What it feels like to work in AI right now

From my perspective it’s just _confusing_ to work in AI right now. We have some massive models that are doing some really neat stuff, and apparently hundreds of millions of people are using them—but I keep wondering: to do _what_, exactly? I’m not asking what the models can do, I’m asking what people want the models to do every day, all the time.<p>I’ve been shown some neat pictures people made that they thought were cool. I don’t know that I need this every day.<p>I’ve seen examples of “write an email to my boss”. It would take me longer to explain to ChatGPT what I want than to write these myself.<p>I’ve seen “write a snippet of code” demos. But I hardly care about this compared to designing a good API; or designing software that is testable, extensible, maintainable, and follows reasonable design principles.<p>In fact, <i>no one</i> in my extended sphere of friends and family has asked me <i>anything</i> about chatGPT, midjourney, or any of these other models. The only people I hear about these models from are other tech people.<p>I can see that these models are significantly better than anything before, but I can’t see yet the “killer app”. (For comparison, I don’t remember anyone in my orbit predicting search or social networking being killer apps for the internet—but we all expected things like TV and retail sales to book online.)<p>What am I missing?

I work on a research team in FAANG. What it really feels like is one company made everyone else obsolete. And we are going to work working on NLP models that underperform ChatGPT by a huge margin. Twiddling my thumbs and keeping quiet while no one wants to recognize the elephant in the room.<p>Also, there is no &quot;working in AI&quot;, a few thousand people are doing real AI at most. The rest of us are calling an API.

This sub and blog post are all talking about ChatGPT which is no doubt amazing and far ahead of the curve. However I would also point to Metas new vision model:<p>SAM (Segment Anything): it is so far beyond any other vision model, I actually believe vision will be solved in a few years now. People don’t realize that there was an industry of publishing paper with incremental improvements in small datasets in CVPR that has been completely invalidated by this paper. I’ve seen engineers in Cruise segmentation team, say Metas new model seems to work better than the in house models they developed and that they should build on top of this. I’ve worked in Tesla Autopilot before and saw it hit a mannequin because we never had mannequin in our dataset before ( we might have had it in the data but it was not a part of our ontology for the network to predict). One approach to mitigate this was OpenAI’s clip that used the English Language as classification labels but Meta’s SAM is so much better where it detects objects without need to specify language. It just understands scenes and objects, at a fundamental level, it can detect anything in a picture if you prompt it right. Honestly it feels a lot like GPT1 which was also ignored by most. If you prompt it right you can get it to segment anything in an image, but prompting it right requires human input. However I can imagine the third or fourth version, with some RL sprinkled in just working zero-shot on complete pixel understanding of any image in the world. This was one of the holy grail of computer vision, that we are seeing solved right in front of our eyes

I feel like the whole ChatGPT bubble has really highlighted what feel like some fundamental shortcomings in the worldview that is represented in tech. That is, there seems to be a winner take all dynamic baked into the tech world. Maybe this arises from the simple fact that tech places a lot of power in the hands of individuals. But there&#x27;s an emergent downside to this which is that it makes those who were already much more powerful even more so. Because who is best suited to take advantage of all the power but those who were already perched up above everyone else and permitted to pick and choose their opportunities?<p>I&#x27;ve found it exceptionally hard to stay positive about all of this. It almost feels as though the advent of LLMs has shined light on a fundamental law of the universe that does not work out in the little person&#x27;s favor. It&#x27;s like survival of the fittest on steroids. Guys, what the heck are we doing??

AI world can be proper dystopian these days. I know someone who accepted a job offer in one of the biggest market research firms in the world. She was making AI models for a high tech company before and was hired to make summarization AIs. Between serving out the (frankly ridiculous) 2 month notice period and joining this new job GPT-4 was launched and prices came down for 3.5. The original plan for building something internally was turned into feed everything to chatGPT and then ask it questions. A junior engineer whipped up this system in a couple of weeks. Its all just API calls after all. Now she spends entire days trying to ask the model the right questions so that it can generate the correct reports. Her entire job has been boiled down to talking to an AI. She is working with the most cutting edge technology, yet its so mundane that everyone is just bored on this project. This could have been a ironically tragic character in a Gibson novel.

I think what’s really depressing here is just how effective scaling seems to be. It just means that any company that’s not willing to pour hundreds of millions of dollars into their AI programs isn’t serious at all and would probably be better off hiring engineers to figure out how to integrate GPTX into their systems than trying to roll their own. I really think we’re going to see a massive collapse of AI&#x2F;data science jobs once it becomes clear that no in house model is <i>ever</i> going to be better than the zero shot performance of these mega models.

