency can be measured in a lot of dimensions and often times spending time and money writing optimized code is inefficient as well.
> there is always the risk of total nonsense hidden somewhere in the middle
This is true of human-generated code as well. Trust me: Reviewing other people's code is my day job.
It's exceptionally rare that a malicious actor is trying to sneak something into the code. The common scenario is the developer who's new to the project not fully understanding how everything works so they copy & paste something they think is necessary but ultimately isn't and could in fact be very wrong. Just like how ChatGPT works.
I've been asking friends in non-programming engineering fields how ChatGPT does in their area of expertise, and I believe programming is the area that ChatGPT is the most accurate. Finding solution to general engineering problems seems blatantly wrong in almost all cases, whereas in programming, it seems to be able to generate mostly correct code for simple, boiler-plate like tasks.
yes, but why? Why is GPT so much better at programming than other tasks?
can it be that programming itself can be so easily predicted in a generative way, while others require more ingenuity and real world model to be solved?
In this case I would totally offload programming to a GPT /LLM AI, while my job is simply to specify largely the business case.
I have to imagine its because so much of its training data is readily available programming docs, tutorials, and general Q&A that there is an amazing abundance of online. How many times have you just pasted an error into google and hoped someone else has asked the exact same question on stack overflow?
As Chomsky points out, the AIs understand syntax not semantics. Code is all (or almost all) syntax whereas human language has both syntax and meaning. Meaning is not something ChatGPT understands.
Is it because programming is a more limited and specific language than the ones people speak? There's less room for double-meanings, slang, meaning, or even sentence structure.
It seems the improvement in the models is the size of the training set. Hence the "Large" in large language models. I am assuming that it is better at programming because there a is a ton of training data available on Github, etc. There isn't a similar set of data on solving a variety of physics problems.
Yup. What seems to be largely missed is that these models have zero understanding, and are actually destroyers of information, not creators. In classic Information Theory, information is basically surprise value — how much unexpected info is in the message? — yet these "AI" systems put out the most expected subset in each instance. This highly averaged output is very recognizable and so very striking, but it is not actually very informative (perhaps except in cases where it is specifically used as a verbose search engine, where the query takes advantage of the breadth of the AI's training).
> In classic Information Theory, information is basically surprise value — how much unexpected info is in the message? — yet these "AI" systems put out the most expected subset in each instance.
Forgive me, but isn't this kind of moving-the-goalposts? Information is the surprise value from the recipient's point of view, which meas the recipient's bayesian prior probability is "expected". Saying "these "AI" systems put out the most expected subset in each instance" assumes that the recipient's priors exactly equal those of the model which would only be the case when the model is talking to itself. (or I suppose to an even more complex model with perfect knowledge of ChatGPT's weights)
The fact that no information is transferred when the model talks to itself should not be surprising and would apply to any AI. (even including a superhuman post-singularity god-like AI)
Yes, the AI's output could be surprising to the point of view of many recipients.
This does not mean anything more than that the AI has a greater breadth of training background, which is likely.
We get the output most likely expected from any of (or the average of) the humans whose writing/drawing/whatever was included in the input set.
What we will not be getting from the AIs is any creative output based on unique understanding, as we would from an intelligent, creative human. Many of hte humans in the input set would see the same prompt and produce an actual novel and meaningful output, not simply a cut-and-paste from prior works. (& yes, seme novel output may come from some randomizing algo, but if it is correct, it is no more correct than the broken clock that is correct twice every day.)
Or, another example, I was involved in a legal deposition where an "AI" transcription system was used instead of a skilled court reporter. The output LOOKED fantastic, until I actually read it, and it was absolute garbage. The standard errata sheet has room for the deponent to put in about a dozen corrections, and most are less than a handful. My errata list was multiple pages. These errors often reversed the meaning of sentences, substitutin "I have ..." for "You have...", dropping or adding "not", or substituting in common names for unusual names (e.g., "Jack Kennedy" for "John Kemeny". note human transcribers always ask for correct spellings of names in the next break, this crap just inserted it like it had a clue).
So, even though the total "experience" or training set of the may go beyond the experience of the reader, so some of the output is surprising, this is no more so than a search engine produces surprise. In fact, I think this is the best use of the AIs, to have them trained on an enormous data set, and provide possibly better results, defined as more on-point, but likely less thorough.
>This does not mean anything more than that the AI has a greater breadth of training background, which is likely.
This has nothing to do with Shannon's definition of information. The light from a star going supernova or a tsunami hitting shore convey information without any kind of agency nor intent being involved.
>What we will not be getting from the AIs is any creative output based on unique understanding, as we would from an intelligent, creative human.
This is a really big assumption here that I don't think is justified. Are you assuming that humans have some non-physical soul which makes us somehow different from any other deterministic information-processing system?
>>The light from a star going supernova or a tsunami hitting shore convey information without any kind of agency nor intent being involved.
Right, and I'm obviously speaking loosely, but the amount of information or information value is the amount of non-redundant 'surprise'.
Receiving the supernova's signal for the first time contains valuable information, but replaying it does not.
>>Are you assuming that humans have some non-physical soul which makes us somehow different from any other deterministic information-processing system?
No, I'm speaking about the current and near-intermediate-term state of AI. This current state is basically a mashup of everything it's 'seen' in training, but without a shred of understanding. It only spits back what is statistically the most likely string of words to occur in any situation.
Ordinary humans do far better because they have understanding. A simple example shows this: Asking ChatGPT this question: "Mikes mum had 4 kids; 3 of them are named Luis, Drake, and Matilda. What is the name of the 4th kid?" [0] ChatGPT utterly fails, even when told the answer is in the question.