At least, people working on ML models that handle these tasks must be feeling terrible. They know their models will be abandoned sooner or later and composed on top an LLM.<p>1. Classification<p>2. Named Entity Recognition (NER)<p>3. Dialog Engine<p>4. Sentiment Analysis<p>5. Tone Analysis<p>6. Language Translation<p>7. Summarization<p>8. Tokenization<p>9. Simple NLP Tasks (part-of-speech tagging, dependency parsing, lemmatization, morphological analysis)<p>10. Sentence Segmentation<p>11. Content Parsing<p>12. Question Answering (Structured &amp; Unstructured)<p>13. Similarity<p>14. Grammar Correction<p>15. Speech to Text (ASR)<p>16. Text to Speech (TTS)<p>17. OCR<p>18. Image Recognition<p>19. Text Test Data Generation

I think this depends a lot on where you are working.<p>I&#x27;ve talked to academics who are getting discouraged that they don&#x27;t see how their approach to AI is going to be possible any more, with so much funding going toward the largest models from industry. On the other hand I&#x27;ve talked to startup founders building AI products whose business is booming because ChatGPT brought so much attention to the entire space.<p>We are certainly living in interesting times ;-)

Very interesting post. Sounds similar to the introduction of cars. Around the 1900s, there were hundreds of car manufacturers all jumping into this new market [1].<p>[1]: <a href="https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Timeline_of_motor_vehicle_brands" rel="nofollow">https:&#x2F;&#x2F;en.m.wikipedia.org&#x2F;wiki&#x2F;Timeline_of_motor_vehicle_br...</a>

I won&#x27;t try to predict the future as far as such ML models will take our jobs or not. However, I&#x27;d like to point out to most people that working on and studying fundamentals won&#x27;t be a time wasted. Even if programming disappears as an employable job, people will still need to understand and learn things. You&#x27;ll still need to know the math&#x2F;physics&#x2F;philosophy&#x2F;history&#x2F;literature&#x2F;art&#x2F;music etc. It might turn out that it&#x27;s going to be more important than ever for us to be more confident and strict in our own skills and knowledge. And I&#x27;d also suggest to broaden our horizons. We should become good at a couple of different things like the polymaths of old.<p>IF you have to spend hours debugging ML generated code for something critical you better know exactly what you&#x27;re doing. If you have to generate a Harry Potter Balenciaga meme video (the whole Balenciaga meme and all the different versions are hilarious btw) you better have some visual&#x2F;art and musical sense in order to make it funny and engaging.<p>I think the end of the world is greatly exagarated. We actually have much bigger problems than &quot;AI&quot;.

The middle class was already flagging, destroying all knowledge work is like the coup de grace.<p>Not sure how this isn&#x27;t clear to people.

Huge crypto vibes<p><pre><code> Every single person I know working in AI these days ... has been sparked by the ChatGPT moment.</code></pre>

One of the most frustrating thing is that GPT-4 training was finished 6 months ago. It might as well be a decade in the current movement of things.<p>What that mean: - OpenAI has such a large head start that they are only worried about safety, world disruption and getting people used to AGI. - The ideas&#x2F;project you have was tested by them months ago and is probably irrelevant already. - Insiders have a huge advantage.<p>One area where there is opportunity outside is making the models run on smaller HW (llama&#x2F;alpaca) and to see what we can do with them.

I fear for my team, really do. We work in audio processing tech. As a small team of five people residing in a third world country, we will never have a chance to compete with big companies because it looks like ML-based audio processing will soon be a commodity. Sad but it&#x27;s a reality!

I do honestly believe (and I am willing to take some hate for this haha) that paradoxically &quot;Ai people&quot;. E.g. People who are up to date with the latest papers, are good at pandas and experimenting can build a training validation testing pipeline know the difference between random forrest and gradient boosting etc. are one of he least qualified people to really take advantage of the GPT style models.<p>I think it will be the frontend people (if their app still needs one and a text interface isn&#x27;t enough). And the backend people (whose API churning speed will double through Copilot), if their APIs are still needed. And most and foremost the techy business people that can translate what this new chat thing means for an actualy real world business, like banking, insurance, agriculture, manufacturing, real estate whatever.<p>I think in this new world of few shot learning there is much much less room for ML Engineers. Its a bit like setting up your own servers in a cloud world. Sure it might make sense for the big guys, or for security reasons etc. but the vast majority are much better off generating 20 examples and using some API avoiding the immense risk and costs associated to building your own model.<p>So, in an interesting twist, AI is taking the AI jobs first.

In the past, we had notable individuals in various scientific or mathematical fields who had multiple successive breakthroughs throughout the duration of their career (Einstein, Feynman, Gödel, Hilbert, etc.)<p>With ML&#x2F;AI, that doesn’t so much seem to be the case. There are the early pioneers (Hinton, LeCun, Bengio, etc.), but it seems more as though they were the first to “discover” neural networks (that actually worked), and then the individual breakthroughs sort of stopped after that. This observation is not a jab at these people—rather, it’s because I wonder if in machine learning, unlike the more traditional areas of math and science, one person is just not able to test groundbreaking new ideas on their own anymore. A lot of the latest progress in ML comes from large companies consisting of teams of researchers who are largely unknown to most of the public.<p>I’m not quite sure what my point is, but personally it’s a bit sad to me that fundamental development in AI now appears to require a vast amount of resources that small teams or individuals don’t have access to. I suppose you could argue this is a similar situation to Bell Labs, but even in that case there were many distinct contributions from well-known individuals working there.

I&#x27;m probably way too late for this thought to get any traction &#x2F; discussion - but I have this weird feeling that openai screwed up and showed it&#x27;s &quot;cool new thing&quot; too early, and publicly.<p>As much as it pains me to say this, I don&#x27;t think the real money is in making this a service, or &quot;the product.&quot; I think the real money is in using AI internally as a puzzle piece of your backend - ie. the secret sauce behind xyz product.<p>I&#x27;m being very narrow here, but you can only do so much integrating what openai has built into your products - eventually &quot;everything&quot; providing data from the same model brings &quot;everything&quot; to the same level. In contrast if you train and create your own models to make xyz do something specific, nobody knows how it was done, or it surely makes it a lot harder to kang.<p>I have zero proof, but I suspect Google for instance has models that would literally obliterate what openai has shown capability wise. They&#x27;re probably not necessarily language models though. Again, nothing to stand on here but I doubt their search and analytics for example are driven by hard coded algorithms these days.<p>Bard may have been released sort of as a &quot;psh, we&#x27;ve been there done that&quot; when in reality they didn&#x27;t, because they never planned to make the models they were&#x2F;are working on &quot;publicly&quot; available to use. It makes me wonder if this is how Google has lead for some long with some areas - now openai sort of screwed it up for everyone by making it a service that can be integrated &#x2F; adopted by nearly anyone.<p>The only people I guess that are really going to know are the devs working for these big orgs, and I&#x27;m sure that lock and key knowledge.

Probably not the right place to post this, but I really want someone to build a ChatGPT service that reviews consumer EULAs and highlights the important stuff, and can tell me what changes with them each time I have to reaccept or get notified. It&#x27;s a subset of making legalese digestible, but bringing more visibility into the density of them would be wondrous.

LLM sizes have been increasing 10X every year for the last few years. Source is NVidia. That pace is staggering. GPT4’s lead is temporary and it’s a shiny ball distraction from doing dev on your own projects rather than being a user on someone else’s. Get to it and get over it. There are plenty of breakthroughs to be had.

can someone show some actual product features where AI is being used productively?<p>I dont wanna say it but I think this is a hype train headed for failure

&quot;Every single person I know working in AI these days (in both the academy and industry) has been sparked by the ChatGPT moment.&quot;<p>This does not ring true to me at all. Anecdotally, it feels the rush to get into AI (in both industry and academia, for both individuals and organizations) peaked around 2016-2020, post-AlexNet&#x2F;ResNet, around the time Transformers became very popular. Hiring for ML research roles in particular definitely slowed down in 2021, and 2022 of course saw a broader course correction across all of tech.<p>That said, I do agree that ChatGPT may be the &quot;first iPhone moment of AI&quot;, in that is the first mainstream, end-user application of deep learning that millions of people have really engaged with.

The article mentions something important and doesn&#x27;t elaborate, that is, to trust the scientific method.<p>I&#x27;ve seen a few times state of the art developments that appeared to upend everything in my field. Only then with the passage of time do we then see that, yes the development is great for x, but older approaches are still better for y. With &#x27;better&#x27; running a gamut of considerations, such as speed, accuracy, complexity, practicality, and so on.<p>ChatGPT is dazzling but I wouldn&#x27;t be surprised if in 10 years we find that traditional NLP techniques have better precision for, e.g., language detection or information extraction than GPT-x (I&#x27;m making this example up), that evaluations from 2023 weren&#x27;t rigorous enough, research was misleading, researchers were biased etc.<p>Hopefully cooler heads can prevail via the scientific method.