As in my example with the deposition, it looks great but it is just a literal mashup of training data, and worse yet, is only outputting the most average data. This produces surprisingly emotionally satisfying results, which may be useful in some contexts (marketing copy?) but isn't close to the reasoning of an ordinary human.
I'm making no claim about whether AI can get there. I suspect it ultimately will. BUT, this layer is only a small component of what must be built to get to actually abstract concepts and wield them like any child. My minor in college was neuroscience and I still follow it a bit; we're a LOOONG way from understanding what it is to get and weild concepts, nevermind how to model or reproduce that in a computer.
That's all I'm saying - the current crop is impressive, maybe even useful in some applications, but nowhere near the kind of "intelligence" that is being touted.
No, but I do think that both of these things are revealing to us just how bad highly technical people and highly intuitive and empathetic people are at properly communicating with one another.
(It is a broad generalization to assume that these traits are mutually exclusive. They do, of course, co-exist in many people. However, it seems to me that the number of people in whom they coexist robustly are few. However, if is these few from whence come true once-in-a-generation geniuses.)
what a weird moment to take a stab at AI, right now when all the projects are starting to bear fruit and we can see advances that remind me of moore's law in the 90s.
these projects have direct commercial applications right now.
What you say about AI is true; however, from where I'm standing, it seems there is still too much greed-driven, mindless "no problem; we'll simply brute-force the problem of inventing a machine that does human-level cognition, which is a thing we admit we do not understand at all"-type enthusiasm, and not enough openness, humility, and critical thinking in the field.
I write educational technical articles for a living. Dev tools, frameworks, security, APIs, infrastructure, web3, etc.
I talk to the AI as if I would interview an expert on a subject matter.
This usually gives a good starting point for an article, if the subject is general enough, and not too new.
It's also good at structuring and rewriting texts. If you already have all the correct data, you can use it to write an outline or something like that.
The problems I saw were that it can't follow a coherent thought for more than a few paragraphs, and the writing style is generally a bit boring.
Also, the system uses sampling of results to sound more interesting and to prevent overfitting, it happens regularly that it tells you crap. One time you get a good answer, then you change one word in your prompt and the results isn't accurate anymore.
But I worked for years as a developer, so I usually notice when things are off, and I also fact check manually with Google when I want to be sure.
No offense but this approach worries me - it seems like a novel mechanism to (perhaps inadvertently) generate and spread false information. It takes a lot of fact checking to make sure everything is right, and if you do the research yourself that's a natural part of the process. It seems way too easy to minimize that effort in a process like this.
I was already worried about ChatGPT-like systems generating mass-produced nonsense and polluting the internet, but if people are also going to edit ChatGPT output just enough to make it seem right (a mechanism I hadn't thought of so far), that might make the nonsense a lot harder to detect.
I totally understand the reasoning though, it sounds like a productive workflow.
So how do you know if it's lying to you if you don't do the research?
In the short-term I'm certain that you have the background knowledge to detect when generated content is not quite correct - but going forward as your own personal knowledge atrophies (since you're letting AI do your research) - will you still be able to make sure it's not making everything up?
Do you have examples you can provide of these technical articles? Because those topics your offered are really broad and very few people are knowledgeable about all of them, so it sounds like you're filling in your knowledge by querying ChatGPT.
Using ChatGPT to fill in knowledge for a technical articles sounds bad. If I'm reading an article about security, I want it written by a security expert not a semi-layman plus a ChatGPT model.
Some things are more critical than others and many of the articles I write are pretty mainstream. After all, clients pay for articles that get them customers, and niche stuff often doesn't help here.
Also, security is a broad topic. I wouldn't write an article that explains how to implement your own secure hash algorithm. Often I just tell people how they could lose their keys or such things.
But yeah, thanks for your input. Learning what readers care about is crucial for my job and I'm always trying to get better at it.
I paid for Tabnine pro since it was 50% off for a year but I won't renew it unless it massively improves.
I mean, it does give good completions sometimes but the time saved isn't that great imho. Maybe chatgpt is better but it feels like AI still have some way to go to actually be so useful you would be less sucessful without it.
Their product MaestroAI is marketed as “for teams” (and of course with the obligatory fading-color call-to-action buttons) presumably to attract VC $$$ but I would love something like this (powered by LLMs) to extract info from all my documents.
Maybe something like this exists? Please no DEVONThink suggestions :)
On the topic of Content is King, I have a different view than the author. I think in the case of these trained AIs, 'content' refers to the training datasets and not the generated outputs.
Trained AIs are in something like the early digital streaming days where there was only one provider in town, so that provider aggregated All The Content. Over the next decade we would see the content owners claw their content back from Netflix, and onto competitor platforms -- which takes us to where we are today. Netflix's third party content has dwindled and forced them to focus on creating their own first party content which can not be clawed away.
When these generative AIs start to produce income, it will be at the expense of the artists whose art was in the training dataset nonconsensually. This triggers the same content clawback we saw in digital streaming. Training datasets will be heavily scrutinized and monetized because the algorithms powering generative AIs aren't actually carrying much water. What is DALL-E without its dataset? Content is King.
The funny thing is I had tried just pasting in code and saying "find the bug" and it wasn't helpful at all, but when I posted in a portion and asked it to explain what the code was doing I was able to work backwards and solve the issue.
Its nice anecdote where the AI felt additive instead of existentially destructive which has been a overbearing anxiety for me this last month.