&gt; so for people focused on learning and doing good, some simple logic implies that there shouldn’t be an AI race.<p>I think the author got the order wrong: acknowledging the race is what reinforces &quot;safety&quot;. If there was no race whats the incentive to talk so much about safety?<p>On the other hand I have a sense (although I won&#x27;t bet on it) that just like Siri hype died out after a few months, so will chatGPT. the author (and the doom seekers who sign open letters) can rest assured for calmer days. After all, one still need to know what&#x2F;how to talk to chatGPT. So much so that now there&#x27;s are jobs for &quot;prompt &quot;engineering&quot;&quot; - it&#x27;s so funny, because on one hand we are surprised by how smart(?) the responses are and yet we need to engineer the right questions to get the &quot;best&quot; answers.

I continue to be underwhelmed by the impact of machine learning (which some people refer to as AI). It certainly has interesting applications, and I expect it be important as a generator of entertainment. I&#x27;m just tired of the repeated waves of hype the surround the field.

The article discusses what it feels like to work in AI currently. The ChatGPT moment has shaken up the entire industry, causing career changes and projects to be abandoned. The pace is very high and everyone is extremely motivated but simultaneously close to burning out. Prioritization is hard, and leadership and vision are strained. The article offers solutions, such as taking solace in the scientific method and being process-oriented, managing up, and managing competition. The author reminds readers that it takes a lot of consistent work and luck to catch a wave in AI. - I used pagechat.com to summarize

The annoying thing about AI is that it can do stuff but with like 90% accuracy. This means that if you want to do something that takes 10 AI steps, your accuracy is down to 35%.<p>This is also why Copilot can&#x27;t really do computer programming.

What does it mean to work in AI?

Go to CHATGpT and ask for a link to some examples and half of them are 404. There is certainly a drawback to having scraped the web to build the model

With chatgpt having the ability to have so much more personality it would be interesting in the future when comments are just overrun by bots with actual human-like comments and reply. Eg. Imagine if someone released a ‘Snark Bot’ that made Snarky Comments and replies. But most likely it’s going to up the game politically to try to sway peoples opinions

&gt; there is a serious be the first or be the best syndrome (with a third axis of success being openness<p>Apple seems to have focused more on being the &quot;best&quot; rather than the &quot;first&quot; for several hardware products (MP3 players, smartphones, tablets, smart watches, wireless earphones) and it seems to have worked out OK.<p>However I have no idea what their AI plans are.

&gt; All of these low-level concerns make working in AI feel like the candle that <i>burns bright and short.</i> I&#x27;m oscillating between the most motivated I&#x27;ve ever been and some of the closest to burnt-out I&#x27;ve ever felt. This whiplash effect is very exhausting.<p>My candle burns at both ends;<p>It will not last the night;<p>But ah, my foes, and oh, my friends—<p>It gives a lovely light!<p>(First Fig, Edna St Vincent Millay)

If there&#x27;s anything I learned from reading articles from the Xiph.org people, it&#x27;s that neural networks work a lot better when you pre-process their inputs with an algorithm. They need far less information, and can run in realtime when most of the work is being done by traditional computation instead of doing it in the neural network.

as a student, it is not a weird but interesting time to be in this space. while the hype is like web3 from a year or two ago, the &quot;advancements&quot; are confusing when you look deeper.<p>while i have a lot to learn, i am struggling to find real innovations being made in all the fancy models today. i feel that the major component of the recent developments is that we now have more money and hardware to throw into the problem. there are some clever methods employed for gpt and image GANs, but the core part is still the same decades-old theory we can finally achieve at larger scale than ever.<p>i&#x27;d like to be enlightened about what i am not understanding here, but it has only made it more important to start from the fundamentals.

I prefer working on the &quot;Garbage In&quot; side of AI instead of the &quot;Garbage Out&quot; side. There are more customers whose problem is that they have too much garbage, than customers who need more garbage.

It&#x27;s depressing for sure. Definitely the cause of some suicidal thoughts. I wonder how many people will actually start harming themselves because the value of their skills and self-worth will be destroyed.

Not sure why any researcher wants to go into this … other than for money. None of them will get credit for it, the big shots from 40 years ago will. Better to work against the grain rather than as part of a herd.

I totally get it. All the AI experts are picking their jaws off the floor. Harsh.

AI developer is the new web developer.